Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Characterization of ultrawideband radiowave propagation within the passenger cabin of a Boeing 737-200… Chuang, James Tzu-Ho 2007

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Notice for Google Chrome users:
If you are having trouble viewing or searching the PDF with Google Chrome, please download it here instead.

Item Metadata

Download

Media
831-ubc_2007-0363.pdf [ 6.25MB ]
Metadata
JSON: 831-1.0100775.json
JSON-LD: 831-1.0100775-ld.json
RDF/XML (Pretty): 831-1.0100775-rdf.xml
RDF/JSON: 831-1.0100775-rdf.json
Turtle: 831-1.0100775-turtle.txt
N-Triples: 831-1.0100775-rdf-ntriples.txt
Original Record: 831-1.0100775-source.json
Full Text
831-1.0100775-fulltext.txt
Citation
831-1.0100775.ris

Full Text

CHARACTERIZATION OF ULTRAWIDEBAND RADIO WAVE PROPAGATION WITHIN THE PASSENGER CABIN OF A BOEING 737-200 AIRCRAFT by J A M E S T Z U - H O C H U A N G B . A . S c , The University of British Columbia, 2004 A THESIS S U B M I T T E D I N P A R T I A L F U L F I L M E N T OF T H E R E Q U I R E M E N T S F O R T H E D E G R E E OF M A S T E R OF A P P L I E D S C I E N C E in T H E F A C U L T Y OF G R A D U A T E S T U D I E S (Electrical and Computer Engineering) T H E U N I V E R S I T Y OF B R I T I S H C O L U M B I A October 2007 © James Tzu-Ho Chuang, 2007 Abstract In 2002, ultrawideband ( U W B ) systems gained prominence in the field of radar, communications and sensor networks when F C C made their ruling on the unlicensed commercial use of the 3.1-10.6 G H z spectrum. Channel modeling for U W B thus became extremely important for the evaluation of newly proposed systems. Currently, based on the envisioned applications, standard channel models based on the Saleh-Valenzuela (S-V ) model have been developed by two dedicated I E E E task groups, 802.15.3a and 802.15.4a for four environments: residential, office, outdoor and industrial. However, with increasing demand for wireless connectivity, wireless devices are being deployed in more areas that have not yet been well characterized. One such environment is the public transportation. In this thesis, we have made three major contributions. First, the identification of clusters is the essential first step needed for the extraction of S-V model parameters; however, this process is still being done through subjective visual inspections. Here, we remove that subjectivity and make the process more consistent by developing an automated cluster identification algorithm based on performing regression analysis on exponentially decaying clusters expressed in semi-logarithmic scale. Second, based on extensive measurements, we characterize the large-scale aspects of U W B propagation in the passenger cabin of a Boeing 737-200 aircraft; an environment that is fundamentally different from environment previously considered due to its confined volume, cylindrical structure and high passenger density. Several noteworthy aspects include: (1) the coverage within the passenger cabin is found to follow a chevron shape contour where it is the greatest along the aisle and the weakest around window seats which suggests that the path gain also depends on the seat location and (2) high passenger density can introduce significant excess path loss. Third, based on more extensive measurements, we model the small-scale aspects of U W B propagation which focused on the shape and duration of the channel impulse response and the small-scale fading of multipath components. In most cases, our results take the form of the parameters of the standard channel models and can be used directly for those planning to simulate, evaluate and deploy U W B systems in the aircraft environment. i i Table of Contents Abstract • ii Table of Contents iii List of Tables vi List of Figures vii List of Abbreviations x Acknowledgments xi Co-Authorship Statement xii Chapter 1 Introduction • 1 1.1 General Background and Motivation 1 1.2 Thesis outline 4 References 5 Chapter 2 Automated Identification of Clusters in UWB Channel Impulse Responses 6 2.1 Introduction '. 6 2.2 Cluster Identification Approach 7 2.3 Description of the Cluster Identification Algorithm 11 2.3.1 Local Smoothing in the Time Domain 12 2.3.2 Initial Search for Clusters 12 2.3.3 Iterative Search for Clusters 12 2.3.4 Recursive Partitioning 13 2.3.5 Generalized Cross-Validation Criterion 14 2.3.6 Additional Rules 14 2.4 Validation of the Cluster Identification Algorithm 19 2.4.1 Validation Using Simulated Channel Impulse Responses 19 2.4.2 Validation Using Measured Channel Impulse Responses 20 2.5 Possible Sources of Error : 24 ii i 2.5.1 Noise and Small-Scale Fading 24 2.5.2 Overlap between clusters ' 25 2.5.3 Anomalous clusters and additional rules 26 2.5.4 Coupling between input parameters 26 2.6 Conclusions 31 References 32 Chapter 3 UWB Radiowave Propagation within the Passenger Cabin of a Boeing 737-200 Aircraft .' 34 3.1 Introduction 34 3.2 Measurement Approach 35 3.2.1 UWB Channel Sounder 35 3.2.2 Channel Sounder Calibration 36 3.2.3 Data Collection 37 3.2.4 Measurement Database -. 39 3.3 Path Loss in the Aircraft Environment 42 3.3.1 Distance Dependence of Path Loss 42' 3.3.2 Three Dimensional Coverage Model 44 3.3.3 Frequency Dependence of Path Loss 51 ' 3.4 Time Dispersion in the Aircraft Environment 54 3.4.1 Delay Spread 55 3.4.2 Number of Dominant Paths 59 3.5 Conclusions 61 References •• 62 Chapter 4 UWB Channel Impulse Response within the Passenger Cabin of a Boeing 737-200 Aircraft . 65 4.1 Introduction 65 4.2 Measurement Approach 66 4.2.1 Channel Sounder Configuration and Calibration 66 4.2.2 Data Collection 67 iv 4.2.3 Measurement Database 68 4.3 Models of Multipath Characteristics 72 4.3.1 Initial Processing of the Channel Impulse Response 72 4.3.2 Modeling Strategy... 72 4.3.3 Shape of the Power Delay Profile 74 4.3.4 Interdependence of MPCs 81 4.3.5 Small-Scale Fading 82 4.3.6 Delay Spread within a Local Area 86 4.4 A Simulation Model for UWB CIR in Aircraft Passenger Cabin 86 4.5 Conclusions 90 References 91 Chapter 5 Conclusions and Recommendations 94 5.1 Conclusions : 94 5.2 Recommendations for Further Work 96 Appendix A Useful Tables 97 Appendix B Detailed Setup of E8362B PN A 98 Appendix C Calibration of Antennas 101 Appendix D Detailed Measurement Plan 103 Appendix E Matlab Code of the Automated Cluster Identification Algorithm 105 v List of Tables Table 2.1. User-supplied parameters used in the validation trials 20 Table 2.2. Comparison of actual and estimated parameters for different environments..21 Table 2.3. Consistency check for CM5 - 50 trials 21 Table 2.4. Estimated parameters from measured data in office and underground mine..21 Table 2.5. Algorithm parameters used for office and underground mine environments.21 Table 2.6. Evolution of RMS Error 27 Table 3.1. UWB Channel Sounder Link Budget 36 Table 3.2. Parameters of the distance dependent path gain model 46 Table 3.3. Parameters of shadowing effects 46 Table 3.4. Parameters of the frequency dependent path gain model 53 Table 3.5. RMS delay spread increase rate, a : 56 Table 3.6. Mean excess delay, RMS delay spread, number of significant paths, and energy captured for different thresholds levels 59 Table 4.1. Power delay profile model parameters 76 Table 4.2. Small-scale fading parameters 83 Table 4.3. Time dispersion parameters within a local area 86 Table A. 1. Dimensions of modern aircraft 97 Table A.2. Large-scale parameters of conventional environments 97 Table A.3. Time dispersion parameters of conventional environments 97 Table B . l . E8362B PNA option 014 port description 98 Table B.2. Setting of E8362B PNA 99 Table B.3. Link budget of the UWB channel sounder 99 vi List of Figures Figure 2.1. Model of the sparse multi-cluster channel impulse response (S-V Model).. 10 Figure 2.2. Model of the dense single-cluster channel impulse response with uniformly distributed ray arrivals 10 Figure 2.3. Flow chart of the automated cluster identification algorithm 16 Figure 2.4. Locally smoothed power delay profile where the black dots represent the smoothed MPCs 17 Figure-2.5. Evolution of the cluster identification process in a C1R generated from CM3 . - Office LOS model. Here, (a) two clusters are broken into (b) three clusters, the second of which is broken on (c) the next iteration to yield four clusters.. 18 Figure 2.6. (a) Identification of clusters in a CIR measured in an office, (b) Example of an anomalous cluster in a CIR measured in an office 22 Figure 2.7. (a) Identification of clusters in a CIR measured in an underground mine, (b) Example of anomalous clusters in a CIR measured in an underground mine 23 Figure 2.8. A PDP with its features buried by small-scale fading 28 Figure 2.9. Components of the residual error 28 Figure 2.10. Identified clusters for CM5. The total number of clusters is 12 and are marked using crosses while the estimated number of clusters is 10 represented by straight lines 29 Figure 2.11. Typical CIR for CM2 29 Figure 2.12. Change in the reduction of RMS error for different smooth interval, N . ... 30 Figure 2.13 Change in reduction of RMS error for penalty coefficient, a 30 Figure 3.1. Location of the transmitting, "A, and receiving, O, antennas on a Boeing 737-200 aircraft for (a) the p-to-mp and (b) p-to-p configurations during the development runs. In the production runs, only one side of the aircraft and only one transmitting antenna location at the font aircraft are considered 40 Figure 3.2. Cross-section of the passenger cabin and the typical antenna mounting positions for the point-to-multipoint and point-to-point configurations 41 Figure 3.3. A photograph of the typical receiving antenna location (armrest) 41 vii Figure 3.4. Distance dependent path gain for p-to-mp configuration. Dotted squares represent receiving antenna mounted on the headrest of aisle seats 47 Figure 3.5. Distance dependent path gain for p-to-p configuration 47 Figure 3.6. Log-normal fit to location variability for p-to-mp configuration 48 Figure 3.7. Log-normal fit to location variability for p-to-p configuration 48 Figure 3.8. UWB coverage within the passenger cabin of a Boeing 737-200 aircraft with the receiving antenna mounted at headrest, (a) is the measured coverage and (b) is the regenerated coverage. The unit for the colorbar on the right is dB. 49 Figure 3.9. Shadowing region and the corresponding distances that describe it 50 Figure 3.10. Multiband-UWB spectral map. 'O ' = Usable bands, ' X ' = Unusable bands, and ' A ' = Usable bands only with detect and avoid schemes 53 Figure 3.11. Power delay profile of a typical LOS channel (ceiling to headrest) 57 Figure 3.12. Power delay profile of a typical LOS channel (ceiling to footrest) 57 Figure 3.13. RMS delay spread obtained for p-to-mp configuration 58 Figure 3.14. RMS delay spread obtained for p-to-p configuration 58 Figure 3.15. CDF of the number of significant paths for p-to-mp configuration 60 Figure 3.16. CDF of the number of significant paths for p-to-p configuration 60 Figure 4.1. Layout of a Boeing 737-200 aircraft. Circles represent the location of the receiving antennas and the triangles represent the location of the transmitting antenna, in each transmitting antenna location, spatial sampling is performed according to the measurement grid. The squares represent the measurement taken from our first measurement campaign as described in [12] ; 70 Figure 4.2. Cross-section of the passenger cabin and the typical antenna mounting positions for the point-to-multipoint and point-to-point configurations 71 Figure 4.3. A photograph of a typical receiving antenna location (armrest) 71 Figure 4.4. APDP measured with receiving antenna mounted at the headrest of row 19. ...77 Figure 4.5. APDP measured with receiving antenna mounted at the armrest of row 19. 77 Figure 4.6. APDP measured with the receiving antenna mounted at the footrest of row 19 78 viii Figure 4.7. Excess amplitude of the LOS path, Kr, as a function of distance for LOS channels -. '. 79 Figure 4.8. Confirming the log-normality of the deviation of excess amplitude of LOS path, sK 79 Figure 4.9. Exponential decay rate, y, as a function of distance for N LOS channels 80 Figure 4.10. Confirming the log-normality of the deviation of exponent decay rate, sy. 80 Figure 4.11. Averaged spatial correlation as a function of antenna separation with the receiving antenna mounted on the footrest of row 84 Figure 4.12. A fit of the path amplitudes against theoretical distributions 84 Figure 4.13. Estimated m-factors as a function of delay for receiving antenna mounted on the headrest of row 19 85 Figure 4.14. The lognormal fit of m-factors for the receiving antenna mounted at the footrest of row 19 85 Figure 4.15. Comparison of the measured and regenerated APDP 88 Figure 4.16. Comparisons of the distribution of the simulated and measured RMS delay spread for LOS and NLOS channels 88 Figure 4.17. Comparisons of the RMS delay spread with respect to distance for NLOS channels 89 Figure B. 1. Connections of the UWB channel sounder 100 Figure C . l . Measurement setup of the antenna transfer functions 102 Figure C.2. Averaged antenna transfer functions : 102 Figure D . l . Detailed measurement plan of p-to-mp configuration for (a) empty aircraft, (b) half full aircraft and (c) full aircraft. ( A = transmitting antenna, O = receiving antenna, • = passengers) '. : 1 Figure D.2. Detailed measurement plan of p-to-p configuration. (A = transmitting antenna, O = receiving antenna) 1 •ix List of Abbreviations A o A : Angle of arrival A P D P : Averaged power delay profile A T F : Antenna transfer function B E R : Bit error rate C D F : Cumulative distribution function CIR : Channel impulse response C M # : Channel model # C T F : Channel transfer function DS-SS : Direct-sequence spread spectrum IFT : Inverse Fourier transform L O S : Line-of-sight M B - O F D M : Multiband orthogonal frequency division multiplexing M L E : Maximum likelihood estimation M P C : Multipath component N L O S : Non-line-of-sight P D P : Power delay profile P N A : Performance network analyzer p-to-mp : Point-to-multipoint p-to-p : Peer-to-peer R M S : Root mean square S-V : Saleh-Valenzuela U W B : Ultra-wideband V N A : Vector network analyzer x Acknowledgments This work was supported by grants from Be l l Canada (through its B e l l University Laboratories R & D program), Nokia Canada, and the Natural Sciences and Engineering Research Council of Canada. We are grateful to the management and staff of the B C I T Aerospace Technology Campus at Vancouver International Airport for providing our research group with access to their Boeing 737-200 aircraft (a donation from West Jet Airlines) during the course of this study. I thank my colleagues N i X i n and Shahzad Bashir and undergraduate students Ivan Chan, Alex Lee, Chris Pang, Cecil ia Yeung, and Chad Woodworm for their considerable assistance during the data collection phase of this study. I also thank Dr. Michelson for the all the useful guidance, suggestions, and patience over the past three years. x i Co-Authorship Statement A version of Chapters 2, 3 and 4 in this thesis w i l l be submitted to I E E E transactions for publication, [1] J. Chuang, S. Bashir and D . G . Michelson, "Automated Identification of Clusters in U W B Channel Impulse Responses," to be submitted to IEEE Transactions on Wireless Communications. [2] J. Chuang and D . G . Michelson, " U W B radiowave propagation within the passenger cabin of a Boeing 737-200 aircraft," to be submitted to IEEE Transactions on Vehicular Technology. [3] J. Chuang and D . G . Michelson, " U W B channel impulse response within the passenger cabin of a Boeing 737-200 aircraft," to be submitted to IEEE Transactions on Vehicular Technology. Each project was identified and initiated by Dr. Michelson. In the first paper, Dr. Michelson provided useful suggestion on validating the algorithm using both simulated and measured data. Both M r . Bashir and Dr. Michelson provided useful discussions during the development of the automated cluster identification algorithm and more importantly, the preparation of the manuscript. In the second and third manuscripts, Dr. Michelson was in charge of acquiring access to a Boeing 737-200 aircraft from British Columbia Institute of Technology, recruiting volunteers to collect the measurements. Dr. Michelson also helped with the planning of all the measurement campaigns that took place in the passenger cabin of the Boeing 737-200 aircraft, and contributed greatly to the organization and editing of the two manuscripts. x i i Chapter 1 Introduction 1.1 General Background and Motivation U W B signals are defined as signals with an absolute bandwidth greater than 500 M H z or with a relative bandwidth greater than 20% of its center frequency. The wide bandwidth have attracted great interest and created many new possibilities in the field of radar, communications and sensor network. Some of the benefits of the U W B signals include: 1. excellent ranging and positioning capabilities, 2. no significant multipath fading due to fine time resolution, 3. multiple access due to wide bandwidth, 4. possibility of very high data rate 5. covert communication due to low transmit power, and 6. possible easier material penetration due to the wide bandwidth transmitted [1]. In 1990s, U W B system gained prominence when W i n and Scholtz published their groundbreaking work on time-hopping impulse radio systems [4]-[6]. In 2002, the ruling made by the Federal Communications Commission (FCC) granted the unlicensed commercial use of U W B spectrum from 3.1 to 10.6 G H z and fueled the development of U W B systems. Currently, two major industrial alliances, Multiband Orthogonal Frequency Divis ion Multiplexing ( M B - O F D M ) Alliance or M B O A and WiMed ia Alliance, are formed in support of the two major U W B schemes, M B - O F D M and direct sequence spread spectrum (DS-SS). Other than the industrial support, two dedicated task groups, I E E E 802.15.3a and I E E E 802.15.4a, are also formed to regulate the P H Y layer designs as well as the channel modeling aspects. In addition, the U W B technology has been adopted by poplar commercial applications like Wireless U S B and Bluetooth 3.0. The performance limitations of any communications systems are determined by the channel it operates in. After the F C C made their ruling in 2002, the channel modeling 1 subcommittee in the 3 a task group immediately started developing a standard channel model for U W B . Their main goal is to find a suitable model for the U W B channel such that different systems can use it to make fair comparisons against one another [7]. During the course of their work, they have adopted the well-known Saleh-Valenzuela (S-V ) model as the basis for their standard channel models. Their decision to use the S-V model was based on whether or not the model is capable of reproducing several key channel parameters such as power delay profile (PDP), R M S delay spread, and number of significant M P C s [7]. In their final report, based on the intended applications for U W B , i.e., high data rate multimedia access and data transfer, the 3a task group created the first four standardized models for two different environments: residential line-of-sight (LOS) , residential non-line-of-sight (NLOS) , office L O S , and office N L O S . Later, the increased interest in creating sensor networks using U W B technology has led to the formation of the 4a task group. The 4a task group focused on new environments such as outdoor (suburban microcell) and industrial (dense scatterers) environments where the sensor networks are most likely to be deployed. Several changes and improvements were made to the standard models developed by the 3 a task group which include introducing a new shape for the P D P for dense scattering environments with a "soft" onset, modeling the number of cluster and M P C s in the S-V model as a random variable, and incorporating the effect of frequency dependent path loss [2]. B y the end of 2004, again through a series of measurement campaigns, the 4a task group, other than modifying the original four channel models, created four more standardized channel models: outdoor L O S , outdoor N L O S , industrial L O S and industrial N L O S . In addition, they have also created a new model for body area networks which includes correlated shadowing. Their work on channel modeling eventually replaced the work done by the 3a task group. Despite the enormous effort put into channel modeling by the task groups, there are still a quite a few unresolved issues regarding the measurement and modeling of U W B propagation. Specifically, there are four major issues, 1. the identification of clusters within the CIR, 2. the impact of antennas on propagation measurements, 3. the effect of run time compensation, and 2 4. the finite bandwidth effect and the effect of windowing [1]. (1) The identification of clusters is the essential first step in extracting the relevant parameters used in the S-V channel model; however, they are still being done through subjective manual visual inspections which leads to rather arbitrary extracted model parameters [1] [9]. (2) The impact of antennas on propagation measurements is in general very difficult to remove. Different researchers have reported using different techniques to remove the effect of antennas and some of have reported results including the effects of antennas [1] [10]. (3) Spatial sampling within a local area is often used to determine small-scale effects; however, due to the fine time resolution, as the receiving antenna is moved from one spatial point to another, the M P C s tend to migrate from one delay bin to another. This effect on the small-scale fading statistics is still unclear [1]. (4) The effect of finite bandwidth which leads to the leakage of power in the time domain is also another major issue. Although techniques such as C L E A N , S A G E , and windowing can be utilized to reduce the effect of leakage, each of them has their own weakness. For example, C L E A N and S A G E algorithms require the assumption of a template waveform [10]. For windowing, although when it is applied the leakage power in the side lobes is greatly reduced, the resolution in time domain solution is also reduced [1]. With more and more mobile devices being built and the demand for wireless connectivity increasing, wireless devices are finding themselves deployed in more and more environments. One such environment is the vehicles used in public transportation, e.g., aircrafts, buses, trains and subways. What all these environments have in common is that they are typically in a confined volume, in a long cylindrical structure and the density of obstacles, such as passengers and luggage, can change quite significantly. These channel conditions thus make public transportation environment unique from the conventional environments previously investigated by the 3a and 4a task groups. Although in the past, numerous propagation measurements that facilitated the deployment of traditional narrowband systems were made within those environments, none of them characterized the channel for U W B signals. There are three main objectives in this thesis. The first objective is to address the main issues in the process of U W B channel modeling process. Due to time constraints, 3 we focused only on the first of the four main issues, the identification of clusters in the S-V model. Our intension is to develop an automated cluster identification algorithm that w i l l be more consistent and robust than the current manual techniques. Our next objective is to characterize the U W B propagation in vehicles used for public transportation. Specifically, we w i l l focus on characterizing the U W B propagation within the passenger cabin of a Boeing 737-200 aircraft a subset of the public transportations environment. Through extensive measurements, we wi l l first characterize the large-scale aspects of U W B propagation. Our data reduction efforts w i l l then focus on comparing those propagation characteristics with other conventional environments. Our final objective is to characterize the small-scale aspect of U W B propagation, again through extensive measurements and create models suitable for the design and evaluation of system performance. The results derived here wi l l be broadly applicable to a wide variety of modern aircraft due to the similarities in their cross-sections (See Table A . l ) . 1.2 Thesis outline This thesis is organized as follows. In Chapter 2, we presented an automated cluster identification algorithm. The algorithm removes subjectivity and unifies the definition of clusters in CIR. Chapter 3 presents the measurement data collected within the passenger cabin of a Boeing 737-200 aircraft, and with the measurement data, analysis of how the structure of the aircraft affects the coverage and reliability of U W B communication. In Chapter 4, a statistical channel model is developed for the U W B propagation within the passenger cabin of a Boeing 737-200 aircraft. The statistical model can be used for system simulations. Finally, in Chapter 5, we draw conclusions, assess the limitations of the present work and offer recommendations for future work. 4 References [1] A . F. Molisch, "Ultrawideband Propagation Channels-Theory, Measurement, and Modeling," IEEE Trans. Veh. Technol, vol . 54, no. 5, pp. 1528-1545, Sep. 2005. [2] A . F. Molisch, et al., " A Comprehensive Standardized Model for Ultrawideband Propagation Channels," IEEE Trans. Antennas Propag., vol . 54, no. 11, Nov. 2006, pp. 3151-3166. [3] A . A . M . Saleh, and R. A . Valenzuela, " A Statistical Model for Indoor Multipath Propagation," IEEE J. on Sel. Areas in Commun., vol . 5, no. 2, pp. 128-137, Feb. 1987. [4] R. A . Scholtz, "Multiple access with time-hopping impulse modulation," Proc. IEEE MILCOM1993, pp. 447-450, 1993. [5] M . Z . W i n and R. A . Scholtz, "Impulse radio: How it works," IEEE Commun. Lett., vol . 2, pp. 36-38, Feb. 1998. [6] M . Z . W i n and R. A . Scholtz, "Ultra-wide bandwidth time-hopping spread-spectrum impulse radio for wireless multiple-access communications," IEEE Trans. Commun., vol . 48, pp. 679-691, Apr. 2000. [7] J. Foerster et al, "Channel Modeling Subcommittee Final Report," Tech. Rep., I E E E 0249r0P802-15_SG3a, 2003. [8] A . F. Mol isch et al, " I E E E 802.15.4a Channel Model-Final Report," Tech. Rep., I E E E 802.1504-0062-02-004a, 2005. [9] O . H . Woon and S. Krishnan, "Identification of clusters in U W B channel modeling," Proc. IEEE VTC 2006 - Spring, M a y 2006. [10] B . A l l en et al, Ultra-wideband Antennas and Propagation, Wiley, 2007. 5 Chapter 2 Automated Identification of Clusters in UWB Channel Impulse Responses 2.1 Introduction The tendency for multipath components (MPCs) in wideband channel impulse responses (CIRs) to form exponentially decaying clusters was first reported by Saleh and Valenzuela over twenty years ago [1]. Their model, often referred to simply as the S-V model, was later adopted by both the I E E E 802.15.3a and 802.15.4a W P A N task groups as the basis for standardized models of the ultrawideband ( U W B ) CIRs that one is likely to encounter in residential, office, outdoor and industrial environments [2]-[5]. Because the M P C s observed in certain dense scattering environments are often so closely spaced in time that clustering can no longer be observed in the CIR, I E E E 802.15.4a also specified a dense single-cluster model to be used in cases such as N L O S channels in office and industrial environments where the sparse multi-cluster S-V model is not appropriate [6]-[8]. Other models for representing U W B CIRs have been proposed including the two-cluster model proposed in [9] and the lognormal-exponential model with a single cluster proposed in [10]. Some have raised concerns that clustering is often not apparent in CIRs measured in many conventional environments [11]. However, experience has shown that the two models proposed by I E E E 802.15.4a cover the vast majority of CIRs observed. Accordingly, we have focused on identifying clusters within a C IR assuming that the underlying model is the multi-cluster S-V model proposed by 802.15.4a. The dense single-cluster model can be treated as a special case. Despite the tremendous effort that has been applied to U W B channel measurement and modeling in recent years (and the importance of cluster identification to the process), A version of this chapter has been submitted for publication: J. Chuang, S. Bashir and D.G. Michelson, "Automated Identification of Clusters in U W B Channel Impulse Responses," submitted to IEEE Transactions on Wireless Communication, 12 October 2007. 6 most researchers still identify clusters in CIRs through time-consuming manual techniques that rely on subjective assessment by the analyst. Although a reliable automated U W B cluster identification algorithm has the potential to make cluster identification less subjective, more consistent and less time-consuming, three factors have limited their adoption: (1) a surprising lack of agreement concerning the criteria for identifying a cluster, e.g., [4] [5] [12], (2) a lack of sufficiently robust algorithms, as noted in [8][13], and (3) the effort required to implement or tune existing algorithms, e.g., [12][14]. Our efforts to develop a better cluster identification algorithm differs from past work in three ways: (1) we focus on the manner in which clusters of M P C s of given start time, duration, and exponential decay profile introduce discontinuities in the shape of the entire CIR, (2) we have explicitly considered the coupling between the parameters of the algorithm and (3) we have verified the correct operation of our algorithm (and the validity of our approach) by conducting estimation trials using both: (a) simulated U W B S-V CIRs with known channel parameters that correspond to residential, office and outdoor environments and (b) measured CIRs collected in both office and underground mine L O S environments. We also acknowledge the potential to use such an algorithm as the basis for either an autonomous reduction tool or as an interactive aid to the analyst. The remainder of this paper is organized as follows: In Section 2.2, we briefly review the U W B standard channel models adopted by the I E E E 802.15.4a task group and summarize our own cluster identification approach. In Section 2.3, we describe the steps that comprise our cluster identification algorithm. In Section 2.4, we present the results of our efforts to validate the algorithm using both U W B CIRs generated using I E E E 802.15.4a's CIR simulation code and measured CIRs from both office and underground mine environments. In Section 2.5, we consider factors that could affect the performance of the algorithm. In Section 2.6, we summarize our contributions. 2.2 Cluster Identification Approach Our approach to cluster identification is based on identification of discontinuities in the shape of the CIR. The I E E E 802.15.4a task group has recommended using either a 7 sparse multi-cluster or a dense single-cluster model, as appropriate, to represent U W B CIRs [3]. The sparse multi-cluster model is based upon the well known S-V model and is given by MO = Z I av e x P ( M , / )S(t-T,-TkJ) (2.1) /=0 k=0 where the M P C s are modeled as Dirac delta functions, d\.), and a^i and fai are the amplitude and phase of the M i M P C in the /th cluster. L is the total number of clusters in the CIR and Kt is the total number of M P C s within the /th cluster. T/ and z>_/ represent the arrival time of the /th cluster and the kth M P C in the /th cluster, respectively. The expected shape of the P D P , here defined as the square magnitude of the CIR, is described as the product of two exponential functions, E {k ' | 2 } 0 0 e x p ( - T ' 1 r ) e x p [~Tk,i1 y) (2-2) where T and y are the cluster and intracluster decay constants respectively as depicted in Figure 2.1. The arrival time of the clusters and M P C s , T/ and x^i, are found to follow the Poisson distribution with arrival rates, A and X, respectively. For the dense single-cluster model, /'. e., without any discontinuities, the shape of the CIR has the following form E\\akj2\cc 1 - ^ e x p \ yrise J exp (2.3) where x denotes the attenuation of the first component, y r i s e determines how fast the P D P rises to its local maximum, y\ represents the decay at later times. The arrival rate of the M P C s is fixed, At, given by the inverse of the signal bandwidth as depicted in Figure 2.2. In both models, the small-scale fading follows the Nakagami distribution. The shape of a cluster may be greatly distorted by the effects of noise and small-scale fading. A s a result, our observation of the power delay profile is, y{r) = f(T) + s (2.4) where fix) takes the form of either (2) or (3) and e is a random variable from a combination of white noise and small-scale fading. B y expressing the P D P on a semi-logarithmic scale, the exponential decays are expressed as straight lines. Then, using linear regression techniques, we find a set of non-overlapping piecewise regression lines, such that 8 b]0 + bnr tx<z <t2 b20 + blxz t2<r < r3 (2.5) where Z?,o and bn are the intercepts and slopes of the best fit lines, respectively, and tt are the estimated arrival times for each of the estimated L clusters. Each regression line in g(f) corresponds to a cluster and the estimated arrival times, th correspond to the discontinuities in the PDP. A s a result, when g(z) is accurately estimated, tt w i l l provide us with the accurate location of the clusters. Then, from the estimated arrival time of each cluster, the channel parameters needed in the S-V model are extracted. 9 03 3 -> 1A, -> 1/A delay Figure 2.1. Model of the sparse multi-cluster channel impulse response (S-V Model) PQ D T3 3 I Yrise / \ A V A s At delay Figure 2.2. Model of the dense single-cluster channel impulse response with uniformly distributed ray arrivals. 10 2.3 Description of the Cluster Identification Algorithm The goal of our cluster identification algorithm is to determine the combination of exponentially decaying clusters that best fit the most significant M P C s in the power delay profile. Whether measured in the time or frequency domain, a measured channel response has a finite bandwidth that is determined by the instrument and/or the measurement process. The result is equivalent to convolving the true CIR with a sine function whose duration is inversely proportional to the bandwidth of the measurement. Before processing a measured CIR using our cluster identification algorithm, one must first remove the effects of the finite bandwidth either by windowing or deconvolution. Other pre-processing steps that must be performed include setting thresholds for noise, suppressing small-scale fading, and so forth, as outlined in [15]. Because the first set of results presented in the next section is based upon simulated PDPs, we were able to omit this pre-processing step. The second set of results is based upon measured PDPs that we suitably processed before applying our cluster identification algorithm. Computing the spatial average of several PDPs over a local area is one method for suppressing small-scale fading, and it should be applied whenever possible to obtain a better estimate of the shape of the PDP. Because the first set of results is based upon randomly simulated PDPs, spatial averaging could not be applied. Accordingly, we developed an alternative method for small-scale fading suppression that involves using an average filter in the time domain to smooth the rapid variations within the CIR caused by small-scale fading. This approach has been used in image processing to reveal edges against background noise [16]. Using such techniques allows one to apply our cluster identification algorithm to either spatially-averaged or instantaneous PDPs. A flow chart of our cluster identification algorithm is depicted in Figure 2.3. It includes the three main steps that we describe in more detail below: (1) local smoothing in the time domain, (2) the initial search for clusters by identifying gaps in the P D P , and (3) the iterative search for clusters based upon trial fitting of intracluster decay curves to the P D P . 11 2.3.1 Local Smoothing in the Time Domain The first step is to suppress small-scale fading through local smoothing in the time domain. The P D P w i l l typically be sampled at an interval, ts, that is inversely proportional to the bandwidth of the measured signal. In the U W B case, the bandwidth w i l l typically range from a minimum of 500 M H z to a maximum of 7.5 G H z yielding 0.13 n s < / , < 2 n s . We smooth the P D P by applying an averaging filter defined by, i N/2 PDPs[m] = — PDP[mN + i] (2.6) M , = - j v / 2 where N is the user-defined smoothing interval and M w i l l nominally equal TV + 1. Downsampling in this manner increases the sampling interval to Nts. In certain cases, a given tap may not contain any energy after we apply an amplitude threshold. Accordingly, we may use a value M < N + 1 in (2.6) to compensate. A representative P D P that has been generated using the standard channel impulse response generator developed by the I E E E 802.15.4a channel modeling committee is given in Figure 2.4. The dots represent the components of the P D P that result when the averaging filter given in (2.6) is applied with N = 9. 2.3.2 Initial Search for Clusters If the interval between successive M P C s exceeds a user-defined threshold, we consider this to be a gap between distinct clusters. The number of such gaps plus 1 gives the initial number of cluster Zj nj t. Such gaps often arise in simulated PDPs but, in our experience, rarely arise in measured ones. The user can usually select an appropriate value for this threshold by inspecting the P D P . Figure 2.5a shows a P D P that has been split into two clusters that are separated by a large gap that appears near T = 200 ns. 2.3.3 Iterative Search for Clusters Our cluster identification algorithm is based upon expression of the M P C s on a semi-logarithmic scale so that exponential intracluster decay profiles w i l l be displayed as straight lines with given slopes. We use linear regression methods to determine the 12 particular combination of straight lines that best fits contiguous groups of significant M P C s (given by PDPs[m] in (2.6)) in a least squares sense. The search begins with the assumption that the initial number of clusters is L - Lmit as determined by the methods described in Section 2.3.2. The algorithm then estimates goodness-of-fit by calculating the R M S error, where M is the total number of smoothed M P C s identified and g[m] is g(r) sampled at intervals Nts. If the resulting R M S error of any cluster exceeds a specified threshold, it is assumed that an unidentified cluster is causing a discontinuity in the PDP. In response, the algorithm increments the number of clusters by 1, tries all possible combinations of L straight lines that one might fit to contiguous groups of M P C s to form L clusters, then selects the combination that yields the best fit, i.e., the lowest R M S error. If the best fit satisfies the error threshold criterion, the process ends, and the number of clusters in the CIR is reported as L . Otherwise, it is assumed that the number of clusters must be incremented further and the process is repeated. A typical sequence is depicted in Figure 2.3.4 Recursive Partitioning A n exhaustive search of all possible cluster configurations w i l l ensure that the best possible solution is ultimately found. However, the number of possible combinations increases very rapidly as the number of clusters increases and soon becomes computationally intractable. During trials conducted upon both simulated and measured CIRs, we have observed that incrementing the number of clusters used to represent a C IR almost always involve subdividing an existing cluster. In statistics, this approach is referred to as recursive partitioning [18]. Here, the decision to further subdivide an existing cluster is also based upon whether or not the R M S error can be further reduced. Making this a rule in our search strategy and eliminating all other possibilities reduces the number of trials that we must conduct from OiM1'1) to just OtM) where M is the total number of smoothed M P C s in the P D P and L is the number of clusters in the CIR. (2.7) 2.5. 13 Implementing such a partitioning strategy dramatically reduces the number of combinations that must be examined compared to an exhaustive search. This partitioning strategy plays an important role in making our algorithm tractable. 2.3.5 Generalized Cross-Validation Criterion A s mentioned before in Section 2.3.3, a R M S error threshold sets the stopping condition for the algorithm. However, in order to ensure that the fast variations of the i P D P are not interpreted as clusters, we make the stopping condition more robust by adopting the generalized cross-validation ( G C V ) criterion, RMSE GCV = l aB M (2.8) that has been used in Error! Reference source not found, and elsewhere. Here, B is the number of parameters that we seek to estimate (in our case, the slope and intercept of each of the identified clusters), M is the number of smoothed M P C s and a is the user-defined penalty coefficient. A s the number of identified clusters increases, the GCV w i l l initially decrease with the R M S error. Past a point defined by the penalty coefficient, the increasing number of parameters B w i l l cause the GCV to increase. The results presented in Section 2.4 suggest that a penalty coefficient of 1 to 2 is adequate for U W B cluster identification. Based on the GCV criterion, the algorithm w i l l stop: (1) when the value of GCV is below the user-defined threshold or (2) when the penalty prevents the value of GCV from being reduced further. 2.3.6 Additional Rules In order to ensure the correct operation of the cluster identification algorithm, we enforce three additional rules: 1. The clusters must exceed a minimum length. 2. The power level at the start of a given cluster must be greater than the residual power level extrapolated from the previous cluster. 3. The slopes of the regression lines that we fit to the M P C s must be negative because the S-V model assumes that the clusters decay. 14 The three additional rules are jointly enforced during recursive partitioning by applying a multiplicative penalty coefficient, Bt, where i denotes the additional rule. The decision to split an existing cluster is now based on the lack-of-fit (LOF) criterion, LOF = ftfrfrGCV = 0}J3263 r R M S E n - , (2.9) x aB M . where the penalty coefficient, /?,, is set to 1 by default and set to an alternative value when the z'th additional rule is violated. Based on our experience, the optimal choices for these alternative values are: B\ = 10, B2 = 2 and = 2. The increase in the L O F criterion that occurs when an additional rule is violated wi l l reduce the likelihood that a cluster w i l l be subdivided. However, it w i l l not completely stop the algorithm from identifying clusters that have violated the rules i f there is sufficient reduction in the R M S error. 15 User Inputs i Measured CIR data 1 - Smoothing Interval [samples] 2 - Maximum separation between MPCs [ns] Local smoothing in the time domain using an averaging filter Initial search for clusters based on gaps in the PDPs 3 - Apply penalty coefficients, a, /?,. 4 - Acceptable RMS error threshold e IdBl a) Exhaustive Search: Find the best linear fit amongst all possible combinations of L clusters b) Recursive Partitioning: Find the best linear fit by subdividing a cluster into two clusters at all possible points within that cluster for every clusters in the PDP Extract model parameters Figure 2.3. F low chart of the automated cluster identification algorithm 1 6 Time [ns] Figure 2.4. Local ly smoothed power delay profile where the black dots represent the smoothed M P C s . 0 50 100 150 200 250 300 Time [ns] (a) Figure 2.5. Evolution of the cluster identification process in a C IR generated from C M 3 - Office L O S model. Here, (a) two clusters are broken into (b) three clusters, the second of which is broken on (c) the next iteration to yield four clusters. 17 Figure 2.5. Evolution of the cluster identification process in a CIR generated from C M 3 - Office L O S model. Here, (a) two clusters are broken into (b) three clusters, the second of which is broken on (c) the next iteration to yield four clusters. 18 2.4 Validation of the Cluster Identification Algorithm 2.4.1 Validation Using Simulated Channel Impulse Responses We validated our algorithm by: (1) generating random CIRs for different scenarios using the standard channel impulse response generator developed by the I E E E 802.15.4a channel modeling committee then (2) estimating the relevant model parameters based upon the clusters identified using our algorithm. The eight standard channel models correspond to line-of-sight (LOS) and non-line-of-sight (NLOS) instances of residential (CM1 and 2), office ( C M 3 and 4), outdoor ( C M 5 and 6) and industrial environments ( C M 7 and 8), respectively. The parameters that define each of the channel models are the average number of clusters, Z m e a n ; the cluster decay factor, T; the cluster arrival rate, A ; the intracluster decay factor, y; the M P C arrival rate, X; and the shadowing term, a. The I E E E 802.15.4a CIR generator presets the interval between samples, ts, to 125 psec. For each channel model, we generated 50 CIRs and converted them to PDPs in accordance with the procedures outlined in [3]. We then estimated the number of clusters, their peak amplitudes, and arrival time by applying our cluster identification algorithm to each P D P in turn using the user-defined parameters given in Table 2.1. Next, we normalized the amplitude and arrival time of the clusters and superimposed them in order to estimate T and A , then repeated the process for the M P C s in order to estimate y and X. Details of the parameter extraction are outlined in [3]. The actual and estimated values of the S-V model parameters are given in Table 2.2. Since our interest is identification of clusters, only parameters related to clusters and their exponential decay are shown. Even though all of the CIRs are affected greatly by both small-scale fading and the randomness of the Poisson distributed cluster arrival time, the estimated parameters are very close to the parameter settings that had been applied to the simulator. Also , the algorithm has correctly recommended that a single cluster be used to represent the environments that were simulated using the single-cluster dense scattering model. A s a further check of the performance of the cluster identification algorithm, the steps described above were repeated fifty times using CIRs based upon the C M 5 19 (Outdoor - LOS) model. We chose this model because it tends to exhibit the greatest number of clusters. We compare the mean and standard deviation of the estimates to the actual values of the parameters in Table 2.3. Once again, our estimates and the actual values generally agree. 2.4.2 Validation Using Measured Channel Impulse Responses To further validate our algorithm, we have also applied it to U W B CIRs that we measured under line-of-sight conditions in both office and underground mine environments. Two CIRs that are representative of those observed in each case are presented in Figure 2.6a and Figure 2.7a. The lines represent the identified clusters, and the dots represent the smoothed M P C s . Our estimates of the model parameters are summarized in Table 2.4. The input parameters used to derive these results are listed in Table 2.5. The office L O S parameters are based upon 49 measured CIRs; the underground mine CIRs are based upon 27 CIRs. In each case, the values are similar to what a human analyst might extract but were obtained with far less time and effort. Moreover, the algorithm functioned correctly and robustly despite the presence of significant departures from the ideal S-V case. Table 2.1. User-supplied parameters used in the validation trials. Residential Office Outdoor industrial Algorithm Parameter CM1 C M 2 C M 3 C M 4 C M 5 C M 6 C M 7 C M 8 LOS N L O S LOS N L O S LOS N L O S LOS N L O S Smooth Interval, N [samples] 9 9 11 11 11 11 7 11 Smooth Interval, N t s [ns] 1.125 1.125 1.375 1.375 1.375 1.375 0.875 1.375 Maximum Separation Time [ns] 20 20 20 20 20 20 20 20 Minimum R M S Error Threshold [dB] 2 2 2 2 2 2 2 2 Penalty Coefficient for Additional Splits, a 1 1 1 1 1 1 1 1 Minimum Cluster Length [samples] 2 2 5 5 5 5 2 5 Penalty Coefficient for Additional Rule 1, /?i 10 10 10 10 10 10 10 10 Excess Amplitude of New Clusters [dB] 2 2 3 3 3 3 2 3 Penalty Coefficient for Additional Rule 2, /?2 2 2 2 2 2 2 2 2 Penalty Coefficient for Additional Rule 3, /?3 2 2 2 2 2 2 2 2 20 Table 2.2. Comparison of actual and estimated parameters for different environments. Residential Office Outdoor Industrial CM1 CM2 CM3 CM4 CM5 CM6 CM7 CM8 LOS N L O S LOS NLOS LOS N L O S LOS NLOS Actual 3.0 3.5 5.4 1 13.6 10.5 4.75 1 Estimate 2.66 2.58 5.5 1 10.1 12.52 4.22 1 Actual 22.6 26.3 14.6 N / A 31.7 105 13.5 N / A Estimate 19.2 20.8 16.3 N / A 33.44 95.4 13.4 N / A Actual 0.047 0.12 0.016 N / A 0.045 0.024 0.071 N / A Estimate 0.034 0.02 0.022 N / A 0.036 0.027 0.059 N / A Actual 12.53 17.5 6.4 11.84 3.7 9.3 0.65 85.36 Estimate 9.02 12.6 9.9 10.65 6.8 11.62 14.64 93.29 Table 2.3. Consistency check for CM5 - 50 trials Channel Parameters Actual Mean of the Estimates Standard Deviation of the Estimates •^mean 13.6 11.2 0.37 r 31.7 34.4 0.61 A 0.045 0.041 5.2e-4 y 3.7 6.3 1.74 Table 2.4. Estimated parameters from measured data in office and underground mine. Channel Parameters Office LOS Underground Mine •^Tnean 7.0 3.05 r 23.32 20.77 A 0.067 0.0302 7 8.79 14.79 Table 2.5. Algorithm parameters used for office and underground mine environments. Algorithm Parameters Office LOS Underground Mine Smooth Interval, N [samples] 9 9 Smooth Interval, Nts [ns] 1.125 1.125 Maximum Separation Time [ns] 20 20 Minimum RMS Error Threshold [dB] 2 2 Penalty Coefficient for Additional Splits, a 1 1 Minimum Cluster Length [samples] 3 3 Penalty Coefficient for Additional Rule \,BX 10 10 Excess Amplitude of New Clusters [dB] 2 2 Penalty Coefficient for Additional Rule 2, B2 2 2 Penalty Coefficient for Additional Rule 3, B3 2 2 21 22 Or 200 Time [ns] (a) 23 2.5 Possible Sources of Error Several factors may affect the performance of the algorithm including: (1) noise and small-scale fading, (2) overlap between clusters, (3) the presence of anomalous clusters and (4) coupling between the input parameters. We discuss each of these in more detail below. 2.5.1 Noise and Small-Scale Fading The L M S linear regression techniques that we use to fit trial PDPs to groups of MPCs can be sensitive to extreme values or outliers that affect the slope and offset of the resulting regression lines. In UWB propagation measurements, small-scale fading is always present and can significantly mask the shape of the PDP [4]. For example, a PDP whose features are buried by small-scale fading is depicted in Figure 2.8 To illustrate the effect of small-scale fading on cluster detection, assume that there are only L clusters in the PDP and we have fit L regression lines. When L*L , the residual error can be expressed as the sum of two components, error (T) = bias (T) + X0. (2.10) as depicted is Figure 2.9. The random variable, Xa, is the combination of small-scale fading and measurement noise and the bias term is the difference between the envelope of the PDP and the current best regression line, g(r). The bias term in decibel scale can be described as, biasdB ( r ) = 10 log 1 0 E {|a(r)f} - y (r) (2.11) where E{\a(f)\2} is the envelope of the PDP as described in (2.2) and (2.3) for the multi-cluster and single cluster cases, respectively. Since the random error cannot be eliminated, the success of the algorithm rests on whether or not we can successfully reduce the RMS error by reducing the bias term with a better estimate, i.e., generate a more accurate g(r). For example, the error in the first regression line in Figure 2.5a can be reduced while the error in the second regression line cannot. As a result, cluster detection will be difficult when small-scale fading is significant. 24 In Table 2.6, how RMS error changes as a function of the number of clusters, L, based on our recursive partitioning is summarized in more detail. Here, we see how the algorithm breaks up clusters in order to identify new clusters. The algorithm stopped at iteration 9 because the RMS error for the entire PDP is no longer reducible. 2.5.2 Overlap between clusters The identification of clusters is also directly related to how the clusters arrive. In the extreme case, when the arrivals of all clusters are too close in time, the clusters become indistinguishable based on the shape, e.g., the dense single-cluster model used in dense scattering environment. Figure 2.10 shows an example of the identified clusters in a channel realization from CM5, the outdoor LOS environment. The lines represent the identified clusters and the crosses indicate where the actual clusters start. While most clusters are easily resolved, two clusters around 190 ns and two more around 250 ns cannot because the clusters within the two groups of clusters are not distinct To a lesser extent, the cluster decay rate T and the MPC decay rate y also affects our ability to distinguish the start of a new cluster. Assuming that the first cluster has a continuous decay profile, the arrival of a second cluster A T after the start of the first inserts a discontinuity of height, e x p ( - A r / r ) exp ( -Ar /x ) or, on a decibel scale, as f \ 1^ (2.13) J r where Ti is the arrival time of the new cluster and k is lOlogio(e). This discontinuity is essentially the bias term in (2.11). As the MPC decay rate increases, the clusters can arrive increasingly closer in time and still be successfully identified. From (2.13), we calculated the average magnitude of the discontinuities in the PDP for the standard channel models assuming that, on average, A T is the inverse of the cluster arrival rate, A . A l l channel models give reasonably distinct clusters except for CM2 where the discontinuities, on average, are only 0.7 dB. The relatively small discontinuities make the algorithm an unsuitable one for identifying the clusters defined 25 by C M 2 . On the other hand, because all channel models have their clusters identified by the same set of procedure and rules, relative to other channel models, it would reasonable to conclude that C M 2 is best represented by the single-cluster model. Figure 2.11 shows a typical C IR for C M 2 . Compared to the CIR in Figure 2.5 for C M 3 , the clusters in C M 2 are difficult to distinguish whether by human analyst or by algorithm. 2.5.3 Anomalous clusters and additional rules In measured C I R data, either because of small-scale fading or the nature of the reflectors in the environment, the shape of the P D P does not always follow the S-V model. A s mentioned before, in these cases, we add a penalty to the R M S error calculated so that the identified clusters w i l l more likely follow the underlying shape of the S-V model. Figure 2.6a and Figure 2.6b depict the identified clusters when additional rules are enforced and not enforced in the office environment. Figure 2.7a and Figure 2.7b depict the identified clusters when additional rules are enforced and not enforced in the underground mine environment. How well the clusters should follow the S-V model is up for the analyst to decide when he sets the penalty coefficients /?,. 2.5.4 Coupling between input parameters Correct operation of the cluster identification algorithm requires that the smoothing interval, the penalty coefficients and the R M S error threshold be appropriately set. The smoothing interval averages out the rapid variations in the P D P . Although the process is aided by increasing the interval, this also increases the risk of smoothing out the discontinuities that we are trying to identify. The analyst must therefore take care when setting the value used. Results obtained by applying the cluster identification algorithm to the CIR in Figure 2.10 for different smoothing intervals N are given in Figure 2.12. The intervals are given in samples where the sampling time is 0.125 ns. Each curve ends when the R M S error has reached its minimum value. In general, when the smoothing interval is very small and the penalty coefficient, a, is not large enough to prevent the creation of apparent clusters that are merely artifacts of small-scale fading, the algorithm wi l l tend to overestimate the number of clusters. A s the smoothing interval increases, the algorithm 26 converges to the correct value and the R M S error drops rapidly. Past a certain point, excessive smoothing blurs the transition between clusters and the number of clusters is underestimated. Figure 2.13 shows the reduction in R M S error as the penalty coefficient, a, increases. Each curve ends when the R M S error has reached its minimum (except for the case a = 1). Increasing the penalty coefficient, a, inhibits cluster splitting and thereby increases the rate at which the algorithm achieves the minimum R M S error. For all cases, the minimum R M S error threshold, here, 2 dB, has little effect because the R M S error never drops to that value. Table 2.6. Evolution of R M S Error #of Iter. RMS Error of Clusters [dB] RM S Erro rof entir e PDP 1.90 1 1.90 1 1.90 I 1.90 I 1.90 i 1.90 I 1.90 I 1.90 1 1.90 9.71 9.13 I 8.49 1 8.07 I 7.29 I 6.40 4 5.82 I 5.49 1 5.13 1 5.05 27 Cluster 1 Cluster 2 O Sig. MPCs Cluster 1 delay Cluster 2 Figure 2.8. A P D P with its features buried by small-scale fading. CQ -a • i u 3 Best Fit E{\a(r)\2} Cluster 1 Cluster 2 delay Figure 2.9. Components of the residual error. 28 Or Time [ns] Figure 2.10. Identified clusters for C M 5 . The total number of clusters is 12 and are marked using crosses while the estimated number of clusters is 10 represented by straight lines. 1 Clusters Time [ns] Figure 2.11. Typical C I R for C M 2 . 29 13 Number of Identified Clusters Figure 2.12. Change in the reduction of R M S error for different smooth interval, N . 11 7 I , , : , 1 5 10 15 20 Number of Identified Clusters Figure 2.13 Change in reduction of R M S error for penalty coefficient, a. 30 2.6 Conclusions We have developed an automated cluster identification algorithm that determines how a U W B channel impulse response (CIR) can be most effectively represented by either of the I E E E 802.15.4a standard channel impulse response models: (1) a single exponentially decaying cluster (a straight line when expressed on a semi-log scale) or (2) a sequence of exponentially decaying clusters described by the Saleh-Valenzuela model, as appropriate. Trials conducted using U W B CIRs generated by a simulation code developed by I E E E 802.15.4a and U W B CIRs measured in office and underground mine environments have confirmed the validity of our approach. Although the algorithm works best when applied to CIRs that have been expressed as spatially averaged PDPs, our use of local smoothing allows us to apply it to instantaneous PDPs with considerable success. Compared to previous work, our algorithm has several key features that contribute to its success: (1) We focus on the manner in which clusters of M P C s of given start time, duration, and exponential decay profile introduce discontinuities in the shape of the entire CIR, (2) We employ recursive partitioning to dramatically reduce the number of cluster combinations that must be checked and thereby make the algorithm tractable. (3) The iterative nature of the algorithm and the manner in which we layer the cluster selection rules makes it possible to use the algorithm as the basis for either (a) an autonomous tool for batch mode processing or (b) an interactive tool for use by analysts. Compared to current manual approaches, the proposed algorithm makes cluster identification more consistent and less time-consuming. Thus, the algorithm should be a useful aid for those engaged in analysis of U W B channel impulse responses. 31 References [I] A . A . M . Saleh and R. A . Valenzuela, " A statistical model for indoor multipath propagation," IEEEJ. Sel Areas Commun., vol . 5, no. 2, pp. 128-137, Feb. 1987. [2] A . F. Mol isch, J. R. Foerster, and M . Pendergrass, "Channel models for ultrawideband personal area networks," IEEE Wireless Commun., vol . 10, no. 6, pp. 14-21, Dec. 2003. [3] A . F. Mol i sch et al., " I E E E 802.15.4a channel model - final report," Tech. Rep., I E E E 802.15-04-0662-00-004a, Nov. 2004. [4] A . F. Mol isch, "Ultrawideband propagation channels - Theory, measurement, and modeling," IEEE Trans. Veh. Technol, vol . 54, no. 5, pp. 1528-1545, Sep. 2005. [5] A . F. Mol isch et al, " A comprehensive standardized model for ultrawideband propagation channels," IEEE Trans. Antennas Propagat., vol . 54, no. 11, pp. 3151-3166, Nov. 2006. [6] U . Schuster and H . Bolsckei, "Indoor U W B channel measurements from 2 G H z to 8 G H z , " Tech. Rep., I E E E 802.15-04-0447-00-004a, Aug . 2004. [7] J. Karedal, S. Wyne, P. Aimers, F. Tufvesson and A . F . Molisch, " U W B channel measurements in an industrial environment," Proc. IEEE Globecom 2004, pp. 3511-3516, Nov. 2004. [8] J. Karedal, S. Wyne, P. Aimers, F. Tufvesson and A . F . Mol isch, "Statistical analysis of the U W B channel in an industrial environment," Proc. IEEE VTC 2004 -Fall, pp. 81-85, Sep. 2004. [9] S. Venkatesh, J. Ibrahim, and R. M . Buehrer, " A new two-cluster model for indoor U W B channel measurements," 2004 IEEE APS Int. Symp. Dig, vol . 1, pp. 946-949, Jun. 2004. [10] S. S. Ghassemzadeh, L . J. Greenstein, T. Sveinsson, A . Kavcic , and V . Tarokh, " U W B delay profile models for residential and commercial indoor environments, IEEE Trans. Veh. Technol, vol . 54, no. 4, pp. 1235-1244, Jul. 2005. [II] L . J. Greenstein, S. S. Ghassemzadeh, S-C Hong, and V . Tarokh, "Comparison study of U W B indoor channel models," IEEE Trans. Wireless Commun., vo l . 6, no. l , p p . 128-135, Jan. 2007. [12] O . H . Woon and S. Krishnan, "Identification of clusters in U W B channel modeling," Proc. IEEE VTC 2006 - Spring, May 2006. [13] C - C . Chong and S. K . Yong, " A generic statistical-based U W B channel model for high-rise apartments," IEEE Trans. Antennas Propag., vol . 53, no. 8, pp. 2389-2399, Aug . 2005. [14] D . Shutin and G . Kubin, "Cluster analysis of wireless channel impulse responses with hidden Markov models," Proc. IEEE ICASSP, pp. 949-952, M a y 2004. [15] A . F. Mol isch, U . G . Schuster, and C - C . Chong, "Measurement procedure and methods on channel parameter extraction," Tech. Rep., I E E E 802.15-04-0283-00-004a, M a y 2004. [16] I. Gijbels, A . Lambert, and P. Qiu, "Edge-preserving image denoising and estimation of discontinuous surfaces," IEEE Trans. Pattern Anal. Mach. Intell., vol . 28, no. 7, pp. 1075-1087, Jul. 2006. [17] S. C. Chan and Z . Zhang, "Robust local polynomial regression using m-estimator with adaptive bandwidth," Int. Symp .Circuits Syst., May 2004, pp. III-333III-336. [18] J. H . Friedman, "Multivariate adaptive regression splines," The Annals of Statistics, vol . 19, no. 1, pp. 1-141, Mar. 1991. 33 Chapter 3 UWB Radiowave Propagation within the Passenger Cabin of a Boeing 737-200 Aircraft 3.1 Introduction To date, a variety of studies have been conducted in support of deployment of personal wireless technology aboard passenger aircraft, includes studies conducted by researchers at the German Aerospace Centre (DLR) (under the aegis of the European Union's WirelessCabin project), The Vahala lab at Old Dominion University, Boeing, Intel and elsewhere, e.g., [l]-[9]. In general, these studies have emphasized: (1) systems engineering studies and field trials for conventional wireless technologies such as cellular telephones, wireless L A N , and Bluetooth, (2) simulation of aircraft interiors using industry standard R F coverage prediction tools, and (3) measurement of R F coverage using client devices. In one instance, the wideband channel impulse response within the cabin was characterized using a commercial channel sounder [7]. For airlines, U W B wireless technology operating within the frequency band between 3.1 and 10.6 G H z holds great promise for enabling deployment of high data rate multimedia and network access services within aircraft passenger cabins and for facilitating operations and maintenance through deployment of sensor networks and precise positioning systems. However, past efforts to develop measurement-based models of the U W B propagation channel have focused on residential, office, industrial and outdoor environments [ 10] [ 11 ]. With its confined volume, cylindrical structure, and high density of occupancy, the passenger cabin of a jet aircraft is fundamentally different from those environments considered previously. The effect of human presence or body shadowing in confined spaces or over U W B channels has been considered in [8] [12]. To the best of our knowledge, however, no previous published study has characterized either the U W B propagation channel within aircraft passenger cabins or explicitly considered the effect of human presence on U W B propagation within such environments. 34 Here, we seek to characterize large-scale aspects of U W B propagation within the passenger cabin of a typical mid-size airliner. Based upon frequency response data collected over the range 3.1 - 10.6 G H z aboard a Boeing 737-200 aircraft with the cabin empty, with a small group of passengers occupying half of the seats in a portion of the cabin, and with that small group of passengers occupying virtually all o f the seats in a smaller portion of the cabin, we have characterized three large-scale aspects of U W B channel that affect coverage and reliability, i.e., the distance dependence of path loss, the frequency dependence of path loss and the location variability. We have also characterized time dispersion parameters such as the R M S delay spread, the number of significant multipath components (MPCs) and the corresponding percentage of energy captured by those M P C s . In Section 3.2, we describe our VNA-based U W B channel sounder, our procedure for calibrating the channel sounder, our data collection procedure, and our measurement database. In Section 3.3, we present the results of our path loss investigation. In Section 3.4, we present the results of our time dispersion investigation. Finally, we summarize our key findings in Section 3.5 3.2 Measurement Approach 3.2.1 UWB Channel Sounder Our U W B channel sounder consists of an Agilent E8362B vector network analyzer ( V N A ) , a laptop-based controller equipped with a GPIB interface, a pair of 15-m L M R -400 UltraFlex coaxial cables, a pair of Electro-metrics 6865 omni-directional U W B biconical antennas, and suitable tripods and fixtures for mounting the antennas at various locations throughout the aircraft. During data collection, a M A T L A B script running on the laptop controlled the V N A and logged the received data. The system link budget is given in Table 3.1. The transmit power was set to 5 dBm. The loss from each measuring cable is 9.1 dB. The intermediate frequency bandwidth of the V N A was set to 3 kHz ; this led to a noise floor of -107.2 d B m and a minimum sweep time of approximately 2 sec. A s a result, channel sounders of this type can only be used in situations where the channel is effectively static. Using the Friis 35 transmission formula with a pathloss exponent 2.2, we determined that our channel sounder can measure responses with an S N R ^ 10 dB at transmitter-receiver separation distances of up to 15m. During data collection, the V N A was configured to sweep from 3.1 to 10.6 G H z over 6317 frequency points. The frequency sampling interval of 1.1875 M H z corresponds to a maximum unambiguous excess delay of 842 ns or a maximum observable distance of 253 m. The frequency span of 7.5 G H z gives us a time resolution of 133 ps or 40 mm. Table 3.1. U W B Channel Sounder Link Budget Links Values Transmitted Power 5 dBm Cable Loss -9.1 dB* Transmit Antenna gain OdBi Path Loss -82.1 dB** Receiver Antenna gain OdBi Cable Loss -9.1 dB* Received Power -95.3 dBm Receiver Sensitivity -107.2 dBm System Margin 11.9 dB * Calculated from datasheet at the highest frequency 10.6 G H z ** Calculated using Friis transmission formula with a path loss exponent of 2.25 at a distance of 15m 3.2.2 Channel Sounder Calibration Before use, the channel sounder must be calibrated so that systematic variations in the amplitude and phase of the measured frequency response due to factors other than the propagation channel can be removed. The process involves two steps. The first step is to use the V N A ' s built-in calibration routines which are based upon a 12-term error model to compensate for all distortion between the points where the transmitting and receiving cables attach to the transmitting and receiving antennas, respectively. The second step, which is much more difficult, is to compensate for the distortion introduced by the antennas themselves. In [13], Mol isch considered how antennas can distort channel measurements and reviewed the various techniques that researchers have used in the past to compensate for 36 such effects. Although it is a relatively simple matter to compensate for the return loss of the antenna over the range of frequencies of interest, one must also account for the manner in which the antenna radiation pattern varies over the range of directions that rays may depart the transmitter and impinge on the receiver. It is generally very difficult to completely remove the effect of antennas from measurements. Not only does the non-uniform frequency response of U W B antennas significantly distort the arriving M P C , but the distortion is different for rays arriving from different directions. A s a result, complete removal of antenna-related distortion requires knowledge of the angle of arrival (AoA) of each M P C . Acquiring this information requires a space-time channel sounder with a real or virtual antenna array at the receiving end. However, the limited space available within the passenger cabin makes it difficult to deploy such an antenna array. Here, we use the approximate approach described in [14] to calibrate our antennas. It is based upon the assumption that all of the M P C s are uniformly incident from all directions. To obtain the correction data, we measured the frequency response between the ports of the transmitting and receiving antennas: (1) with both antennas in open space, (2) with the receiving antenna located in the principal plane of the transmitting antenna and (2) with the receiving antenna successively mounted in all possible orientations with respect to the transmitting antenna. We then average the set of frequency responses in order to determine the mean effective gain at each frequency. We refer to the result as the antenna frequency response (AFR) . We use the result to remove a good portion of the effect of the antenna by taking the ratio of the measured channel frequency response (CFR) and the average antenna frequency response. We measured the antenna frequency response of our antennas on the rooftop of the Electrical and Computer Engineering building at U B C . We avoided reflections from the surrounding environment by carefully placing the antennas away from local scatterers and mounting them on 3.4-metre high tripods. We minimized ground reflection by placing the antennas such that the transmission path spanned the gap between the two perpendicular wings of the building. A s expected, we found that the magnitude of the averaged A F R decreases with increasing frequency with a frequency dependent pathloss coefficient K o f approximately 0.8. 3.2.3 Data Collection 37 We collected our channel frequency response measurements within the passenger cabin of a Boeing 737-200 aircraft. The aircraft is 21 m in length overall, 3.54 m in width and 2.2 m in height. It can carry over 100 passengers. Plan and cross-sectional views of the passenger cabin are shown in Figure 3.1 and Figure 3.2, respectively. A photograph of the receiving antenna mounted at the armrest is shown in Figure 3.3. Because modern mid-sized airliners have similar cross-sections, our results should be generally applicable to a wide range of modern aircraft such as the A320 family from AirBus , the ARJ21 family from A C A C and the C R J series from Bombardier, as depicted in Table A . 1. In our measurement campaign, we considered two wireless system configurations: point-to-multipoint (p-to-mp) and peer-to-peer (p-to-p). In the p-to-mp configuration, the transmitting antenna, in the manner of an access point, was mounted on the ceiling, as suggested in Figure 3.2, and the user terminals, in the manner of mobile devices, were placed at the headrest, armrest and footrest level of the passenger seats throughout the aircraft as shown in Figure 3.3. In the p-to-p configuration, the transmitting antenna was mounted on the headrest, armrest and footrest level of a passenger seat. After accounting for reciprocity, this configuration yielded six unique antenna combinations. During our development runs for the point-to-multipoint configuration, we verified the static nature of the channel and the consistency of our measurements by comparing ten consecutive channel frequency response measurements over selected paths within the cabin. This verified that we could take just one sweep per location during production runs and thereby dramatically reduce the number of measurements needed to characterize the aircraft. We introduced further redundancies into our measurement database by: (1) putting the transmitting antenna at different locations and (2) measuring on both sides of the bilaterally symmetric passenger cabin, as shown in Figure 3.1. In particular, we collected channel response data at every other seat on both sides of the aircraft from row 4 to row 19, i.e., at 53 different seats, for each of three transmitting antenna locations. Once we had verified that the expected symmetries and similarities appeared in the results, we were able to take advantage of them to further reduce the number of measurements required to characterize the aircraft, e.g., by focusing only on the port side of the aircraft and using only one transmitting antenna location. 38 3.2.4 Measurement Database In our development runs, for an empty aircraft using the p-to-mp configuration, we collected data at 24 seats, i.e., at every other seat on one side of the aircraft from row 4 to row 19. When measurement data were collected with people aboard the aircraft, 19 volunteers occupied seats on the port side. In the first instance, they occupied every other seat, i.e., seats next to where the receiving antennas are placed, from row 4 to 16. A total number of 19 seats were sampled for this instance. In the second instance, the volunteers occupied every seat from row 4 to 10 and we sampled at every seat which gave a total number of 21 seats sampled. For p-to-p configuration, two transmitting antenna locations at the window and aisle seat of row 4 were examined. Again, the receiving antenna is mounted at every other seat on the port side but only from row 4 to 18 to sample the aircraft. The sampling strategy gave us a total number of 22 seats sampled for each of combination of transmitting and receiving antenna mounting positions. In total, the passenger cabin is sampled in great detail at 564 different locations. 39 C D (a) (b) Figure 3.1. Location of the transmitting, A , and receiving, O, antennas on a Boeing 737-200 aircraft for (a) the p-to-mp and (b) p-to-p configurations during the development runs. In the production runs, only one side of the aircraft and only one transmitting antenna location at the font aircraft are considered. 40 66 IN (1.68 M) INTERIOR TRIM-TO-TRIM 139.2 IN (3.54 M) xAntem f 1 1 1 1 \ I I / 86.6 N (2.20 M) 62.2 IN (1.58 M) 148 IN (3.76 M) Figure 3.2. Cross-section of the passenger cabin and the typical antenna mounting positions for the point-to-multipoint and point-to-point configurations. Figure 3.3. A photograph of the typical receiving antenna location (armrest). 41 3.3 Path Loss in the Aircraft Environment The U W B path gain model adopted by the I E E E 802.15.4a channel modeling committee is given by, Gp{d,f) = k f j\-"f j-\-lK (3.1) where d and / are distance and frequency, respectively, do and fc are the reference distance and frequency, n and K are the distance and frequency exponents, and A: is a constant that represents the path gain at the reference distance and frequency. The two independent variables, d and / , are assumed to be independent from one another and as a result, they are modeled separately. In this study, we refer to the path gain as the mean received power multiplied by the gains of the transmitting and receiving antennas divided by the transmitting power as measured by the V N A . That is, G J£&- (3.2) p p v y 3.3.1 Distance Dependence of Path Loss Distance dependent path loss is the most fundamental and most important parameter to characterize for wireless devices and is especially important for U W B devices due to their regulated low transmit power. The path gain model parameters derived here are very important for system designers to determine the coverage and reliability of wireless systems within the passenger cabin of an aircraft. Also , the peer-to-peer model is useful for understanding the effect of interference from other wireless devices aboard the aircraft. Following the recommendation by the I E E E 802.15.4a channel modelling committee, we took the average of the measured complex channel frequency responses across the entire span from 3.1 to 10.6 G H z in order to obtain the mean distance dependent path gain, 1 M •> 42 where M is the number of frequency steps. Path gain decreases with increasing transmitter-receiver separation due to the combined effects of spatial spreading and obstruction by cabin fixtures, seats and human presence. In decibels, the path gain with respect to distance is, GP(d) = GP0-lOn\og] \doj + Xa,d>d0 (3.4) where GPo, the intercept point, is the path gain at the reference distance, do, and I0nlog\o(d/do) is the excess path gain referenced to 1 m, n is the path gain exponent and Xa is a zero-mean Gaussian random variable with a standard deviation of a. B y fitting a linear regression line to the path loss data with respect to distance, we are able to determine the path loss exponent n and the intercept point GPo- The standard deviation of Xa is found by subtracting the mean values from the path loss data and fitting the differences to a Gaussian distribution. Figure 3.4 shows how path gain varies with respect to distance for the different receiving antenna mounting positions in the p-to-mp configuration. From Figure 3.4, we can see, as expected, that the path gain data points for the headrest channels are always greater than the armrest and footrest channels. Note that some data points from the armrest and footrest channels are receiving higher power because they are placed too close to the aisle as shown in Figure 3.3. Also , some of the headrest positions are not exactly line-of-sight (LOS) because some of the positions are shadowed by overhead compartments. If we only consider the case where the receiving antenna is placed on the headrest of an aisle seat, then the path loss exponent drops to 1.83 with an intercept of -39.5 dBm and location variability of 0.42 dB. Depending on the mounting position of the receiving antenna, path loss is expected to increase between 5 and 10 dB at greater distances. The difference shown here is expected to increase as more obstacles such as passengers and luggage are brought aboard the aircraft. The fit of Xa to a Gaussian distribution is shown in Figure 3.6. The slightly deviation of the data from a Gaussian distribution is because of the difference in mounting positions creating two significantly different channel conditions. For peer-to-peer configuration, because both the transmitting and receiving antennas are below the height of the seats for all the antenna mounting combinations considered, the path gain parameters derived are all very similar. The only exception is 43 when both the transmitting and receiving antennas are mounted at the headrest level. Figure 3.5 shows how path gain varies with distance for the p-to-p configuration. We have also, in Figure 3.5, marked the headrest to headrest data separately to show the difference. Unlike the point to multipoint case, the location of the receiver, whether it is on an aisle or window seat does not matter. In terms of shadowing, Xa, as shown in Figure 3.7, fits the Gaussian distribution very well . Table 3.2 gives a complete summary of how path gain varies for all the cases considered. In free space environments, the path loss exponent is equal to 2 as a consequence of spatial spreading. In conventional environments with L O S condition, the path loss exponent for U W B signals is often less than 2 because other than the L O S component, additional energy is being collected from the reflections from scatterers [15]. For example, in the industrial L O S channels with a lot of metallic scatterers, the path loss exponent is only 1.2 (See in the Appendix). In the aircraft passenger cabin environment, the channel is enclosed within a metallic cavity so we might expect that the path loss exponent to be around the same as industrial L O S channels. In practice, however, the measured path loss exponent is not as low as expected. We believe this is due to the large amount of R F absorbers within the cabin, e.g., seats, overhead compartments, etc. On the other hand, because there are no metallic objects that can completely block signals and energy often finds its way to the receiver due to the dense set of scatterers in the environment, the path loss exponent for non-line-of-sight (NLOS) channels is never too high (See Table A . 2 in the Appendix for comparisons). 3.3.2 Three Dimensional Coverage Model Because of the large yet predictable variance in path loss observed with respect to the different scenarios considered, a simple distance dependent path loss is no longer adequate. For instance, Figure 3.8a shows the two-dimensional coverage within the passenger cabin based on measurements. The degree to which the coverage is affected by the overhead compartments and the seats is apparent. To more accurately account for the path gain measured within the passenger cabin, path gain is decomposed into three components: (1) L O S path gain assuming there are no 44 obstacles within the aircraft, (2) deterministic shadowing due to the furnishings within the aircraft, and (3) random shadowing. The complete path loss equation is, G P ( ^ ) = G P 0 - 1 0 n L O S l o g 1 0 \d0; -GP,+Xa, d>d0 (3.5) where we simply added a GPS term to (3.4) to account for the deterministic shadowing and modified the path loss exponent, «LOS, such that it is derived from a reference case with no obstacles in between, i.e., with the receiving antenna mounted on the headrest of the aisle seats. To model the deterministic shadowing, three shadowing conditions, shadowing by overhead compartment only, by seats only, and by both, are identified. Each of these shadowing conditions corresponds to one or more antenna mounting configuration, e.g., in p-to-mp configuration, the receiving antenna is shadowed by both the overhead compartments and the seats when mounted at the armrest and is shadowed only by the overhead compartments when mounted at the headrest. The effect of shadowing is then quantified as, GPs(dx,dy) = adx + 3d2 (3.6) where the two new distances, dx and dy, account for how far the receiving antenna is in the shadowing region as shown in Figure 3.9. The corresponding coefficients, a and B, are then derived by applying linear regression techniques with respect to the reference case for each row and then averaged over all rows. Finally, the random shadowing is modeled using a zero mean Gaussian random variable. Table 3.3 gives a complete summary of the coverage models in the aircraft. The use of this model simplifies the different scenarios into a compact form and at the same time, it models more accurately the shadowing effects caused by the overhead compartments and seats. Figure 3.8b shows the reconstructed coverage within the passenger cabin without the random component. 45 Table 3.2. Parameters of the distance dependent path gain model. System Mounting Point Path loss 1-m intercept, Location Configuration exponent, N P L 0 [dBm] variability, a [dB] Point to Multipoint A l l 2.2 -40.5 2.7 C-to-H 2.1 -39.3 1.6 C-to-A C-to-F 2.2 2.3 -40.7 -42.6 2.3 1.3 A l l 2.2 -41.7 1.8 H-to-H 2.0 -40.7 1.3 H-to-A 1.7 -44.5 1.8 Peer to Peer H-to-F 1.7 -45.7 0.7 A-to-H 1.8 -43.5 1.0 A-to-A 2.3 -41.3 1.5 A-to-F 2.3 -41.4 1.6 F-to-H 2.6 -38.5 0.8 F-to-A 2.6 -38.9 1.2 F-to-F 2.5 -39.9 1.3 (C = Ceiling, H = Headrest, A = Armrest, F = Footrest) Table 3.3. Parameters of shadowing effects. Shadowing Condition a P Example [dB/m] [dB/m] Overhead Compartments 3.5 - p-to-mp, C-to-H Seats 1.1 3.30 p-to-p, H-to-A Both 1.5 5.07 p-to-mp, C-to-A (C = Ceiling, H = Headrest, A = Armrest, F = Footrest) 46 -45 -50 ~ -55 CD I "60 03 Q. -65 -70 -75 - Best Fit • Headrest o Armrest V Footrest 3 4 D D ^ a 7vO 5 6 7 8 9 10 Distance [m] Figure 3.4. Distance dependent path gain for p-to-mp configuration. Dotted squares represent receiving antenna mounted on the headrest of aisle seats. -45 Distance [m] Figure 3.5. Distance dependent path gain for p-to-p configuration. 47 0.999 F 0.001 I -I -5 -4 -3 -2 -1 0 1 2 3 4 5 Location Variability, Xo- [dB] F i g u r e 3.7. L o g - n o r m a l fit to locat ion var iabi l i ty for p-to-p conf igurat ion. 48 Headrest Measured 4 6 8 Distance [m] 10 12 -70 -65 -60 -55 -50 -45 (a) £ 1 4 6 8 10 Distance [m] 12 14 -70 -65 -60 -55 -50 -45 (b) Figure 3.8. U W B coverage within the passenger cabin of a Boeing 737-200 aircraft with the receiving antenna mounted at headrest, (a) is the measured coverage and (b) is the regenerated coverage. The unit for the colorbar on the right is dB. 49 Figure 3.9. Shadowing region and the corresponding distances that describe it. 50 3.3.3 Frequency Dependence of Path Loss The I E E E 802.15.4a channel modeling committee has adopted the following relationship for modeling the frequency dependence of path gain where, In a free space environment, the frequency dependence of path loss comes from the antennas only. For a typical omnidirectional antenna with constant gain, path loss increases with frequency with K = 2 [16]. This is a consequence of the effective aperture of the antenna scaling with frequency. In a real channel, frequency dependence can also be introduced by one or more of the following physical aspects of the channel: (1) diffraction across blocking objects, (2) scattering from rough surfaces, (3) wall penetration, with material reflection coefficients being frequency-dependent, (4) frequency-selective reflection from metallic objects of specific geometric shapes such as railings and gratings; and (5) vector superposition of overlapping signal waveforms in a dense multipath channel, altering the frequency content of individual M P C [16]. Frequency dependent path loss is a serious problem, because it greatly affects the coverage at different frequencies. It is especially important in systems that use Multiband Orthogonal Frequency Division Multiplexing ( M B - O F D M ) . For instance, Figure 3.10 shows a typical frequency dependent U W B C T F and the different government regulations on the unlicensed use of U W B systems for the United States, Europe, Japan, and Korea. Here, we see that in order to implement a worldwide design, the system designer needs to know how path gain changes with frequency to adapt to the different regulations set by the different regions of the world. Also , it has been reported in Error! Reference source not found, that depending on the value of K, the channel w i l l either act as an integrator or a differentiator that severely distorts the transmitted waveform and degrades the performance of coherent receivers. (3.7) 51 Based on (3.7), K is estimated by converting the frequency axis of the C T F into logarithmic scale and fitting a regression line across the C T F . The slope divided by 2 of regression line is then the desired K. Table 3.4 gives a complete summary of K observed at the different antenna mounting combinations where juK is the mean of K observed and oK is the standard deviation. The results are given in terms of both the radio and the propagation channels where K for the radio channel is defined as the channel including the effects of the antennas, i.e., from anything in between the transmitting and receiving antenna connectors, and K for the propagation channel is defined as the channel after we have applied the antenna calibration procedure described in Section 3.2 and removed the effects of the antennas. The most significant result here is the huge difference in the K observed for the radio and propagation channel. The K observed for the propagation channel is expected because in environments with similar geometry such as tunnels, path loss has been shown to decrease with increasing frequency due to the waveguide effect [17]. This unique effect also has not been reported previously in the U W B channel modeling. Comparing the K observed at the different receiving antenna mounting positions, we see that K varies slightly with respect to the different mounting configurations. This slight variation is mainly cause by the different antenna mounting positions having slightly different A o A distributions. We can also see that the variation in K is greater for the p-to-mp configuration, and this is because the difference in height between the transmitting and receiving antennas is greater for the p-to-mp configuration and this greater difference in height caused greater change in the A o A distribution. The propagation channel has less to do with this variation because the materials with the aircraft are consistent everywhere. For p-to-p configuration, the K obtained indicates that path loss is decreasing with the square of frequency. Comparing the K measured with the K used in the I E E E 802.15.4a channel models, we can see that the aircraft environment is closest to rich scattering environment like the industrial environment. Although in [18], we have observed a slight dependency between K and the separation distance possibly due to a cumulative effect of the materials within the aircraft, because the dependency is small, we w i l l assume, for simplicity, that this distance dependency is negligible and is just a part of aK. Our analysis here is applicable due to the large amount of scenarios considered 52 Table 3.4. Parameters of the frequency dependent path gain model. System Mounting Radio Channel Propagation Channel Configuration Point oK oK Point to Multipoint A l l 0.43 0.40 -1.52 0.40 C-to-H 0.09 0.33 -1.86 0.33 C-to-A C-to-F 0.49 0.71 0.37 0.22 -1.46 -1.24 0.38 0.22 A l l 0.77 0.17 -1.11 0.17 H-to-H 0.51 0.17 -1.36 0.17 H-to-A 0.81 0.24 -1.11 0.24 Peer to Peer H-to-F 0.93 0.08 -0.94 0.08 A-to-H 0.78 0.09 -1.09 0.09 A-to-A 0.79 0.17 -1.08 0.17 A-to-F 0.71 0.15 -1.16 0.15 F-to-H 0.73 0.13 -1.14 0.13 F-to-A 0.78 0.11 -1.09 0.11 F-to-F 0.75 0.11 -1.13 0.11 (C = Ceiling, H = Headrest, A = Armrest, F = Footrest) CO o O O O o o o o o o o o o o us A A O X X X o o o o X X X X EU X A O X X X X X o o o o o X Japan A A O X X X X X o 0 o o o X Korea 3.168 6.864 10.56 Frequency [GHz] Figure 3.10. Mul t iband-UWB spectral map. ' O ' = Usable bands, ' X ' = Unusable bands, and ' A ' = Usable bands only with detect and avoid schemes. 53 3.4 Time Dispersion in the Aircraft Environment The channel impulse response (CIR) can be described in term of a power delay profile (PDP) as such, ^ * H % ) | 2 = X M ( * - * * ) | 2 (3-8) k where ak is the amplitude coefficient and S(r — T0 is the impulse function at different delays. Figure 3.11 and Figure 3.12 depict the shape of typical PDPs measured for L O S and N L O S channels, respectively. A general sense of the shape, duration and structure of the CIR can be obtained by estimating the R M S delay spread, rrms, and the number of dominant paths. Such information is very helpful when evaluating the performance of U W B wireless communications systems. For instance, the ratio of R M S delay spread of a channel to the symbol period is often strongly correlated with the bit-error-rate (BER) experienced by a wideband system. The number of dominant paths is also very useful information for the design of rake receivers in U W B systems because it determines how many rake fingers are required. Whether measured in the time or frequency domain, a measured channel response has a finite bandwidth that is determined by the instrument and/or the measurement process. The result is equivalent to convolving the true CIR with a sine function whose duration is inversely proportional to the bandwidth of the measurement. Before processing a measured CIR, one must first remove the effects of the finite bandwidth either by windowing or deconvolution. Here, we apply a Kaiser window with /? = 7 to the channel frequency response (CFR) in order to suppress dispersion of energy between delay bins. Then, we convert the C F R s into channel impulse responses (CIRs) using an inverse Fourier transform (IFT). Specifically, the transformation was done directly in the complex baseband and without any zero padding in the frequency domain. The CIRs are then normalized to unit energy. Next, in order to remove the initial propagation delay, we define the start of the CIRs. For L O S channels, we define the start of the CIR as the first M P C that arrives within 10 dB and 10 ns of the peak M P C . For N L O S channels, we define the start of the CIR as the first M P C that arrives within 10 dB and 50 ns of the peak M P C . We remove 54 the propagation delay by setting the start time of the first arriving M P C to zero. These procedures are based upon the recommendations contained in Appendix IV of the final report of the I E E E 802.15.4a channel modeling committee [19]. 3.4.1 Delay Spread The mean excess delay, rmean, o f a P D P is defined as the normalized first-order moment, r = ^ = (3.9) mea" k The R M S delay spread,- rrms, is defined as the square root of the second central moment of a PDP, T =JT2 -(T f (3.10) rms \ mean \ mean } v ' where 1mean \ p n I \ (3-11) L , P » \ T k ) k Before any of the S-V model parameters are extracted, all M P C s with amplitudes that are more than 25 dB below the peak M P C are removed. This ensures that only significant M P C s are considered. The results obtained from the aircraft measurements are shown in Figure 3.13 and Figure 3.14. For both p-to-mp and p-to-p configurations, R M S delay spread increases with distance which agrees with the measurements of the aircraft environment that were presented in [7] and results for other environments that were presented in [20]. This increase in xrms as a function of distance is related to the decrease in power at greater distances and is very well known. A summary of how xrms changes with respect to distance, d, is presented in Table 3.5 where y is the distance dependent exponent as dJ. The large spread in rrms shown in Figure 3.13 and Figure 3.14 can be easily explained with the different channel conditions associated with the different antenna mounting points. For p-to-mp configuration, whenever the receiving antenna is placed at the headrest, the channel can roughly be classified as a L O S channel. A s a result, the 55 P D P w i l l always have a strong spike as shown in Figure 3.11, and the presence of the strong spike w i l l result in a much smaller TRMS as shown in Figure 3.13. The large spread in xrms in the headrest case itself is a result of shadowing caused by the overhead compartments. Whenever the receiving antenna is placed at the armrest or footrest, the channel is almost always N L O S except for a few rare cases along the aisle. A s a result, irms is always larger. However, since N L O S channels do not have a strong path, the variation in TRMS is much smaller. For peer-to-peer configuration, since L O S condition only exists for headrest to headrest configuration, only the data points corresponding to that particular case showed much lower TRMS-Although the increase of R M S delay spread with respect to distance is observed, the measured R M S delay spread is not as large as the results obtained in [7]. This is expected because as pointed out in [21], the range of R M S delay spread depends heavily on the measured bandwidth. Although the aircraft environment is enclosed in a relatively small metallic cavity, the R M S delay spread is still comparable to values obtained for residential and office environments. On the other hand, i f we compare the R M S delay spread observed to larger environments, such as outdoor or industrial settings, we can see that the R M S delay spread for the aircraft environment is much smaller. Table 3.5. R M S delay spread increase rate, a. System Configuration Mounting Point a Point to Multipoint A l l 1.08 C-to-H 0.63 C-to-A C-to-F 1.35 1.00 A l l 1.85 H-to-H 0.98 H-to-A 1.29 Peer to Peer H-to-F 1.60 A-to-H 1.61 A-to-A 1.88 A-to-F 1.84 F-to-H 1.95 F-to-A 2.07 F-to-F 2.04 (C = Ceiling, H = Headrest, A = Armrest, F = Footrest) 56 0 50 100 150 200 250 Delay [ns] Figure 3.12. Power delay profile of a typical L O S channel (ceiling to footrest). 57 35 30 I 25 •o CO CD „ „ I. 20 - Best Fit • Headrest o Armrest V Footrest 4 5 6 7 8 9 10 Distance [m] Figure 3.13. R M S delay spread obtained for p-to-mp configuration. CO c ~o CO 0 k_ Q . 00 >^  J2 0 Q CO 35 30 25 20 15 10 5 0 Best Fit • Headrest to Headrest o All Others 4 5 6 7 8 9 10 Distance [m] Figure 3.14. R M S delay spread obtained for p-to-p configuration. 58 3.4.2 Number of Dominant Paths To find the dominant paths, we first set 5 different threshold levels: 5 dB, 10 dB, 15 dB, 20 dB, and 25 dB from the maximum peak in a given P D P . Then, any M P C s in the PDPs are considered to be a dominant path i f the energy of the M P C is greater than the threshold. Figure 3.15 and Figure 3.16 are the C D F s of the number of significant paths seen in the P D P for the p-to-mp and p-to-p configurations respectively. Table 3.6 summarizes the number of dominant paths observed for each of the receiver cases and the percentage of energy captured with those identified paths. If we compare the number of dominant paths in the aircraft environment with other conventional environment under the same threshold level, we can see that the number of paths is significantly less when compared with environments like the residential or office environment. The number of path is about the same however, when we compare with industrial environment that has a rich amount of scatterers. This result is a natural consequence of propagation inside a large conducting cavity within which the signal can reverberate [8]. Table 3.6. Mean excess delay, R M S delay spread, number of significant paths, and energy captured for different thresholds levels. System Threshold tmean [US] ^rms Num. of % Power Configuration [dB] [ns] Paths 5 6.4 4.0 15 30 Point 10 9.6 7.3 75 55 to 15 13.1 11.0 180 74 Multipoint 20 16.2 14.7 324 87 25 18.5 17.9 501 95 5 12.4 7.7 33 27 Peer 10 17.3 12.1 159 63 to 15 21.6 16.2 354 84 Peer 20 24.5 19.5 572 94 25 26.2 21.8 801 98 59 Figure 3.15. C D F of the number of significant paths for p-to-mp configuration. Number of Paths (NP) Figure 3.16. C D F of the number of significant paths for p-to-p configuration. 60 3.5 Conclusions With its confined volume and cylindrical structure, and the dense and regular layout of its seating, the passenger cabin of a mid-sized airliner is physically quite different from the residential, office, industrial and outdoor environments considered by the I E E E 802.15.4a propagation committee. Based upon an extensive measurement database that includes both point-to-point and point-to-multipoint configurations, we have determined the range of distance and frequency dependent path gain exponents (1.7 to 2.6 and -1.09 to -1.86) and the R M S delay spread (4 to 22 nsec) that characterize this environment. Based upon these results, we conclude that the passenger cabin most closely resembles the previously characterized industrial environment. However, our measurements also reveal that the dense and regular layout of the passenger seats cause the two-dimensional coverage pattern to take the form of distinctive chevron-shaped contours with path loss increasing most rapidly along the window seats and least rapidly along the aisle seats. This suggests that a comprehensive path loss model for such environment could include antenna separation and distance offset from the cabin aisle, as input parameters. A three-dimensional model could be generating by including distance below the headrest as another input parameter. In most cases, our results take the form of the parameters of the corresponding models recommended by the I E E E 802.15.4a channel modelling committee and can be used directly in simulations of U W B propagation in an aircraft interior. Accordingly, our results w i l l assist: (1) those planning U W B deployments and field trials in aircraft, (2) those wishing to verify the results of electromagnetic simulations of aircraft interiors, (3) those wishing to simulate U W B aircraft systems with realistic channels, and (4) those wishing to predict the level of interference that one might experience from wireless devices in other locations within the passenger cabin. 61 References N . R. Diaz and M . Holzbock, "Aircraft cabin propagation for multimedia communications," Proc. EMPS 2002, 25-26 Sep. 2002. M . Youssef and L . Vahala, "Effects of passengers and internal components on electromagnetic propagation prediction inside Boeing aircrafts," 2006 IEEE APS Int. Symp. Dig., pp. 2161-2164, 9-14 Jul. 2006. M . Jafri, J. E ly and L . Vahala, "Comparative analysis of interference path loss coupling patterns on B-737 vs B-757 airplanes," Digital Avionics Systems Conference, vol . 1, pp. 6.B.5 - 61 - 10, 30 Oct. - 3 Nov. 2005, G . Hankins, L . Vahala and J. H . Beggs, "Propagation prediction inside a B767 in the 2.4 G H z and 5 G H z radio bands," 2005 IEEE APS Int. Symp. Dig, vol . 1A, pp. 791-794, 3-8 Jul. 2005. C. P. Niebla, "Coverage and capacity planning for aircraft in-cabin wireless heterogeneous network," Proc. IEEE VTC 2003 - Fall, pp. 1658-1662, 6-9 Oct 2003. G . A . Breit, H . Hachem, J. Forrester, P. Guckian, K . P. Kirchoff, B . J. Donham, " R F propagation characteristics of in-cabin C D M A mobile phone networks," Digital Avionics Systems Conference, pp. 9.C.5-1—9.C.5-12, 30 Oct. - 3 Nov. 2005. N . R . Diaz and J.E.J. Esquitino, "Wideband channel characterization for wireless communications inside a short haul aircraft," Proc. IEEE VTC 2004 - Spring, pp. 223-228, 17-19 M a y 2004. M . P. Robinson, J. Clegg, and A . C. Marvin, "Radio frequency electromagnetic fields in large conducting enclosures: effects of apertures and human bodies on propagation and field-statistics," IEEE Trans. Electromagn. Compat, vol . 48, no. 2, pp. 304-310, M a y 2006. R. Bhagavatula, R. W . Heath and S. Vishwanath, "Optimizing M I M O antenna placement and array configuration for multimedia delivery in aircraft," Proc. IEEE VTC 2007 - Spring, pp. 425-429, Apr. 2007. 62 [10] A . F. Molisch, J. R. Foerster and M . Pendergrass, "Channel models for ultrawideband personal area networks," IEEE Wireless Commun., vol . 10, pp. 14-21, Dec. 2003. [11] A . F . Mol isch et ah, " A comprehensive standardized model for ultrawideband propagation channels," IEEE Trans. Antennas Propagat., vol . 54, no. 11, pp. 3151-3165, Nov. 2006. [12] T .B . Welch et al, "The effects of the human body on U W B signal propagation in an indoor environment," IEEEJ. Sel. Areas Commun., vol . 20, no. 9, pp. 1778-1782, Dec. 2002. [13] A . F . Mol isch, "Ultrawideband propagation channels: Theory, measurement, and modeling", IEEE Trans. Veh. Technol, vol . 54, no.9, pp. 1528-1545, Sep. 2005. [14] C. C. Chong, Y . K i m and S. S. Lee, "Statistical characterization of the U W B propagation channel in various types of high-rise apartments," IEEE Wireless Commun. Networking Conf., vol . 2, pp. 944-949, Mar. 2005. [15] J. A . Dabin, A . M . Haimovich, and H . Grebel, " A statistical ultra-wideband indoor channel model and the effects of antenna directivity on path loss and multipath propagation," IEEEJ. Sel. Areas Commun., vol . 24, no. 4, pp. 752-758, Apr . 2006. [16] W . Q. Mal ik , D . J. Edwards and C. J. Stevens, "Frequency-dependent pathloss in the ultrawideband indoor channel," Proc. IEEE Int. Conf. Commun. 2007, vol . 6, pp. 800-804, Mar. 2007. [17] S. R. Saunders and A . Aragon-Zavala, Antennas and Propagation for Wireless Communications Systems, Wiley, 2007. [18] J. Chuang, N . X i n , H . Huang, S. Chiu, and D . G . Michelson, " U W B radiowave propagation within the passenger cabin of a Boeing 737-200 Aircraft," Proc. IEEE VTC 2007 - Spring, pp. 496-500, Apr. 2007. [19] A . F. Mol isch et al., " I E E E 802.15.4a channel model - final report," Tech. Rep., I E E E 802.15-04-0662-00-004a, Nov. 2004. [20] S. S. Ghassemzadeh, R. Jana, C. W . Rice, W . Turin and V . Tarokh, "Measurement and modeling of an ultra-wide bandwidth indoor channel," IEEE Trans. Commun., vol . 52, no. 10, pp. 1786-1796, Oct. 2004. 63 J. Karedal, S. Wyne, P. Aimers, F. Tufvesson, and A . F. Mol isch, " U W B channel measurements in an industrial environment," Proc. IEEE Globecom 2004, vol . 6, pp. 3511-3516, 29 Nov. - 3. Dec. 2004. 64 Chapter 4 UWB Channel Impulse Response within the Passenger Cabin of a Boeing 737-200 Aircraft 4.1 Introduction ^ In order to effectively predict and compare the performance of ultrawideband ( U W B ) wireless communication systems, the shape and duration of the channel impulse response (CIR), and the small-scale fading statistics experienced by individual multipath components (MPCs) , need to be accurately modeled. The placement of the fingers in rake receivers, the guard-time required in O F D M systems, and the design of mitigation techniques such as insertion of cyclic prefixes are all affected by time-varying time dispersion due to the propagation channel. The shape of the CIR also affects the performance of U W B ranging and positioning algorithms because it determines how well the algorithm w i l l be able to detect the first arriving M P C . In recent years, the channel modeling subcommittees of the I E E E 802.15.3a and 802.15.4a task groups devoted considerable effort to the modeling of U W B channel impulse responses under both line-of-sight (LOS) and non-line-of-sight (NLOS) conditions in residential, office, outdoor and industrial environments at ranges up to 15 m. The standard channel models that they developed allow fair comparison between alternative U W B systems under a range of channel conditions and deployment scenarios [1][2]. U W B systems operating between 3.1 and 10.6 G H z hold great promise for enabling high data rate multimedia and network access within the passenger cabin of an aircraft or facilitating operations and maintenance through deployment of low power UWB-based sensor networks. However, with its confined volume and cylindrical structure, the passenger cabin of an aircraft is fundamentally different from previously A version of this chapter has been submitted for publication: J. Chuang and D.G. Michelson, " U W B Channel Impulse Responses within the Passenger Cabin of a Boeing 737-200 Aircraft," submitted to IEEE Transactions on Vehicular Technology, 12 October 2007. 65 modeled environments. Although several research groups have made considerable progress in characterizing aircraft passenger cabins in support of deployment of more conventional wireless technologies such as Bluetooth and I E E E 802.11 wireless L A N [3]-[10], no other groups have yet reported results regarding U W B propagation in the aircraft passenger cabin environment. Previously, we have reported upon the large-scale aspects of U W B propagation within the passenger cabin of a typical mid-sized airliner [11][12]. Here, we focus on the small-scale aspects of U W B propagation. Based upon frequency, response data collected over the range 3.1-10.6 G H z between numerous locations aboard a Boeing 737-200 aircraft, we have characterized two principal aspects of small-scale propagation of the U W B channel: (1) the shape and duration of the channel impulse response (CIR) and (2) the small-scale fading statistics experienced by individual multipath components (MPCs) . Due to the similarities in the cross-section of other mid-sized airliners, our results w i l l also be generally applicable to the A320 family from AirBus , the ARJ21 family from A C A C , and the C R J series from Bombardier. The remainder of this chapter is organized as follows. In Section 4.2, we describe the configuration and calibration of our VNA-based channel sounder, our procedure for collecting data in the aircraft, and our measurement approach. In Section 4.3, we present our proposed model for the shape and duration of the P D P seen in the aircraft passenger cabin environment and the small-scale fading statistics measured. In Section 4.4, we describe how we modified the standard channel model simulation code developed by I E E E 802.15.4a to apply to the aircraft passenger cabin environment. We validate our model by regenerating CIRs and comparing the essential channel parameters with those derived from the measured CIRs. Finally, we summarize our key findings in Section 4.5. 4.2 Measurement Approach 4.2.1 Channel Sounder Configuration and Calibration The channel sounder configuration and calibration procedure used in this study are the same as those used in our recent study of the large-scale aspects of U W B 66 propagation. Because a more complete description of the setup and calibration can be found in [12], we w i l l only summarize the essentials in this section. Our channel sounder is based upon an Agilent E8362B vector network analyzer. Two Electrometrics 6865 U W B omni-directional biconical antennas are attached to the V N A through two 15-m long L M R - 4 0 0 UltraFlex coaxial cables. During the measurements, the start and stop frequencies are set to 3.1 and 10.6 G H z , respectively, the number of frequency points is 6317, the intermediate bandwidth is set to 3kHz and the transmit power is set to 5 dBm. We verified that the channel is static by taking 10 consecutive sweeps in several locations throughout the aircraft. A s expected, we found no significant differences between the sweeps. The V N A and the coaxial cable up to the antenna connectors are calibrated using the 12-term error model that is implemented by the V N A ' s calibration facility. The antennas are calibrated separately using an approximate approach described in [13]. This approach is based upon the assumption that the M P C s are arriving from all directions. A n averaged antenna transfer function (ATF) - essentially a measure of the average distortion (averaged over all angle of arrivals of the M P C s ) caused by the antennas - is obtained from a set of A T F measurements collected under free space conditions. Details of how we obtained our antenna correction data are summarized in [12]. 4.2.2 Data Collection We collected our measurement data within the passenger cabin of a Boeing 737-200 aircraft. The aircraft, which is 21 m in length, 3.54 m in width and 2.2 m in height, can carry over 100 passengers. Plans and cross-section of view of the passenger cabin are shown in Figure 4.1 and Figure 4.2, respectively. Here, we have considered both point-to-multipoint (p-to-mp) and peer-to-peer (p-to-p) wireless system configurations. In the point-to-multipoint configuration, the transmitting antenna is mounted on the ceiling in the manner of an access point, as suggested by Figure 4.2, and the receiving antenna is placed at the headrest, armrest and footrest level of the passenger seats throughout the aircraft. The different receiving antenna mounting positions not only represent the typical use cases such as using a cell phone (headrest), a laptop (armrest), or devices that might be contained in passengers' carry-on baggage (footrest) but also represent both line-of-67 sight (headrest) and non-line-of-sight (armrest and footrest) channel conditions. In the peer-to-peer configuration, the transmitting antenna is mounted on the armrest level. In this case, all the receiving antenna positions are considered as N L O S . In order to check the consistency of our results, we have introduced redundancies to the measurements that we collected during our development runs. For example, we compared the results obtained with the transmitting antenna successively placed in three different locations or with the receiving antenna successively mounted on opposite sides of the bilaterally symmetric aircraft. Exploiting symmetry effects have also allowed us to dramatically reduce the number of measurements needed to characterize propagation within the aircraft. 4.2.3 Measurement Database In the point-to-multipoint configuration, we collected data in the middle seats on the port side of the aircraft at row 4, 11 and 19. In each case, the transmitting antenna was mounted on the ceiling of the passenger cabin in the aisle beside row 2. A t each row, we collected 49 spatial samples by moving the transmitting antenna across a 7-by-7 grid with a spacing of 5 cm. B y collecting 49 spatial samples in this manner, we obtained enough data to obtain a good estimate of the amplitude statistics within a local area [14] [15]. The 5-cm spacing was chosen because it corresponds to half the wavelength of the lowest frequency. This ensures that the samples are sufficiently independent [14] [15]. We chose to move the transmitting antenna instead of the receiving antenna (as suggested in [15]) because: (1) it is too difficult to move the receiving antenna when it is mounted close to the seats and (2) this alternative approach was shown in [13] to successfully characterize the channel. The slant distances from the transmitting antenna to each of the rows considered are approximately 2, 7 and 13 m, respectively. We have also taken 9 measurements with a 3-by-3 5 cm spacing grid at rows 7 and 15. The measurements taken at these two rows are used only to characterize the shape of the P D P rather than estimate the small-scale fading statistics. According to [16], 9 measurements are sufficient to average out the small-scale fading and permit the true shape of the P D P to be recovered. In the peer-to-peer configuration, we collected measurements at row 11 68 with a 7-by-7 grid with 5 cm spacing. Additional data from measurements collected [12] are also used to derive the results in this chapter. 69 5 c n _L T A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A A 5cm • • • • • • • • • • • ) • ( • • ) • [n • ) • (• • ) • (• • ) • ( • •) • (• • ) • (• • ) • ( • Figure 4.1. Layout of a Boeing 737-200 aircraft. Circles represent the location of the receiving antennas and the triangles represent the location of the transmitting antenna. In each transmitting antenna location, spatial sampling is performed according to the measurement grid. The squares represent the measurement taken from our first measurement campaign as described in [12]. 70 INTERIOR TRIM-TO-TRIM 139.2 IN (3.54 M) ,20 M) L 148 IN (3.76 M) -1 Figure 4.2. Cross-section of the passenger cabin and the typical antenna mounting positions for the point-to-multipoint and point-to-point configurations. Figure 4 .3. A photograph of a typical receiving antenna location (armrest). 71 4.3 Models of Multipath Characteristics 4.3.1 Initial Processing of the Channel Impulse Response Whether measured in the time or frequency domain, a measured channel response has a finite bandwidth that is determined by the instrument and/or the measurement process. The result is equivalent to convolving the true CIR with a sine function whose duration is inversely proportional to the bandwidth of the measurement. Before processing a measured CIR using our cluster identification algorithm, one must first remove the effects of the finite bandwidth either by windowing or deconvolution. Here, we apply a Kaiser window with a beta value of 7 in order to suppress dispersion of energy between delay bins. Then, we convert the channel frequency responses (CFRs) into channel impulse responses (CIRs) using an inverse Fourier transform (IFT). Specifically, the transformation was done directly in the complex baseband and without any zero padding in the frequency domain. The CIRs are then normalized to unit energy. Next, in order to remove the initial propagation delay, we define the start of the CIRs. For L O S channels, we define the start of the CIR as the first M P C that arrives within 10 dB and 10 ns of the peak M P C . For N L O S channels, we define the start of the CIR as the first M P C that arrives within 10 dB and 50 ns of the peak M P C . We remove the propagation delay by setting the start time of the first arriving M P C to zero. After the initial delays have been removed, we align the spatial samples with respect to a given location and average them directly in the time domain to give the small-scale averaged power delay profile (APDP) . Finally, before any of the S-V model parameters are extracted, all M P C s with amplitudes that are more than 25 dB below the peak M P C are removed. This ensures that only significant M P C s are considered in the channel model. These procedures are based upon the recommendations contained in Appendix IV of the final report of the I E E E 802.15.4a channel modeling committee [15]. 4.3.2 Modeling Strategy Our first step is to decide which channel impulse model is the most suitable for the aircraft passenger cabin environment. Two standard U W B models were adopted by the 72 I E E E 802.15.3a and 4a task groups [1][2]. The sparse multi-cluster model is based upon the well known S-V model and is given by L K h (0=Z Z akj e x p (M,/ ) ^  (' - T / - ^ ) (4.1) ;=o *=o Here, the M P C s are modeled as Dirac delta functions, 8(.), and ak,i and 0*,/ are the amplitude and phase of the /th M P C in the M i cluster. L is the total number of clusters in the CIR and A' is the total number of rays within the /th cluster. T/ and xk,i represent the arrival time of the /th cluster and the /th M P C in the M i cluster, respectively. This model is used for residential and outdoor environments. The shape of the P D P is described as the product of two exponential functions, £ { | « t > / | 2 } <* exp(-T, / r )exp ( -T 4 > ; Iy) (4.2) where Y and y are the cluster and ray decay constants respectively. In the S-V model, the arrival times of the clusters are modeled by a Poisson distribution such that p ( r / | 7 ; _ 1 ) = A e x p [ - A ( r / - 7 ; . 1 ) ] (4.3) while the arrival times of the M P C s are modeled using a mixed Poisson distribution such that / ' ( ^ I V 0 , ' ) = ^ e x p B ( r ' . ' " r H . ' ) + ( l - / ? ) / l 2 e x p -^(T.J-T^J) (4.4) where the A , Ai and fa are the mean arrival rate of the clusters and M P C s within the P D P and P is the mixture probability. The dense single-cluster model depicted is used to describe dense scattering environment with a "soft" onset, e.g., the office and industrial environments under N L O S conditions. In these environments, one can no longer discern clustering within the CIR and the envelope of the P D P can be described as, E\\akj CC 1 - ^ e x p rise J exp - r k,l rx J (4.5) where x denotes the attenuation of the first component, y r i s e determines how fast the P D P rises to its local maximum, yi represents the decay at later times. If the scattering environment is sufficiently dense, e.g., an industrial N L O S environment, then every time resolution bin contains an M P C . Accordingly, the P D P can be modeled as a tapped delay 73 line with fixed arrival time, At, given by the inverse of the signal bandwidth. Where scatterers are less dense but the single cluster response still applies, e.g., an office N L O S environment, then the arrival rate of the M P C s is best modeled as a Poisson distribution. Although two models are currently being used by I E E E 802.15.4a task group, we have to stress that the models are only somewhat representative of the environment because they are only meant for making fair comparisons between the different system schemes and are not used for predicting performance. For example, a simplification made in the standard models is that all model parameters are derived from an average across the entire range of separation distances. Furthermore, the simplification has led to the use of a constant exponential decay rate to model the industrial N L O S environment from 2 to 8 m. The constant decay, which keeps the shape of P D P fixed, severely limits the range of R M S delay spread simulated by the channel model while many researchers have reported that the R M S delay spread, a measure of the time dispersion in the PDP, increases with distance [9][17][18]. In [17] and [19], the authors have further established the fact that the extended range of R M S delay spread is mainly caused by a change in the cluster decay rate, T, in the S V model or the exponential decay rate. Here, we seek to model the multipath characteristics more accurately by introducing the relationship between distance and the shape of the PDP. This strategy has also been used in [17]. Even though we feel that the industrial N L O S channel model is somewhat inadequate, it is still a good starting point because we have adopted its tapped delay line model for the ray arrival rates and its small-scale fading distribution. 4.3.3 Shape of the Power Delay Profile Figure 4.4 shows the A P D P of a L O S channel where the transmitting antenna is mounted on the ceiling and the receiving antenna is mounted on the headrest of a seat. A s in the case of industrial N L O S channels, there are no distinct clusters within the aircraft A P D P and the M P C s form a continuous exponential decay. The only exception is a few strong spikes or impulses early in the A P D P that likely correspond to specular reflection of some sort. A P D P s measured at other receiving antenna mounting positions that exhibit similar characteristics are shown in Figure 4.5 and Figure 4.6. The only difference between the different antenna mounting positions is the existence of a large spike (LOS 74 component) at the beginning of the PDP. Also , the A P D P s do not display the gentle rise to the maximum peak as described by the standard model in (4.5). Our proposed model of the P D P in aircraft passenger cabins is as follows. First, because there is no gentle rise to a maximum, we model the shape of the A P D P with just an exponential decay, (4.6) v r where y is the exponential decay rate. For L O S channels, we need to model the excess amplitude of the L O S path from the exponential decay. Excess amplitude is defined as Kr=- 5^  (4.7) 14 V k 1 t|LOS where PjqLos is the power of the L O S component and the denominator is the expected power at the beginning of the exponential decay described using (4.6). Using (4.6) and (4.7), we modeled the A P D P s for all three receiver scenarios and for all locations. Next, based on [17] and [19], we model how the shape of the P D P changes with distance where, y = Y0+Pr\ogd + er (4.8) and Kr=K0-\0BK\ogwd + eK (4.9) In (4.8) and (4.9), yo and KQ are intercepts and fiy and BK are the slopes. ey and EK are zero-mean Gaussian random variable with standard deviations, ay and OK, respectively, and d is the separation distance between the transmitting and receiving antennas. Using linear regression techniques after converting d into logarithmic scale, we estimate the parameters in (4.8) and (4.9). The results are shown in Figure 4.7 and Figure 4.9. Both shape parameters shown are changing with respect to distance in logarithmic scale. Also , from Figure 4.8 and Figure 4.10, we can see that both ey and EK fit well against a zero-mean normal distribution in logarithmic scale. A summary of all the model parameters is given in Table 4.1. 75 Table 4.1. Power delay profile model parameters Model Parameters LOS N L O S (headrest) (armrest, footrest) K0 24.18 N / A PK -0.35 N / A 1.32 N / A yo 15.08 12.79 Pi 5.00 7.94 ay 0.99 1.74 d 2 to 13 m 2 to 13 m 76 O r -10 • _ -20 Delay [ns] Figure 4.4. A P D P measured with receiving antenna mounted at the headrest of row 19. 77 O r -10 „ -20 _! I I ! I I I I 0 50 100 150 200 250 300 350 Delay [ns] Figure 4.6. A P D P measured with the receiving antenna mounted at the footrest of 19. 78 Distance [m] Figure 4.7. Excess amplitude of the L O S path, Kr, as a function of distance for L O S channels. -6 -4 -2 0 2 4 6 Deviation of Excess Amplitude of LOS Path, ^ [dB] Figure 4.8. Confirming the log-normality of the deviation of excess amplitude of L O S path, E K . 79 80 4.3.4 Interdependence of MPCs When designing systems that utilize spatial diversity, we need to model the dependence of correlation between antenna elements in the environment of interest so that we can realize the desired diversity gain with the minimum antenna spacing. A t the same time, we need to model temporal correlation so that we know how to model the M P C s that have arrived very close in time in the CIR. The correlation between signals measured at two spatial sampling points is defined as, /=r, where r(t, xt) and r(t, xj) are the received signals at grid points x, and x y, and d is the absolute distance between the 2 of 49 grid points [21]. Figure 4.11 shows the averaged correlation coefficient as a function of distance for different parts of the received signal. The averaging comes from calculating the correlation coefficient of all possible combinations of xt and Xj and taking the mean of the correlations when the separation distances are the same. From Figure 4.11, we can see that i f we only look at the first 13.3 ns, the correlation is much stronger and this is because of the few strong M P C s coming from a narrow angular spread. This result has an important implication for multi-antenna application and that is i f the receiver captures only the few dominant components, then a much larger antenna separation is needed to obtain low correlation. The temporal correlation is given by E{{ak-dk){ak+x-dk+x)} where E{.} denotes expectation, ak and ak+\ are the amplitude of the M i and (k + l)th M P C respectively, and ak and ak+\ are their mean values [13]. For all receiving antenna positions considered, the temporal correlation for the different delay taps are all below 0.4 with a mean correlation of only 0.13. Because the delay bins have low correlation, we assume them to be uncorrelated with each other and we can now simplify the channel model by assuming that the path amplitudes at each delay bin are independent random variables. 81 4.3.5 Small-Scale Fading To determine the distribution for small-scale fading, 49 amplitude data, \cik,i\, at a certain delay, TO, measured using spatial sampling are fit using the maximum likelihood estimation ( M L E ) against several probability distributions commonly used in propagation modeling, including lognormal, Nakagami, Rayleigh, Ricean, and Weibull . Figure 4.12 shows the empirical cumulative distribution function (CDF) and the fit against theoretical C D F s . In Figure 4.12, other than the lognormal distribution, all other distributions fit the empirical C D F equally well . From this result, we can conclude that Rayleigh distribution is the best distribution, and the reason all other distributions except for the lognormal distribution fit equally well is because their shape factors, e.g., the m-factor in Nakagami distribution, have transformed them into near replicas of the Rayleigh distribution. This result is different from [19] where a lognormal distribution is used to model the fading statistics. The Rayleigh distribution is not an unexpected outcome because in the aircraft passenger cabin environment, the Rayleigh distribution has been reported to be the small-scale fading statistics in environments with dense scatterers like the industrial environments [22]. Next, we use the Nakagami distribution to characterize the small-scale distribution at different delays because this distribution has the ability to model Rayleigh distribution as well as other types of propagation scenarios. Moreover, because the 4a task group has adopted Nakagami distribution for small-scale fading, expressing our results in terms of Nakagami distribution also allows fair comparisons with the other environments. The Nakagami distribution is given by the following equation, T(m) where m ^ Vi is the Nakagami m-factor, T(m) is the Gamma function, and Q is the mean-squared value of the amplitude. We estimate the m-factor of the Nakagami distribution using the inverse normalized variance estimator with respect to the 49 spatial samples collected for all delays [22]. The estimated m-factor is, m= ^ , (4.13) and 82 m n x2m-] exp m n -x (4.12) where N is the number of spatial sampling points and ht is the path amplitude. The scatter plot of estimated m-factors at different delays for the receiving antenna mounted on the headrest is shown in Figure 4.13. Although there are a few MPCs at the beginning of the PDP (typically when delay is less than 30 ns) that exhibit the Nakagami distribution, the Rayleigh distribution is still applicable for the majority of the PDP and for all antenna mounting positions in the passenger cabin. The estimated m-factors are found to follow a lognormal distribution by many researchers and the IEEE 802.15.4a task group [16][22]. The lognormal distribution is give by ft. ^ 2 ^ (4.15) fu(m) = ^7z= e x P (\nm-nm)2 lot where pm and am are the mean and variance of the m-factors on a natural logarithmic scale. Here, we also find that the w-factors follow a lognormal distribution. However, like the correlation coefficients, the fit to a lognormal distribution is not as accurate in the beginning of the PDP due to the few strong impulses near the leading edge of the response. The fit of w-factors to the lognormal distribution is shown in Figure 4.14. Although in [15], both jum and am are modeled as a function of delay, we did not find any evidence of a relationship. The initial spike in the PDP for the LOS channel is modeled using a fixed m-factor, mo, also given in natural logarithmic scale, and is typically much larger. A summary of the small-scale characteristics is presented in Table 4.2. Table 4.2. Small-scale fading parameters Receiving Antenna Mounting Mm 0~m m0 Position [dB] [dBl [dB] Headrest at row 4 0.055 0.286 4.03 Armrest at row 4 0.012 0.293 0.92 Footrest at row 4 0.083 0.267 2.13 Headrest at row 11 0.066 0.287 2.91 Armrest at row 11 0.092 0.284 2.87 Footrest at row 11 0.092 0.269 2.14 Headrest at row 19 0.101 0.301 2.23 Armrest at row 19 0.145 0.290 1.91 Footrest at row 19 0.187 0.318 1.98 83 c 92 'o IE CD O o c o JS 0 1 l_ o O tn CO o t_ O - 0 — Entire PDP - A— Before 13.3 ns - H — Af ter 13.3 ns 10 20 30 40 Distance between spatial sampling points [cm] Figure 4.11. Averaged spatial correlation as a function of antenna separation with the receiving antenna mounted on the footrest of row Z_i i i i i i ' I 0.01 0.02 0.03 0.04 0.05 0.06 0.07 Path Amplitude, |a | Figure 4.12. A fit of the path amplitudes against theoretical distributions. 84 10 Delay [ns] Figure 4.13. Estimated m-factors as a function of delay for receiving antenna mounted on the headrest of row 19. m-factors Figure 4.14. The lognormal fit of m-factors for the receiving antenna mounted at the footrest of row 19. 85 4.3.6 Delay Spread within a Local Area Mean excess delay, rmean, and R M S delay spread, Th™, are two important parameters that help to characterize the shape and duration of the P D P . In the case of simple digital modulation schemes, the ratio of T r m s to symbol period is also known to be strongly correlated with the bit error rate (BER) . In Table 4.3, we compare the mean excess delay and R M S delay spread within a local area. The relatively small standard deviations of the mean excess delay and R M S delay spread suggest that the channel is consistent within a local area. Table 4.3. Time dispersion parameters within a local area. Receiving Antenna Mounting Mean of x m e a n Std of x m e a n [ns] Mean of Std of [ns] Position [ns] [ns] Headrest at row 4 0.81 0.42 2.0 1.0 Armrest at row 4 5.0 2.51 5.1 1.57 Footrest at row 4 17.2 1.65 12.8 1.32 Headrest at row 11 0.83 0.62 1.94 1.22 Armrest at row 11 24.2 1.62 19.6 1.17 Footrest at row 11 31.0 1.71 22.19 1.24 Headrest at row 19 4.78 1.15 7.81 0.82 Armrest at row 19 29.04 2.12 21.75 1.16 Footrest at row 19 32.83 1.3 23.38 0.93 4.4 A Simulation Model for UWB CIR in Aircraft Passenger Cabin After modeling the U W B propagation for environments like residential, office, outdoor and industrial, 4a channel modeling subcommittee developed a simulation code that can generate CIRs typical in those environments. The M A T L A B - b a s e d code has four parts: (1) assignment of the channel model parameters, (2) generation of CIRs using random processes that simulates the arrivals of the clusters and rays and the path amplitudes based the shape of the P D P and small-scale fading distribution, (3) prediction of the frequency dependent path loss, and (4) conversion from continuous time to discrete time models. We have modified the channel simulation code developed by 4a so that it can generate channel impulse response typical of the aircraft passenger cabin 86 environment. First, we added two more sets of channel model parameters corresponding the L O S and N L O S channel conditions in the aircraft passenger cabin environment. In particular, we have added the parameters of small-scale fading. Then, in the part where CIRs are being simulated, instead of using the shape of the P D P as described in (6), we used a single exponential decay as described in (5) in the simulator and inserted codes that randomly generate the exponential decay rate and excess amplitude of the L O S path in (6) and (7) as a function of distance using (8) and (9). The modified version of the channel simulator can be downloaded from [23]. To validate the code, CIRs are generated using the modified channel simulator. Because our model depends on distance, distances from the aircraft measurements are used when we are regenerating the CIRs. A comparison of measured and simulated A P D P is shown in Figure 4.15. From the simulated CIRs, we also need to examine how closely the model can reproduce the C D F of the R M S delay spread. The comparisons between simulated and measured R M S delay spread are shown in Figure 4.16 and Figure 4.17. From Figure 4.16, we can see that the model has successfully reproduced the C D F of R M S delay spread. If we look at the scatter plot in Figure 4.17, we can see that the R M S delay spread is increasing as a function of separation distance between the transmitting and receiving antennas. Because the clusters of R M S delay spread represent the small-scale measurements we did for rows 4, 11 and 19, we can see that the range of R M S delay spread in the aircraft passenger cabin environment is much greater than what one might expect from within a local area. 87 CO 2, D_ Q CL < •o 0 L_ 3 in co 01 CD 73. Q. Q CL < T3 0 -4—' J9 E CO 200 100 Delay [ns] Figure 4.15. Comparison of the measured and regenerated A P D P . 200 XI CO o 0 .a E o 20 30 40 RMS Delay Spread [ns] Figure 4.16. Comparisons of the distribution of the simulated and measured R M S delay spread for L O S and N L O S channels. 88 CO c co (D CL co >. JS a> Q CO or CO .g. T3 CO d) i_ Q. CO >. is 0) Q CO 30 25 20 15 10 30 25 20 15 10 1 • NLOS simulation NLOS Fit 3 4 5 6 7 8 910 Distance [m] • NLOS Measred NLOS Fit 2 3 4 5 6 7 8 9 1 0 Distance [m] Figure 4.17. Comparisons of the R M S delay spread with respect to distance for N L O S channels. 89 4.5 Conclusions We have measured a multiplicity of U W B channel impulse responses within the passenger cabin of a typical mid-size airliner for both point-to-multipoint and peer-to-peer configurations. Based upon analysis of the results, we have proposed a statistical model that describes the multipath characteristics of the channel including the shape of the power delay profile and both the spatial and temporal distribution of the small-scale fading. We have observed several noteworthy aspects of U W B channel impulse responses within the passenger cabin of a mid-size airliner: (1) the shape of the P D P generally follows I E E E 802.15.4a's dense single-cluster model, but with negligible rise time and, on many occasions, one or more impulses or spikes near the leading edge of the response, (2) delay spread tends to increase with path length and as the receiving antenna drops from the headrest to the footrest, (3) small-scale fading of multipath components (MPCs) tends to follow a Nakagami distribution with a lognormally-distributed m-parameter, as has been found in other environments, and (4) the mean excess delay and rms delay spread within a local area (several wavelengths) are consistent. In most cases, our results take the form of the parameters of the corresponding models recommended by the I E E E 802.15.4a channel modelling committee and can be used directly in simulations of U W B propagation in an aircraft interior. Accordingly, our results w i l l assist: (1) those planning U W B deployments and field trials in aircraft, (2) those wishing to verify the results of electromagnetic simulations of aircraft interiors, (3) those wishing to simulate U W B aircraft systems with realistic channels. 90 References [1] A . F. Mol isch, J. R. Foerster and M . Pendergrass, "Channel models for ultrawideband personal area networks," IEEE Wireless Commun., vol . 10, pp. 14-21, Dec. 2003. [2] A . F. Mol isch et al., " A comprehensive standardized model for ultrawideband propagation channels," IEEE Trans. Antennas Propag., vol . 54, no. 11, pp. 3151 -3165, Nov. 2006. [3] N . R . Diaz and M . Holzbock, "Aircraft cabin propagation for multimedia communications," Proc. EMPS 2002, 25-26 Sep. 2002. [4] M . Youssef and L . Vahala, "Effects of passengers and internal components on electromagnetic propagation prediction inside Boeing aircrafts," 2006 IEEE AP-S Int. Symp. Dig, pp. 2161-2164, 9-14 Jul. 2006. [5] M . Jafri, J. E ly and L . Vahala, "Comparative analysis of interference path loss coupling patterns on B-737 vs B-757 airplanes," Digital Avionics Systems Conference, vol . 1, pp. 6.B.5 - 61 - 10, 30 Oct. - 3 Nov. 2005. [6] G . Hankins, L . Vahala and J. H . Beggs, "Propagation prediction inside a B767 in the 2.4 G H z and 5 G H z radio bands," 2005 IEEE AP-S Int. Symp. Dig, vol . 1A, pp. 791-794, 3-8 Jul. 2005. [7] C P . Niebla, "Coverage and capacity planning for aircraft in-cabin wireless heterogeneous network," Proc. IEEE VTC 2003-Fall, pp. 1658-1662, 6-9 Oct 2003. [8] G . A . Breit, H . Hachem, J. Forrester, P. Guckian, K . P . Kirchoff, B . J . Donham, " R F propagation characteristics of in-cabin C D M A mobile phone networks," Digital Avionics Systems Conference, pp. 9.C.5-1—9.C.5-12, 30 Oct. - 3 Nov. 2005. [9] N . R . Diaz and J.E.J. Esquitino, "Wideband channel characterization for wireless communications inside a short haul aircraft," Proc. IEEE VTC 2004 - Spring, pp. 223-228, 17-19 M a y 2004. [10] R. Bhagavatula, R. W . Heath and S. Vishwanath, "Optimizing M I M O antenna placement and array configuration for multimedia delivery in aircraft," Proc. IEEE VTC 2007 - Spring, pp. 425-429, Apr. 2007. 91 [11] J. Chuang, N . X i n , H . Huang, S. Chiu and D . G . Michelson, " U W B radiowave propagation within the passenger cabin of a Boeing 737-200 aircraft," Proc. IEEE VTC 2007 - Spring, pp. 496-500, Apr. 2007. [12] J. Chuang and D . G . Michelson, "Large-scale aspects of U W B propagation within the passenger cabin of a Boeing 737-200 aircraft," Submitted to IEEE Wireless Commun., 15 Aug . 2007. [13] C. C. Chong, S. K . Yong, " A generic statistical-based uwb channel model for high-rise apartments," IEEE Trans. Antennas Propagat., vol . 53, no. 8, pp. 2389-2399, Aug . 2005. [14] A . F . Mol isch, "Ultrawideband propagation channels: Theory, measurement, and modeling", IEEE Trans. Veh. Technol., vol . 54, no.5, pp. 1528-1545, Sep. 2005. [15] A . F. Mol isch et al., " I E E E 802.15.4a channel Model - Final report," Tech. Rep. Doc. I E E E 802.15-04-0662-02-004a, 2005. [16] C. W . K i m , X . Sun, L . C. Chiam, B . Kannan, F. P. S. Chin, and H . K . Garg, "Characterization of ultra-wideband channels for outdoor office environment," IEEE Wireless Commun. Networking Conf, vol . 2, pp. 950-955, 13-17 Mar. 2005. [17] J. A . Dabin, A . M . Maimovich and H . Grebel, " A statistical ultra-wideband indoor channel model and the effects of antenna directivity on path loss and multipath propagation," IEEEJ. Sel. Areas Commun., vol . 24, no. 4, pp. 752-758, Apr . 2006. [18] S. S. Ghassemzadeh, R. Jana, C. W . Rice, W . Turin, and V . Tarokh, "Measurement and modeling of an ultra-wide bandwidth indoor channel," IEEE Trans. Commun., vol . 52, no. 10, pp. 1786-1796, Oct. 2004. [19] S. S. Ghassemzadeh, L . J. Greenstein, T. Sveinsson, A . Kavcic and V . Tarokh, " U W B delay profile models for residential and commercial indoor environments," IEEE Trans. Veh. Technol, vol . 54, no. 4, pp. 1235-1244, Jul. 2005, [20] L . J. Greenstein, S. S. Ghassemzadeh, S-. C. Hong and V . Tarokh, "Comparison study of U W B indoor channel models," IEEE Trans. Wireless Commun., vol . 6, no. l , p p . 128-135, Jan. 2007. [21] B . Al l en et al, Ultra-wideband Antennas and Propagation, Wiley, 2007. [22] J. Karedal, S. Wyne, P. Aimers, F. Tufvesson, and A . F. Mol isch, "Statistical analysis of the U W B channel in an industrial environment," Proc. IEEE VTC 2004 - Fall, vol . 1, pp. 81-85, 26-29 Sept. 2004. [23] J. Chuang and D . G . Michelson, "Simulation Code for U W B CIRs in Passenger Aircraft," U B C Radio Science Lab, Aug . 2007, http://rsi. ece. ubc. ca/codes/'aircraft, html 93 Chapter 5 Conclusions and Recommendations 5.1 Conclusions This thesis has been concerned with the identification of clusters within ultrawideband ( U W B ) channel impulse responses (CIR) and the U W B propagation characteristics within the passenger cabin of a Boeing 737-200 aircraft. In this thesis, we have made three major contributions to the modeling of U W B propagation. First, we developed an automated cluster identification algorithm which makes the identification of clusters more consistent and less subjective. The algorithm also determines how a U W B CIR can be most effectively represented by either the single exponentially decaying cluster model or the standard Saleh-Valenzuela (S-V) model. Compared to past work, ours is innovative in four ways: (1) we implemented a novel significant M P C search algorithm that reduces small-scale fading that is inherent in channel measurements, (2) our bounded search drastically reduced the number of searches required and made the algorithm more tractable, (3) only five input parameters are needed which is far less complicated than schemes proposed previously and (4) the iterative nature of the algorithm and the manner in which the cluster selection rules are layered makes it possible to be used either as an autonomous tools or an interactive aid for the analysts. Furthermore, The validity of our approach was verified using both the simulated U W B CIRs generated from the simulation code developed I E E E 802.15.4a task group and the measured CIRs from office and underground mine environments. Second, we characterized large-scale aspect of U W B propagation within the passenger cabin of a Boeing 737-200 aircraft which includes distance and frequency dependent path loss and time dispersion parameters such as R M S delay spread and number of significant M P C s . From the measurements, we determined that the passenger of a Boeing 737-200 aircraft is unique from the other conventional environments in three ways. (1) The coverage within the passenger cabin is found to follow a chevron shape contour where the coverage is the greatest along the aisle and the weakest around 94 window seats. This result suggests that the path gain within the passenger cabin depends not only on the separation distance but also on the seat location and the mounting point. (2) When antenna effects are not as strong, e.g., in p-to-p configurations, frequency dependence of path loss is found to decrease with the square of frequency, (3) x r m s is found to vary greatly from 3 to 29 ns, with respect to the L O S and N L O S channel conditions associated with the different antenna mounting points and (4) the presence of human increased the path loss and decreased delay spread. Third, based on extensive measurements, we developed statistical models that describe the small-scale aspects of U W B propagation within the passenger cabin of a Boeing 737-200 aircraft. Specifically, we looked at two major aspects: the shape of the power delay profile (PDP), and the distribution of small-scale fading. Four noteworthy aspects were discovered: (1) the shape of the P D P generally follows the dense single-cluster model used to describe the channels like the industrial non-line-of-sight (NLOS) environment^ (2) delay spread tends to increase with path length and as the receiving antenna lowers, (3) small-scale fading can be modeled using Nakagami distribution with a lognormally-distributed m-factor, as found in other environments and (4) the mean excess delay and rms delay spread within a local area (several wavelengths) are fairly consistent. In most cases, our model follows the form of the parameters of the corresponding channel models recommended by I E E E 802.15.4a and can be used directly in simulations. B y developing the automated cluster identification algorithm in Chapter 2, we have resolved a major issue regarding the U W B channel modeling process. In Chapter 3 and 4, the measurements and modeling of U W B propagation in the passenger cabin of an aircraft is an essential first step in the modeling of propagation in the public transportation environment. The results are going to help (1) those planning U W B deployments and field trials in aircraft, (2) those wishing to verify the results of electromagnetic simulations of aircraft interiors, (3) those wishing to simulate U W B aircraft systems with realistic channels, and (4) those wanting to understand the effect of interference from other wireless devices aboard the aircraft. 95 5.2 Recommendations for Further Work While this study has demonstrated the practical issues in identification of clusters in U W B channel modeling and the modeling of the U W B propagation in the aircraft environment, there are still many limitations to our work and we recommend that several topics of considerable practical interest be pursued in the very near future. In the automated cluster identification algorithm, we recommend adding smoothing techniques, such as applying a running average, used in image processing. In channel measurements, we feel that more measurements are needed still to fully characterize the aircraft environment. Due to time constraints and other administration issues, only limited data were collected to characterize the effect of human presence. Nevertheless, we have demonstrated that the effect of human presence is important and should be examined with great scrutiny. More importantly, it should be extended to the small-scale aspects of U W B propagation. Finally, in order to investigate the commonality between other types of vehicles used in public transportation, measurement campaigns should be carried out for other types of aircrafts as well as other types of vehicles. 96 Appendix A Useful Tables Table A . 1 . Dimensions of modern aircraft. Manufacturer Aircraft Family Recent model Cabin Width Cabin Height Boeing 737 737-900 3.54 m 2.20 m Boeing 757 757-300 3.54 m 2.20 m A C A C ARJ21 ARJ21-700 3.14m 2.03 m AirBus A320 A321 3.7 m 2.22 m Bombardier CRJ CRJ 1000 2.53 m 1.85 m Embraer E-Jet E-175 2.74 m 2 m Table A . 2 . Large-scale parameters of conventional environments. Environment PG0 n Xa K [dBm] [dB] CM1 Residential-LOS -43.9 1.79 2.22 1.12±0.12 CM2 Residential-NLOS -48.7 4.58 3.51 1.53±0.32 CM3 Office-LOS -35.4 1.63 1.9 0.03 CM4 Office-NLOS -57.9 3.07 3.9 0.71 CM5 Outdoor-LOS -45.6 1.76 0.83 0.12 CM6 Outdoor-NLOS -73.0 2.5 2 0.13 CM7 Industrial-LOS -56.7 1.2 6 -1.103 CM8 Industrial-NLOS -56.7 2.15 6 -1.427 Table A . 3 . Time dispersion parameters of conventional environments. Environment tfnean trms NP 10dB [ns] [ns] CM1 Residential-LOS 16.36 17.75 15.28 CM2 Residential-NLOS 19.68 19.09 35.04 CM3 Office-LOS 8.816 10 20.14 CM4 Office-NLOS 17.49 13.25 54.84 CM5 Outdoor-LOS 26.31 30.24 23.16 CM6 Outdoor-NLOS 70.94 75.4 30.44 CM7 Industrial-LOS 5.55 8.65 15.94 CM8 Industrial-NLOS 114.3 90 355.4 97 Appendix B Detailed Setup of E8362B PNA Figure B . l shows the connections of the U W B channel sounder used to measure the U W B channel in the passenger cabin of a Boeing 737-200 aircraft. The P N A is set to work in the forward operation, i.e, to measure the S-parameter, S21, by taking the ratio of receiver B ' s measurement over receiver R l ' s measurement. B y utilizing option 014 on the P N A , we avoided entering the directional coupler across the ports i and j , and went directly into receiver B . The bypass gave us an extra 12 dB gain. A detailed description of each port on the P N A is given in Table B . l . The settings of the P N A are summarized in Table B.2. The extra gain allowed us to increase the coverage up to 15 metres with at least 10 dB of signal-to-noise ratio. The link budget of the channel sounder is shown in Table B.3 . Table B . l . E8362B P N A option 014 port description. Ports Description a Source output b Receiver R l input c Source output d Directional coupler through e Port 1 f Directional coupler arm g Receiver A input h Receiver B input i Directional coupler arm j Port 2 k Directional coupler through 1 Source output m Receiver R2 input n Source output 98 Table B.2. Setting of E8362B P N A . PNA Settings Values Start Frequency 3.0 GHz Stop Frequency 10.6 GHz IF Bandwidth 3 kHz Number of Sampling Points 6401 Input Power 5 dBm Table B.3 . L ink budget of the U W B channel sounder Links Values Transmitted Power 5 dBm Cable Loss -9.1 dB* Transmit Antenna gain OdBi Path Loss -82.1 dB** Receiver Antenna gain OdBi Cable Loss -9.1 dB* Received Power -95.3 dBm Receiver Sensitivity -107.2 dBm System Margin 11.9 dB •Calculated from data sheet at the highest frequency 10.6 GHz **Path loss is calculated using Friis formula with an exponent of 2.25 at a range of 15m. 99 E8362B Performance Network Analyzer 15mLMR-400 Coaxial Cable 15 m LMR-400 Coaxial Cable ^ Channel EM6865 UWB Omni direction Bicone Antenna EM6865 UWB Omni direction Bicone Antenna Figure B . l . Connections of the U W B channel sounder. 100 Appendix C Calibration of Antennas In order to calibrate for the effect of antennas, the angle of arrival (AoA) of multipath components ( M P C ) is needed. To determine the A o A , virtual or real antenna array are often used. However, due to the limited space available in the passenger cabin, it is too difficult to deploy such an array. We have, instead, decided to use an approximate approach to calibrate the effect of the antenna. First, we have to assume that the M P C s are arriving uniformly from all directions. Next, we measure the antenna transfer function (ATF) with respect to all orientations of the receiving antenna in open space. The measured A T F s are then averaged to form the averaged A T F . This averaged A T F is then deconvolved from the measured channel transfer function (CTF) directly in the frequency domain. Figure C . l shows a picture of the measurement setup used to measure the A T F s . The measurements were done on the roof top of an L-shaped building with the antennas mounted on a 3.4 metres high tripod to minimize reflections from the environments. The ground reflection is also minimized by placing the antennas along the two wings of the building, and as a result, the antennas are 14.4 metres apart. The measured averaged A T F is shown in Figure C.2. In total 24 A T F s were taken with respect to the different elevation angles over the range from 0 to 345 degrees with 15 degree increments. Also , 12 A T F s were taken with respect to the different azimuth angles over the range from 0 to 330 degrees with 30 degree increments. Since there are very little cancellations in the A T F s , we can conclude that the antennas are placed in a nearly free space environment and thus, only contain the effect of the antennas. Note that, the averaged A T F is showing the typical response of an receiving antenna with constant gain where the effective aperture of the antenna is decreasing with increasing frequency. The decrease in effective aperture leads to the decrease in the received power with increasing frequency. 101 Figure C . l . Measurement setup of the antenna transfer functions. Frequency [GHz] Figure C.2. Averaged antenna transfer functions. 102 Appendix D Detailed Measurement Plan ro • • • (a) (b) (c) Figure D . l . Detailed measurement plan of p-to-mp configuration for (a) empty aircraft, (b) half full aircraft and (c) full aircraft. (A = transmitting antenna, O = receiving antenna, • = passengers) 103 C D P' C E O Figure D.2. Detailed measurement plan of p-to-p configuration. (A = transmitting antenna, O = receiving antenna) 104 Appendix E Matlab Code of the Automated Cluster Identification Algorithm % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % 0 / o 0 / o % main.m - begin %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% env = 'demo'; load('demo'); h = h_demo; % parameters of the algorithm Nloc = length(h(l,:)); % number of CIR from 'h' zone = 7; % local maxima search zone [samples] sep = 8; % maximum separation time before two local maxima are considered as two clusters [ns] power_diff = 2; % power difference of old and new clusters [dB] minclusterlength = 3; % this is the length with respect to how many local maxima min_error_threshold = 1; %2.7 % minimum error threshold [dB] sig_peak = 0.1;% when finding local maxima, the local maxima have to be greater than the rest by [dB] % othe parameters of the algorithm maxnumclusters = 100; % a fail-safe stopping condition for the while loop minerrorreduction = 0.01; % stopping condition when increasing the number of clusters does not significantly decrease the residual error [dB] (normally not used) max_delay = 2500; % maximum delay, rays beyond this delay are not considered [ns] add_ruleCheck = 0; % set to zero if you need to enforce the additional rules. add_boundaryCheck = 0; % set to zero if you need to check the boundaries of the clusters, saveresults = 0; % set to zero i f you do not want to save the results in a workspace. step_by_step = 0; % set to 1 if you want to see the evolution of cluster ID process, showresult = 1; % for output graph control enable_axis_control = 1; xmin = 0; xmax = 350; ymin = -80; ymax = 0; for(i= l:Nloc) threshold = max(20*logl0(abs(h(:,i)))) - 100; noise_floor(i) = threshold; % noise floor, rays below this value are not considered end; % loop through all the CIRs for(i= l:Nloc) No_of_CIR = i % put the CIR into the algorithm [bp Nbp pdp lp t_lp Carr Camp Cnum Rarr Ramp Rnum Rbp Rarrindex... error_vector] = clusterAlgorithmTest(h(:,i), t, zone, sep, power_diff,... 105 max_num_clusters, .min_error_threshold, min_cluster_length, min_error_reduction, sig_peak,. noise_floor(i), max_delay, add_ruleCheck, add_boundaryCheck, step_by_step); % for plotting purposes only pdp = 20*logl0(abs(h(:,i))); if(show_result == 1) % show identification result figure(4); elf; hold on; plot(t, pdp); plot(Carr_sim_plot, Camp_sim_plot, 'r.1, 'MarkerSize', 16); plot(t_lp(bp), pdp_lp(bp), 'ms', 'MarkerSize1, 8); for(k= l:Nbp) start = bp(k); if(k == Nbp) stop = length(pdplp); else stop = bp(k+l)-l; end; coef = polyfit(t_lp(start:stop), pdp_lp(start:stop), 1); line = coef(l)*t_lp(start:stop) + coef(2); plot(t_lp(start:stop), line, 'k', 'linewidth', 1.5); end; hold off; xlabel(Time [ns]'); ylabel('PDP [dB]'); % title([env,' -', int2str(i)]); if(enable_axis_control == 1) axis([xmin xmax ymin ymax]); end; % show RMS error figure(5); for(nz= l:length(error_vector)) if(error_vector(nz) ~= 0) break; end; end; nz hold on; plot(nz:length(error_vector), error_vector(nz:length(error_vector)), '-v'); xlabel('Number of Identified Clusters'); ylabel('RMS Error for entire PDP [dB]'); axis([0 14 0 20]); end; % save them all into a big workspace if(save_results == 1) Cnum_all(i) = Cnum; if(i==l) Carr_all(:,l) = Carr'; Camp_all(:,l) = Camp'; else if(length(Carr_all(:,l)) > Cnum) Carr_all(l:Cnum,i) = Carr; 106 else Carr_all = [Carr_all; zeros(Cnum-length(Carr_all(:,l)),i-l);]; Carr_all(:,i) = Carr'; end; if(length(Camp_all(:, 1)) > Cnum) Camp_all(l :Cnum,i) = Camp; else Campal l = [Camp_all; zeros(Cnum-length(Camp_all(:,l)),i-l);]; Camp_all(:,i) = Camp'; end; end; save([env, 'dervied'], 'Cnum_all', 'Carr_all', 'Cnumall '); end; end; %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% main.m - end %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % 0 / o % % % % % cluster iteration.m - begin %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% function [bp Nbp pdp_lp t i p Carr Camp Cnum Rarr Ramp Rnum Rbp Rarrindex error_vector]... = clusterAlgorithmTest(h, t, zone, sep, power_diff, max_num_clusters,... min_error_threshold, minclusterlength, min_error_reduction, sig_peak, noisefloor,... maxdelay, .addruleCheck, add_boundaryCheck, stepbystep); % this function is used the incrementing the number of clusters in the % automated cluster identification algorithm. At the end of this script, % the identified clusters and major rays along with their arrival time and % amplitude will be saved in a workspace. % INPUTS % h —> CIR in linear voltage units % t --> time axis in [ns] % zone —> local maxima search zone in [samples] % sep —> maximum separation time % power_diff --> power difference of new and old clusters % dip --> not used % maximum_num_clusters --> safety parameter so the loop won't run forever % min_error_threshold —> when the algorithm stops incrementing the number of clusters [dB] % noise_floor —> rays that drops below this threshold will not be used [dB] % max_delay —> rays that are beyond this delay value will not be used [ns] % error_weight --> weight used to change the relative importance of postive residue and negative residue % OUTPUTS % bp --> the breakpoints in the power delay profile or the start/end of clusters (used for generating plots) % Nbp —> the number of breakpoints (used for generating plots) % pdp_lp —> power delay profile of identified local maxima (used for generating plots) % t_lp —> time of the identified local maxima (used for generating plots) % Carr —> cluster arrival time [ns] % Cnum --> number of clusters (Nbp + 1) % Camp --> cluster amplitude, strongest ray within a cluster [dB] % Rarr --> ray arrival time [ns] % Ramp --> ray amplitude [dB] % Rnum --> number of rays within a cluster 107 % Rbp --> index of the breakpoints [samples] % Rarrjndex —> number of rays in each cluster % errorvector —> how error decreases as number of clusters increase % illegalcluster —> status flag showing clusters that have not passed the additional rules Ndata = length(h); % convert to pdp pdp = abs(h).A2; pdp = 10*logl0(pdp); % find local maxima index = 1; for(i = 1 :Ndata - zone - 1); if(i <= zone) start = 1; else start = i-zone; end; % conditions to find local maximum include: % 1. It needs to be the maximum within the search zone. % 2. It needs to be greater than the noise floor. % 3. It needs to arrive before the delay cutoff. % 4. the maximum needs to be greater than the second greatest point by 0.5 dB. zone_max = max(pdp(start:i+zone)); if(pdp(i) = zonemax & pdp(i) > noise_floor & t(i) < maxdelay & ... length(find(pdp(start: i+zone)>zone_max-sig_peak))== 1); pdp_lp(index) = pdp(i); tlp(index) = t(i); index_lp(index) = i ; index = index + 1; end; end; Nip = length(pdp_lp); % remove any isolated points (first and last) if(t_lp(Nlp)-t_lp(Nlp-l) > sep) pdp lp = pdp_lp(l :Nlp-l); t_lp = t_lp(l:Nlp-l); indexlp = index_lp(l:Nlp-l); Nip = length(pdplp); end; if(t_lp(2)-t_lp(l)>sep) pdp_lp = pdp_lp(2:Nlp); t_lp = t_lp(2:Nlp); index_lp = index_lp(l:Nlp-l); Nip = length(pdp_lp); end; % remove any isolated points (in the middle) index = 1; remove = 0; for(i = 2:Nlp-l) if((t_lp(i)-t_lpO-1 )>sep)&(t_lp(i+1 )-t_lp(i))>sep) remove(index) = i ; index = index + 1; 108 end; end; iffremove ~= 0) for(i = l:length(remove)) pdp_lp = [pdp_lp(l:remove(i)-l) pdp_lp(remove(i)+l:length(pdp_lp))]; U P = [t_lp(l:remove(i)-l) t_lp(remove(i)+l:length(t_lp))]; if(i ~= length(remove)) % reset the index remove(i+l:length(remove)) = remove((i+l):length(remove)) - 1; end; end; Nip = length(pdplp); end; % find the smallest RMS error (the iteration starts here) % initialization N c = 1; old_bp = 1; % for all number of clusters, 1 is always present and it is the start of the first cluster all_cluster_error_below_threshold = 0; % check to see if there are large gaps. If there are automatically % increase the number of clusters according to the number of gaps + 1. for(i = 3:Nlp-l) if(t_lp(i+l)-t_lp(i) > sep) % if there is large gap old_bp = [old_bp i+1]; % then record the breakpoint Nc = Nc + 1; % incrememt the number of clusters number_of_clusters = Nc % msg in console end; end; % run the RMS error test for the first time to determine if we actually % want to go into the while loop next iffNc == 1) % calculate the RMS error for the case of one cluster just in case one cluster is the best fit coef = polyfit(t_lp, pdplp, 1); % considering everything as one cluster and fit a line through it line = polyval(coef, t i p ) ; min_error = sqrt(mean((pdp_lp - line).A2)); % find the variance if(min_error < min_error_threshold) all_cluster_error_below_threshold = 1; bp = old_bp; end; error_vector = min_error; else m i n e r r o r c luster = zeros(l, Nc); N l p c luster = zeros(l, Nc); min_error = 0; for(k= l:Nc) start = old_bp(k); if(k = Nc) stop = Nip; else stop = old_bp(k+l)-l; end; coef = polyfit(t_lp(start:stop), pdplp(startistop), 1); % least square fit min_error_gamma_cluster(k) = -10*logl0(exp(l))/coef(l); line = polyval(coef, t_lp(start:stop)); min_error = min_error + sum((pdp_lp(start:stop)-line).A2); min_error_cluster(k) = minerrorcluster(k) + sqrt(mean((pdp_lp(start:stop)-line).A2)); 109 Nlp_cluster(k) = stop-start+1; end; % min_error_gamma_cluster mean_min_error_gamma_cluster = mean(min_error_gamma_cluster) minerror = sqrt(min_error/Nlp); if(isempty(find(min_error_cluster > minerrorthreshold))) all_cluster_error_below_threshold = 1; else all_cluster_error_below_threshold = 0; end; error_vector = zeros(l, Nc); error_vector(Nc) = minerror; if(all_cluster_error_below_threshold == 1) bp = old_bp; end; end; Nip = Nip % minerrorcluster % while(min_error > min_error_threshold | all_cluster_error_below_threshold == 1) while(all_cluster_error_below_threshold == 0) % find the smallest RMS error given the number of clusters, Nc [bp min_error min_error_cluster] = clusterRMSerrorTest(pdp_lp, t i p , Nc, oldbp, . . . min_error_threshold, min_cluster_length); % check to see i f the algorithm failed or not if(bp==0) stop ins^ ~~ ' ' ^ ^ ^ ^ ' ^ ' ^ ' i ' ' ^ ' ^ ' ^ ' ^ ' ^ ' ^ new t)re<iks C3.ii be found ^^^^^^^^^^^^t*^'^'^^E^e'^^e^£'^^'^t5^'^^!^' bp = o ldbp; break; end; % min_error_cluster % update the search - increment the number of clusters to reflect the current number of clusters % pass on breakpoint info to the next iteration. Nc = N c + 1; old_bp = bp; % message in the console numberofclusters = Nc % save the changes in the min_error as the number of assumed clusters % increase error_vector(Nc) = min_error; % check to see if the R M S error of each cluster is below the threshold if(isempty(find(min_error_cluster > min_error_threshold))) all_cluster_error_below_threshold = 1; else all_cluster_error_below_threshold = 0; end; % stopping condition in case the algorithm tries to break into too many 110 % clusers when there are not that many local maxima iffNc > length(pdp_lp)/3) stop_msg = '***• Number of clusters greater than significant paths can provide ***** break; end; % stopping condition in case something goes wrong if(Nc > max_num_clusters) stopmsg =***** Number of clusters greater than the user defined maximum number break; end; % used to show the step by step breakdown of the cluster identification % process if(step_by_step == 1) figure(15); elf; hold on; plot(t, pdp); plot(t_lp, pdpjp, W , 'MarkerSize', 4); plot(t_lp(bp), pdp_lp(bp), 'ms', 'MarkerSize', 8); for(k= l:Nc) start = bp(k); if(k == Nc) stop = length(pdp_lp); else stop = bp(k+l)-l; end; coef = polyfit(t_lp(start:stop), pdp_lp(start:stop), 1); line = coef(l)*t_lp(start:stop) + coef(2); plot(t_lp(start:stop), line, 'k', 'linewidth', 1.5); end; hold off; xlabel(Time [ns]'); ylabel('PDP [dB]'); title([int2str(Nc),' Clusters']); axis([0 max(t_lp)+10 min(pdp) max(pdp)+5]); pause; end; end; % message in the console % number_of_clusters = Nc systemmsg = '*********** Cluster identification process completed ***************** % save the number of breakpoints into a variable Nbp = length(bp); % apply additional rules if(add_ruleCheck== 1) [bp] = ruleCheck(pdp_lp, t_lp, bp, power_diff, sep, 1,1); systemmsg = '********** Additional rules successfully applied ****************** end; if(add_boundaryCheck == 1) 111 [bp] = clusterBoundaryCheck(pdp_lp, t_lp, bp); end; Nbp = length(bp); % update the number of breakpoints just in case % returning 4a parameters % stats for the clusters Carr = t_lp(bp); % absolute cluster arrival time = where the breakpoints are for(i= l:Nbp); start = bp(i); if(i == Nbp) stop = Nip; else stop = bp(i+l)-l; end; Camp(i) = max(pdp_lp(start:stop)); % cluster amplitude = strongest path within a cluster end; Cnum = Nbp; % number of clusters % stats for the rays Rbp = indexlp(bp); % reset the start of the clusters as the all rays not just the local maxima Rarrjndex = ones(l, Cnum); Rarr = zeros(length(t), Cnum); Ramp = zeros(length(t), Cnum); for(k = 1 :Cnum) start = Rbp(k); if(k == Cnum) stop = length(pdp)-l; else stop = Rbp(k+l)-l; end; for(m = start: stop) if(m == 1) % check right only for the first point if(pdp(m) > pdp(m + 1)) Rarr(Rarr_index(k), k) = t(m); Ramp(Rarr_index(k), k) = pdp(m); Rarrindex(k) = Rarr_index(k) + 1; end; else % check left and right for all points if((pdp(m)-pdp(m + 1))>0.5 & (pdp(m)-pdp(m - l))>sig_peak & pdp(m) > noise_floor & t(m) <... max_delay) Rarr(Rarr_index(k), k) = t(m); Ramp(Rarr_index(k), k) = pdp(m); Rarrindex(k) = Rarrjndex(k) + 1; end; end; end; Rarrindex(k) = Rarr_index(k) - 1; % you've added one more than needed end; Rnum = sum(Rarr_index); % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % 0 / o % % cluster iteration.m - end %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 0^0 /OZO/0 /0 /0 /0 /0 /0 /0^0 /0^0 /0 /0 / / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o / o 112 cluster RMS error.m - begin %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% function [bp minerror minerrorcluster] = clusterRMSerrorTest(pdp_lp, t i p , Nc, oldbp, . . . min_error_threshold, min_cluster_length); % this function is used for iterating through all combinations of straight % lines given the number of clusters. It will find what the min error is % and pass it back to clusterAlgorithmTest to see i f incrementing the % number of clusters is still necessary. % INPUT % pdp lp —> local maxima identified in the PDP % t_lp —> time of the local maxima identified in the PDP % Nc —> number of clusters for this iteration % old_bp -> breakpoint identified w/ one less cluster % min_error_threshold --> minimum error threshold [dB] % min_cluster_length --> minimum length of a cluster w.r.t. the local maxima % OUTPUT % bp —> breakpoints or the start of a cluster % min_error —> min RMS error found [dB] % min_error_cluster --> min RMS error for each cluster found [dB] Nip = length(pdp_lp); min_error = inf; min_error_cluster = zeros(l, Nc+1); algorithm_fail = 0; % status flag % pre-calculate the residuals of errors to save computations error_cluster_old = zeros(l, Nc); for(k= l:Nc) start = oldbp(k); if(k == Nc) stop = Nip; else stop = old_bp(k+l)-l; end; coef = polyfit(t_lp(start:stop), pdp_lp(start:stop), 1); % least square fit gammaclusterold(k) = -10*logl0(exp(l))/coef(l); line = polyval(coef, t_lp(start:stop)); error_cluster_old(k) = sum((pdplp(start:stop)-Iine).A2); Nlp_cluster_old(k) = stop-start+1; rms_error_cluster_old(k) = sqrt(error_cluster_old(k)/Nlp_cluster_old(k)); % calculate the R M S error end; for(i = 3:Nlp-l) newbp = zeros(l, Nc+1); % initialize a fresh set of breakpoints error = 0; % error tracker - this is updated per combintation of fit and saved as min_error errorcluster = zeros(l, Nc+1); % error tracker for individual clusters Nlp_cluster = zeros(l, Nc+1); % cluster length tracker % checking to see i f any of the old breakpoints are where % this new iterating breakpoint is conflict = 0; for(k= l:length(old_bp)) 113 if(abs(old_bp(k)-i) <= min_cluster_length) % if it is within close to the +/- 1 of the old breakpoint conflict = 1; break; end; end; % proceed w/ the RMS error calculation if(conflict == 0) newbp = [oldbp i]; % add the new breakpoint at the end newbp = sort(new_bp); % sort it in ascending order new_bp_i = find(new_bp == i); % find where the new index is errorstatus = 0; % check to see if this cluster is already below the error threshold if(rms_error_cluster_old(new_bp_i-l) < minerrorthreshold) error_status = 1; end; if(error_status = 0) % proceed only i f the breakpoint is at a cluster with error greater than threshold % add up the errors before the broken line since only the % error associated with the cluster containing the new % breakpoint has changed for(k = 1 :new_bp_i-2) error = error + errorclusterold(k); error_cluster(k) = error_cluster_old(k); Nlpcluster(k) = Nlp_cluster_old(k); gamma_cluster(k) = gamma_cluster_old(k); end; % calculate error for the broken cluster into two clusters for(k = new_bp_i-l :new_bp_i) start = newbp(k); if(k== Nc+1) stop = Nip; else stop = new_bp(k+1)-1; end; coef = polyfit(t_lp(start:stop), pdp_lp(start:stop), 1); % least square fit gammacluster(k) = -10*logl0(exp(l))/coef(l); line = poIyval(coef, t_lp(start:stop)); error = error + sum((pdp_lp(start:stop) - line).A2); error_cluster(k) = error_cluster(k) + sum((pdp_lp(start:stop) - line).A2); Nlp_cluster(k) = stop-start+1; end; % add up the errors after the broken line for(k = new_bp_i+l:Nc+l) error = error + error_cluster_old(k-l); error_cluster(k) = error_cluster_old(k-l); Nlp_cluster(k) = Nlpclusterold(k-l) ; gammacluster(k) = gamma_cluster_old(k-l); end; % complete the 'RM'S error calculation, i.e., take the mean and % then square root everything, error = sqrt(error/Nlp); for(k= l:Nc+l) error_cluster(k) = sqrt(error_cluster(k)/Nlpcluster(k)); 114 end; % update the minimum error and breakpoint i f necessary if(min_error > error) minerror = error; bp = new_bp; min_error_cluster = errorcluster; minerrorgammacluster = gammacluster; end; end; end; end; if(min_error ~= inf) % min_error_gamma_cluster meanminerrorgammacluster = mean(minerrorgammacluster) end; % i f least squares fit fails, assign something to all outputs if(min_error == inf) bp = 0; algorithm_fail = 1; end; cluster RMS error.m - end 115 

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            data-media="{[{embed.selectedMedia}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.831.1-0100775/manifest

Comment

Related Items