Open Collections

Microfluidic olfaction detector for outdoor applications : a pattern recognition approach

University of British Columbia

There exists a growing demand for efficient gas sensing devices in applications involving indoor and outdoor air monitoring, food quality assessment, law enforcement, medical diagnoses and others. Advances in the field of replicating mammalian olfaction are still at infancy as compared to technologies developed to function close to the human senses of vision, hearing, and touch. Microfluidic olfaction detectors (MFODs) are an emerging class of detectors that are low-cost, compact, and highly selective, having a great potential for a wide range of gas monitoring applications. MFODs are, however, susceptible to fluctuations in environmental conditions, and suffer from prolonged sampling times, which hinder their introduction to outdoor applications. This thesis aims at overcoming these two obstacles by developing robust pattern recognition systems using machine learning techniques. To take into account the effect of temperature on MFOD, we perform a machine learning comparison of MFOD performance at three temperature ranges (5, 15, and 25°C) using a limited dataset (as it is challenging to collect substantial amount of data in many applications). By transforming the raw responses using dimensionality reduction techniques and inputting temperature as features, an efficient classifier was able to learn from the small datasets and reached a 95.5% classification accuracy for methane vs. methane-ethane mixure. A unique pattern recognition system was then designed to reduce the recognition time of a MFOD. This system involved: (i) the ensemble of four heterogeneous base machine learning models, (ii) a Long-Short Term Memory Neural Network (LSTM) to construct a portion of the transient response, and (iii) features from an additional fast response sensor. The system was applied to different transient response times to further identify the minimal transient time needed for the proposed system. This effort resulted in facilitating gas discrimination and quantification with less than 20% of the typical MFOD response time using methane and ethane as a case study. By pairing the pattern recognition systems with the MFOD, the performance is shown to improve significantly while keeping the production cost low. The proposed models significantly add to the capability of MFODs in detecting gases for outdoor applications.

Text

2022-01-31

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/77210

Mechanical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/