Open Collections

Applications of artificial olfaction and machine learning for detection of volatile gas

University of British Columbia

Electronic noses are used in diverse applications where sensory information about chemical compounds in the air can provide valuable insight, just as humans use their sense of smell this way. In this thesis, the role of machine learning for data analysis in three main applications of electronic noses is examined. These applications are the regression of arbitrary binary mixtures, detection of natural gas leaks, and detection of THC in human breath. To emphasize the importance of machine learning in the overall workflow of an electronic nose, the data collection apparatus selected was a single metal oxide semiconducting sensor embedded in a coated microchannel to provide selectivity. This apparatus produces a time-series response as the target gas sample diffuses through the microchannel toward the sensor. A comparison of machine learning models and time-series data representation methods is examined for each application. Using this method, a 98.75% accuracy in classifying between methane, ethane, and binary mixtures is obtained, as well as a 12.0% mean relative error in regression estimates for arbitrary mixtures. For the natural gas detection application, simulated natural gas mixtures containing 1% and 3% ethane were discriminated from samples of pure methane with 86.7% and 93.3% accuracy, respectively, and concentrations of both methane and ethane were regressed with a maximum of a 19.3% mean relative error. Finally, samples of methanol were discriminated from samples of denatured THC dissolved in methanol with concentration 1.0 mg/mL. This classification was performed with 100% accuracy using a support vector machine.

Text

2021-06-01

10.14288/1.0392037

Master of Applied Science - MASc

Applied Science, Faculty of; Mechanical Engineering, Department of

2020-09

Graduate