Statistical Shape Mapping Pre-Processing of Metal Oxide Gas Sensor Response for Machine Learning Detection Improvement

Abstract

IntroductionModulation of metal oxide (MOX) gas sensors working temperature offers possibilities of highly selective detection of various gases, when coupled with signal processing and machine learning [1]. However, stability of response is still a major issue [2], hampering widespread practical application of MOX sensors in the fields of industrial safety, public security, ecological control, non-invasive medical diagnostics etc. Most of the modern approaches of signal pre-processing, aimed at elimination of baseline drift effects, signal amplitude fluctuations due to sensor aging or background variations of analyzed air, act through selection of the most significant features of the response, leading to the loss of some useful part of the signal [3]. This may lead to the unstable system performance in variable ambience conditions, which are characteristic to urban air or air at industrial facilities.The present work introduces shape-space mapping pre-processing of MOX gas sensor response for further application of machine learning in order to improve the identification and quantification of common flammable and pollutant gases.ExperimentalThree different gas detection tasks were studied. First - identification of propane vs hydrogen in the flow of real urban air; second – identification of methane vs propane in same conditions, third – determination of CO concentration in clean humid air. Laboratory made SnO2-based MOX gas sensors [4] were used for the solution of the first and second tasks. Sensors were linearly thermally modulated between 100 and 500 oC in the constant flow of outdoor air, to which gases were admixed. Statistical analysis of response curve shape (statistical shape mapping - SSM - in 2D space of sensing element electrical resistance and temperature) was performed as a pre-processing procedure prior to machine learning algorithm use. The performance of the obtained ML model was evaluated with the use of completely independent set of data, obtained in separate days of experiment [5]. The ability of SSM to improve the gas identification accuracy of the ML model was compared to raw sensor response, normalized response as well as response, pre-processed via principal component analysis (PCA), discrete wavelet transform (DWT) and polynomial curve fitting (PCF) approach. Same comparison has been made during the solution of the third task with the use of publicly available database of response of thermally modulated commercially available MOX gas sensors [6,7].Results and discussionThe results of the ML model testing in the case of methane vs propane discrimination task (Table 1) shows that improvement of sensitivity (AuPd/SnO2 sensor [4]) improves the accuracy of detection. Normalization of sensor response doers not necessary lead to the reduced error of model answer and other pre-processing techniques lead to the decreased accuracy of detection compared to the raw sensor response, used for the machine learning. The only exception is SSM pre-processing approach, improving the accuracy of identification in the case of all three used sensors. The day-to day error variation does not correlate with the atmospheric pressure, air temperature, humidity or regional air quality index (AQI), indicating local uncontrolled pollution as a source of erroneous gas identification [8]. Similar results are obtained in the case of hydrogen vs propane identification task [9]. For CO quantification task by commercial sensors with rectangular impulse type working temperature modulation much better performance of ML model is observed in the case of SSM pre-processed response of TGS 3870-A04 gas sensor type (Figaro), than in case of SB-500-12 (FIS) sensors compared to other pre-processing technique [8].ConclusionsThe application of statistical shape space pre-processing to the signal of temperature modulated metal oxide gas sensors improves the selectivity of gases identification with the use of ANN based machine learning algorithm in comparison to other methods – PCA, DWT, PCF, data normalization. It helps to eliminate effects of sensor response drift, amplitude fluctuations, while the major source of erroneous gas identification may rise from the lack of shape features in the original sensor response and the uncontrolled local background air pollution. Improvement in general sensitivity of sensor, is favorable for shape space signal pre-processing in order to enhance the selectivity of chemically similar gases detection at low concentrations in realistic air conditions. The proposed approach can be considered as versatile as its performance is successfully tested on different type of sensors and gases in different range of concentrations in air with different types of working temperature modulation.AcknowledgementsThe reported study was funded by RFBR according to the research project № 18-33-20220.