Metal Oxide Gas Sensors Signal Shape Processing for Selective Detection of Hydrocarbons in Realistic Air Conditions

Abstract

Introduction The selective detection with high spatio-temporal resolution of hydrocarbons leakage as a result of pipelines inconsistency is a valid industrial demand [1]. Perspectives are associated with deployment of distributed networks of chemical sensors, transferring data on ambience pollution to the data processing center [2]. Miniature micromachined metal oxide semiconductor gas sensors possess a great perspective of practical use in this regard, however their long term operation in real atmosphere conditions requires improvement [3]. In this work we demonstrate improved selectivity of propane vs. methane detection in low concentrations in the ambient air of highly urbanized location by the SnO2-based semiconductor gas sensors with the modulated working temperature, using statistical analysis of “sensor response/sensor working temperature” shape representation. The obtained results of gases discrimination with artificial neural network (ANN) machine learning algorithm, used such pre-treated data as input samples, demonstrate advantages over similar classification methods without signal pre-processing, or pre-processing with previously reported methods of PCA, wavelet transformation and polynomial curve approximation. Sensor fabrication, data collection and processing Three SnO2-based materials have been used – pure SnO2, gold-modified SnO2 and bimetallic Au and Pd modified SnO2. The gas sensing materials synthesis via single step flame spray pyrolysis technique and sensors fabrication has been reported in a recent paper [4]. A fixed air flow from outside of the building of the Department of Chemistry of Moscow State University through PTFE sensor chamber was used as a background. Methane or propane were admixed to the background air flow through precise mass-flow controllers in three different concentrations from certified gas bottles. Sensors were operated in a working temperature modulation mode between 500 oC and 150 oC with period length of 60 sec. Sensitive layer resistance was measured with the 10 Hz frequency. Collected data was separated in two parts: first, collected in December 2018 was used for data processing algorithms training. Second, collected during January and February 2019, was used for the testing of elaborated data processing models. The total data set of 22440 measurement cycles,11220 of which represent an ambient air, while two data subsets of 5610 measurement cycles represent same air with admixture of methane or propane, were used for processing. The multilayer perceptron ANN-model with 2 hidden dense layers has been used for data processing model development and gases discrimination. The batch normalization and dropout techniques were implemented to avoid model overfitting. According to statistical shape analysis approach, each characteristic point of sensor response is characterized by certain gas sensor working temperature and sensitive layer resistance. All characteristic points of each single measurement cycle during analysis are converted into Kendall’s pre-shape space with removed translation and scaling factor, but with remaining rotation as well as the shape. During further processing the data samples are projected onto the unit-sphere with a pole of mean shape of sensor responses towards air without additives. Later inverse transformation of data samples leads to icon representation of response shape, eliminating possible effects of sensor baseline and response amplitude drift. Results and Conclusions Direct application of collected data to artificial neural network algorithm for methane vs. propane selective detection without any pre-processing gives a decent level of accuracy (Fig.1). However, it can be seen by the rise of discrimination error after normalization of data samples, that the main component of the sensors response, used by algorithm for gases discrimination, is of amplitude nature. It means, that the developed signal processing model may work in unstable manner when any additional gas is involved in the measurements. The use of statistical shape analysis allows to significantly improve the accuracy of discrimination of methane vs. propane in comparison with other pre-processing techniques. The remained error of gases identification is mostly related to the single day of measurements, which is associated with urban air pollution during winter heating season. References [1] L.E. Mujica; M. Ruiz; J.M. Mejia, Leak Detection and Localization on Hydrocarbon Transportation Lines by Combining Real-time Transient Model and Multivariate Statistical Analysis. Struct Hlth Monit 2015, 2350-2357. doi: 10.12783/SHM2015/292 [2] D. Spirjakin; A.M. Baranov; A. Somov; V. Sleptsov, Investigation of heating profiles and optimization of power consumption of gas sensors for wireless sensor networks. Sensor Actuat a-Phys 2016, 247, 247-253, doi:10.1016/j.sna.2016.05.049. [3] Collier-Oxandale, A.M.; Thorson, J.; Halliday, H.; Milford, J.; Hannigan, M. Understanding the ability of low-cost MOx sensors to quantify ambient VOCs. Atmos Meas Tech 2019, 12, 1441-1460, doi:10.5194/amt-12-1441-2019. [4] Krivetskiy, V.; Zamanskiy, K.; Beltyukov, A.; Asachenko, A.; Topchiy, M.; Nechaev, M.; Garshev, A.; Krotova, A.; Filatova, D.; Maslakov, K., et al. Effect of AuPd Bimetal Sensitization on Gas Sensing Performance of Nanocrystalline SnO2 Obtained by Single Step Flame Spray Pyrolysis. Nanomaterials-Basel 2019, 9, doi:10.3390/nano9050728. Figure 1

Understanding Structure–Property Relationships of MOFs for Gas Sensing through Henry’s Constants

“…Detecting gases at dilute conditions is of great importance in the product quality monitoring and analysis for many applications. Some of these major sectors include sensing volatile organic compounds (VOCs) in foods and cosmetics, hydrocarbon leak detection in industries, environment pollutant monitoring, automobile combustion control, detecting poisonous gases in a warfare zone, etc. − Usually, sensors are deployed for such tasks, and conventionally, solid state metal oxide based materials have been the preferred sensing component . The sensor here works by a physical or a chemical change in this sensing material.…”