Research On Gas Recognition Based On Deep Learning

With the rapid development of machine olfaction, the demand of research on fields such as industrial production, environmental protection, security checks, medical diagnostics, has been urgently raised. As a typical application of machine olfaction, gas recognition system combines cross-sensitivity chemical sensor array and an effective pattern recognition algorithm for the detection, identification or quantification of various odors.The material of sensitive membrane vibrates by absorbing the gas molecule, and data collected by the sensor array are the multivariate time-series signals with complex structure. These signals are influenced by the type and concentration of the gas, the external environment(such as temperature, humidity), and other factors, so that they usually have the complex waveform and are difficult to analyze. In practical application,hand-designed features combined with some means, such as wavelet decomposition and support vector machines, are often used to identify gas.There is a phenomenon called sensor drift, which means that sensor signals tend to show a significant variation and eventually changes the cluster distribution in the data space. The signals become more difficult to analyze due to sensor drift, and degrade the classification accuracy gradually. Currently, many physical and signal processing methods are used to reduce drift.Firstly, this work compared recognition accuracy of various methods. Specifically speaking, the process of extracting features included manual design features, wavelet decomposition and principal component analysis, and the recognition algorithm used support vector machines, k-nearest neighbor and so on. And then we try to use the deep learning methods to improve recognition accuracy. Compared with other methods, the deep learning method can automatically extract features, thus removing the complexity of designing the hand-made features. Meanwhile, unlike previous drift compensation method, the deep learning method is no longer for specific applications, which means it has very broad applicability and good application prospects.Finally, this paper explores a brand new direction, namely using deep learning method to solve the sensor drift problem. Experimental results show that the deep learning method can learn the features that are more robust to drift than the originalinput, and significantly improve the recognition accuracy under sensor drift, which demonstrated its feasibility in the field of drift compensation.