Metal Oxide Gas Sensors Response Processing by Statistical Shape Analysis and Machine Learning Algorithm for Industrial Safety Applications

Abstract

Development of new signal processing approaches is essential for improvement of the reliability of metal oxide gas sensor performance in real atmospheric conditions. Advantages of statistical shape analysis (SSA) pre-processing method in combination with deep learning artificial neural network based machine learning classification algorithm are presented in this regard. The results of the presented method application are compared to simple signal pre-processing techniques for amplitude and baseline disturbance compensation in the task of selective detection of hydrogen and propane. Laboratory made sensors based on highly sensitive Au and Pd modified nanocrystalline SnO2 were used. Modulation of sensor working temperature between 150 and 500 °C was applied. A nearly 30% enhanced accuracy of identification of hydrogen or propane at a concentration range of 30–550 ppm under variable real atmospheric conditions has been demonstrated. The key feature of the presented SSA pre-processing approach is absence of the temperature modulated sensor signal characteristic feature extraction process. Instead, the characteristic pattern of sensor response towards various gases is revealed by the application of the correction procedures of translation, scaling and rotation. As a result of such pre-processing no useful signal shape components are lost, while the effects of signal amplitude and baseline drift are eliminated, facilitating recognition.