Research On Gas Sensor Data Processing Based On Machine Learning Method

In recent years,electronic nose(E-nose),which is composed of a gas sensor array and a pattern recognition model,has been widely used as an alternative odor analysis method in various fields of production and life.With the further research of the recognition algorithm,the data extraction method of gas information is becoming more and more mature.At present,there is no universal feature method that is suitable for different applications.It is necessary to design and select features from different schemes based on specific problems to solve the problem of poor universality.In addition,according to the characteristics of the data after feature processing,the recognition method also needs to be selected according to the actual engineering needs.Therefore,the overall gas identification process needs to continuously adjust the best decision combination to obtain high classification accuracy.Therefore,in response to the problems mentioned above in various engineering applications of machine olfactory systems,this paper focuses on the integration of machine learning technology and gas recognition technology.Firstly,from the perspective of data analysis,data preprocessing is performed for Z-score standardization and polynomial conversion.In feature processing,the variance selection and principal component dimension reduction of the extracted time-domain features were carried out,and three different feature engineering schemes were combined.Then,by combining four different machine learning algorithms,including K-Nearest Neighbors,Logistic Regression,Support Vector Machine,and Backpropagation Neural Net,the impact of each feature scheme on classification performance is verified and analyzed,the feature scheme with variance selection for feature processing has the best classification effect under the back propagation neural network.We demonstrate the importance of selecting appropriate feature engineering and classification models in gas classification tasks.In addition,considering the actual application of the machine olfactory system,the data collected after the sensor drift is different in distribution,which seriously reduces the performance of the recognition system.To solve the problem of inconsistent data sample distribution and low recognition accuracy caused by sensor drift,a multi-task learning method based on transfer samples to drift of calibration data.Each different batch of data is defined as a domain,in which a selected set of transfer samples is measured,and each target domain’s transfer samples are aligned with the transfer samples in the source domain in their respective projection space to reduce the impact of the target domain drift.By comparing with the existing sensor drift compensation algorithms,this method was verified in the auxiliary under the condition of having achieved better accuracy.And the algorithm can also study the training of the source domain data model and the multiple target domain data forecasting model,which proves the efficiency and convenience of the algorithm.