Improvement Of Sensor Selectivity Of Nano Sensor Array Based On Machine Learning

At present,machine olfaction has been applied to food safety,medical and health,consumer electronics,environmental monitoring and aerospace and other fields,which are inseparable from our production and life.The processing of gas sensor data is very important in machine olfaction technology.In particular,the use of gas sensor data for gas identification and gas concentration prediction is the main goal of machine olfaction research.In a gas identification or gas concentration prediction system,sensor data needs to be collected first,and then these data are processed using machine learning methods to identify the gas or obtain the gas concentration.At present,there are many machine learning algorithms for gas sensor data processing,but these tasks often use only one or two algorithms for comparison,and do not compare multiple algorithms.In practical industrial applications,multiple methods must be tried and compared to obtain the best method in a specific application.The dissertation first analyzes and compares various gas classification methods.In the feature engineering phase,Z-score standardization and polynomial-based feature transformation are used for data preprocessing,and variance selection and principal component analysis are used for feature extraction.Then,the experimental analysis compares five classification methods: proximity algorithm,logistic regression,decision tree,support vector machine and artificial neural network.The experimental results show that the artificial neural network performs better on the gas sensor data of each batch,which indicates that the artificial neural network with strong nonlinear expression ability is robust in gas classification.In order to further improve the accuracy of gas classification,this paper proposes a classification algorithm combining artificial neural network and logistic regression.We found in practice that there are more suspicious samples in the prediction results of artificial neural network models,and the prediction accuracy of these suspicious samples is often low.To solve this problem,the algorithm uses logistic regression to predict suspicious samples in the prediction results of the artificial neural network,thereby improving the classification accuracy.Experimental results show that using this algorithm can effectively improve the classification accuracy of gas,so using this algorithm for gas identification has a wide range of application scenarios.Finally,considering that the prediction of gas concentration is also very important in machine olfaction,the prediction of gas concentration belongs to the regression task.Therefore,this paper gives the evaluation indicators and commonly used regression methods of gas concentration regression.The performance of support vector machine and artificial neural network in gas regression task was analyzed and compared through experiments.Among them,Z-score normalization and principal component analysis were used.The experimental results show that Z-score normalization and principal component analysis combined with support vector machines have a better performance in gas concentration regression tasks and can be considered for use in the engineering field.