Studies On Respone Interference Of VOCs Gas Mixture And Recognition With Neural Network

Volatile organic compounds(VOCs)are the main pollutants in the indoor environment which affect human health.They are also the biomarkers to detect human diseases,so study on detection of VOCs has been more and more attention.Neural network with sensor array is often used to solve two kinds of VOCs recognition problems:gas classification and concentration prediction.The sensor array gets information of gas mixture,and the neural network obtains the recognition conclusion by analyzing the information.Classification accuracy,prediction error and training speed of neural networks are closely related to number of layers,number of nodes in each layer and connection between nodes.In order to improve recognition performance of VOCs mixtures by neural networks,In this paper,response interference of gas mixtures,determining method of neural network structure and formation of sensor array have been studied.The application of extreme learning machine(ELM)improves the speed of recognition of gas mixture with sensor array.Extreme learning machine with sensor array was used to predict concentration of each component in a mixture of fixed four components.The results show that extreme learning machine training time is 0.03 seconds,compared with training time of the back-propagation(BP)neural network(14 seconds)and radial basis function(RBF)neural network(10 seconds),speed were increased 466 times and 333 times respectively.The training sample increased from 22 to 91,and training time of extreme learning machine keeps unchanged.The reason is that most weights of ELM are randomly generated during the training process,and other weight are obtained by solving equations,which replaces the iterative of traditional neural network learning,and improves the speed.Based on the sensor response mechanism,test phenomenon of ethanol,acetone,and their mixtures is analyzed.Test phenomenon of gas mixture is that the algebraic value of sensor response to two individual gases ethanol and acetone greater than sensor response value of their mixture.Based on the analysis of response interference,a new gas type definition method is proposed.The new method classifies all gases in the mixture except arget gas as a gas type,and classifies gas type only with different concentrations of target gas.The new method can effectively avoid dimensions problem of traditional gas type definition method,traditional methods define gas type with two dimensions according to different types and different concentrations of gas.The new method solves the dependence of neural network structure on specific gas samples.BPNN,ELM and support vector machine(SVM)with sensor array are used to classify the single component,two component and four component gas mixtures with different concentrations respectively.The new method successfully reduce gas species from 108 to 4,and number of output nodes from 11 to 2,the neural network with 2 output nodes eliminates the correlation between neural network structure and the species and concentration of other gases in gas mixture,except for concentration of target gas.According to the property that L1 regularization term can make partial weight of neural network approach zero,the pruning method is proposed to determine number of hidden nodes in neural network.Considering mathematical meaning of the regularization term,response characteristics of sensors and composition of sensor array,the rules of neural network hidden layer node pruning are established.Number of hidden layer nodes determined by the pruning method is 16,compared with 15 hidden layer node determined by exhaustive,test mean square error of two neural networks by the two methods for the same sample is 0.16 different.In determining number of hidden nodes,pruning method solves the randomness problem determined by experience and the complexity problem determined by the exhaustive method.In order to ensure that sensor array can obtain sufficient,accurate and effective gas sample information,the rules of sensor array are set.Considering the characteristics of each sensor in sensor array and the sensor response to gas sample,the selection of sensors is made in terms of sensor's response to low concentration target gas,amount of information and significance of the information obtained by sensors.