Research On The Novel Mixture Gas Recognition Algorithms For Smart Electronic Nose System

Electronic nose(E-nose)is widely used in the gas recognition of daily life because of its low cost,easy operation and low power.Recently,although a considerable of researches on the recognition of mixed gases in E-nose have emerged and exhibit good results in the classification of mixed gases,their experiments were generally carried out on the stable condition and their input were the stable features on the gas sensors array,like the response time,recover time and sensitivity.After a series of simple data preprocess operations,using the mainstream classification method can obtain a better result.However,in order to apply these research results to practical production and life,there are two questions as follows: 1)the form and concentration of mixed gases are complex,random and irregular;2)there are relatively few studies on real-time prediction of mixed gas concentration in prior work,and more are predictions of concentration in the complete reaction process.Therefore,the main contents of this paper are to identify the types of mixed gases and predict the real-time concentration of mixed gases in E-nose under random conditions.As for the problem of mixed gases classification,we present a novel multi-label one dimensional convolutional neural network(1D-DCNN)algorithm for classifying binary mixture gases among Ethylene,CO and Methane.The proposed 1D-DCNN can automatically classify the features directly extracted from the raw data set.In addition,the 1D-DCNN processes the raw data set in a multi-label way.Moreover,the comparison is further extended to 10-fold cross validation.It is indicated that the projection of features automatically extracted by 1D-DCNN in PCA are easier to classify than that of the original response.Last but no least,it is observed that the average recognition accuracy of the proposed 1D-DCNN(96.30%)significantly outperforms the conventional methods of SVM,ANN,KNN and RF.As for the real-time prediction of mixture gases concentration,we proposed a novel architecture consisting of a hybrid convolutional neural network and recurrent neural network(CNN-RNN)to predict the concentration of mixed gases in real-time.The proposed CNNRNN successfully predicts the real-time concentration of mixed gases including ethylene,CO and methane within every 0.5s step.In the experimental part,relative difference method and correction delay method are adopted in data preprocessing,and long short-term memory(LSTM)and root mean square error(RMSE)are used as the process unit in RNN part and figure of merit respectively.The experimental results show that,in the prediction for the mixing of CO and ethylene,the RMSE of CO is 35.11 ppm and that of ethylene is 1.33 ppm.In the prediction of the mixed gases between methane and ethylene,the RMSE of methane is 11.88 ppm and that of ethylene is 1.15 ppm.Besides,the experiment is extended to three comparative methods.It is indicated that the proposed CNN-lstm significantly outperforms the combination structure of CNN and gate recurrent unit(CNN-gru),echo state network(ESN)and linear regression(LR)method in the real-time prediction of concentration.