Research On Fast Recognition Of Electronic Nose Self-Correction Based On Deep Learning

There are a large number of occasions in daily production and life that need to identify and monitor various flammable,explosive and toxic gases.Among them,the advanced electronic nose system has a gas sensor array,corresponding signal acquisition circuit part,and algorithm processing part,which can cope with various complex gas sensing and detection requirements compared with the traditional single gas sensor.However,the reaction of the gas sensor portion of the front end of the electronic nose system with the gas molecules is generally time consuming,which seriously reduces the response processing speed of the entire system.This is a very practical and challenging problem for the application of the electronic nose system to situations where fast detection is required,such as a large number of monitoring systems for flammable and explosive gases.Some researchers have tried to improve the processing speed of the whole system through various transient feature extraction methods,but the corresponding algorithm recognition rate is not satisfactory.Therefore,this thesis mainly focuses on the transient signal processing part of the electronic nose system,and independently designs a highly innovative convolutional recurrent neural network.It is also the first combination of convolutional neural network and recurrent neural network used in gas detection and identification of the electronic nose system.The network successfully implements effective feature extraction of transient gas response signals and achieves high accuracy and fast identification in a short time.We use the convolution kernel to automatically extract features from the input data.To better suit gas identification applications,we have optimized the structure of the convolution kernel.In addition to this,a Long Short-Term Memory neural network(LSTM)is added after the convolutional neural network,which extracts sequence information from the transient response curves and can record gradients over time.In addition,the Long Short-Term Memory neural network(LSTM)added after the convolutional neural network which can extract sequence information from the instantaneous curve and can record the gradient in the time direction.Finally,in order to solve the drift problem that is widely-existing in mainstream gas sensor devices,we use a data-preprocessing method based on a single time step instead of directly using the raw data.The database used in this paper is based on the response of 5 batches of 8 different sensors for 4 common flammable gases over a period of time.This sample set is not only used to train our proposed convolutional recurrent gas neural network,but also used to train a variety of widely used gas recognition algorithms(random forest,gradient lifting tree,nearest neighbor algorithm,support vector machine,linear discriminant analysis).Through the comparison of experimental results,we find that the proposed algorithm is effective and superior in the fast identification of gas: the system recognition accuracy based on 0.5 second time can reach 84%,and the system recognition accuracy based on 4 seconds can reach 98%.In addition,this paper also explores the recognition ability of the proposed network model in the case of gas anti-drift.The results show that the model has excellent anti-drift ability,and the results obtained under two kinds of accepted test conditions are better than other typical algorithms.Finally,the paper also compares the output characteristics of the input layer,convolutional layer and recurrent layer of the model,and explores the influence of each layer on the feature expression.