Application Of Deep Learning In Electronic Nose Gas Recognition

Electronic nose,as an emerging olfactory bionic technology,has made certain progress but still faces some challenges in practical applications.Among them,the features revealed during the early response stage of gas detection are insufficient,resulting in a lower recognition rate at this stage,which limits the application of electronic noses in rapid response scenarios such as chemical safety.In addition,the output of gas sensors is affected by factors such as the environment and aging,causing sensor drift and a subsequent decrease in recognition rate.To address these two challenges,this study conducts an in-depth investigation from the perspective of signal processing using deep learning techniques,proposes effective solutions,and achieves certain research results.To solve the problem of low recognition rates in the early stages of gas response,this study utilizes a self-built electronic nose system and collects an industrial pollution gas detection dataset as the research object,aiming to improve the performance of early gas detection from the perspective of dataset expansion.First,in response to the limited number of samples in small datasets,this research proposes four data augmentation methods suitable for electronic nose samples,namely random sensor dropout,random sensor response time shifting,and overall response time shifting and scaling.The goal is to improve the generalization performance of subsequent classification algorithms by expanding data sources.Second,during the feature learning stage,to account for the spatial irrelevance of the electronic nose’s multichannel time-series signals,this study applies the gramian angular field method to electronic nose data,transforming the timeseries signals into two-dimensional images and consequently converting the gas recognition problem into an image classification problem.This creates opportunities to introduce relevant methods from the visual domain.Moreover,this method better reveals the relationships between points on the time-series signals.Third,in the pattern recognition stage,this study proposes a convolutional neural network model,GAS-Net,which combines the advantages of both Res Net and Inception network models.Using residual connections and multi-scale receptive field convolutional kernels,the model integrates the popular attention mechanism from the visual domain and employs genetic algorithms for network structure optimization.The experimental results demonstrate that the proposed method significantly improves the recognition accuracy during the early response stage of gas detection,achieving a performance comparable to that obtained when using the entire response stage time-series signals for recognition.This study extends the application potential of electronic noses in scenarios requiring rapid response,such as industrial production safety.To address the issue of decreased recognition accuracy caused by sensor drift,this study selects a publicly available sensor drift dataset as the research object and proposes two automatic drift compensation models based on drift factors and LSTM networks.First,to tackle the vanishing or exploding gradient problem during the training process of traditional LSTM networks and to reduce the adverse impact of outliers on classification performance,this study proposes an improved network model called LNLSTM.By incorporating layer normalization techniques into the traditional LSTM network,the convergence speed and generalization performance of the model are significantly enhanced,achieving a certain degree of compensation for sensor drift.Second,in order to specifically enhance the drift compensation capabilities of the network,this study builds upon the LN-LSTM framework by introducing the time variable as an additional input for sensor responses.Additionally,the Multilayer Perceptron(MLP)and quadratic functions are integrated into the LSTM network units.Drift compensation value estimation is designed based on both MLP and quadratic functions,allowing the network to learn compensation parameters during the training process.This enables automatic real-time drift compensation for responses with known time variables.The contribution of this study is to improve the reliability of the electronic nose in practical applications and also extend its service life.