Research On Electronic Nose Drift Compensation Method Based On Cross-domain Learning In The Subspace

An electronic nose is an electronic device that mimics human olfactory perception and can be used to detect and identify odors.However,drift phenomenon may occur in the process of long-term use,which will destroy the stability of electronic nose and reduce the service life of electronic nose.Therefore,electronic nose drift compensation has become one of the key research directions in the field of machine olfaction.The traditional drift compensation methods generally start from the point of view of signal preprocessing or drift component correction.Although they have certain compensation ability for sensor drift,they also have certain limitations.In recent years,great progress has been made in the research of machine learning,which has a broad development prospect in the electronic nose drift compensation task.The data of electronic nose are collected in different environments and different periods,so it is difficult to satisfy the assumption that the training set and the test set of traditional machine learning methods follow the same distribution.The emergence of domain adaptive is just to solve this practical problem.Therefore,the main research focus of this thesis is the electronic nose sensor drift problem.For the data set Ⅰ with time drift and the data set Ⅱ with both time drift and systematic difference,two drift compensation methods based on domain adaptive are proposed as follows:(1)The existing domain adaptive methods usually use source domain data to train the model,but the underutilization of source domain tags may lead to the decline of performance in the target region.To solve this problem,this thesis proposes a robust domain correction potential subspace learning method.The method learns tag space and feature subspace together,uses source domain tags to guide feature extraction,and applies row sparse constraints to improve robustness and stability.At the same time,the discriminant data layout is constructed by mapping the source domain and target domain through linear discriminant analysis and local preservation projection,thus preserving the discriminant information and local geometry information of the original data.(2)Some drift compensation methods try to reduce the domain distribution difference through some distribution differences and discrimination distance,but this may lead to the decline of feature discrimination ability.Therefore,a domain correction subspace learning method based on orthogonal constraints is proposed in this thesis.This method unified the data distribution,improved the data distribution difference before and after the drift,and retained the local geometric structure of the data for the target domain data without labels.In addition,the mapping relationship between subspace and label space is maintained,and orthogonal constraints are imposed on the projection matrix,so as to reduce the redundancy between features and improve the efficiency of feature extraction.Finally,the multiplication iterative rule is skillfully used to solve the model,and the solving process is explained and analyzed in detail.In order to fully evaluate the model performance,two common gas sensor datasets and different experimental settings are used in this thesis to verify the validity and reliability of the proposed methods.