Study On Sensor Drift Compensation Algorithm In Eelctronic Nose System Via Domain Adaption

In the past two decades,electronic nose system has been widely used in the fields of environmental,agricultural monitoring,medical diagnosis,etc.However,during the long-term use of the electronic nose,sensor sensitivity may be influenced by many factors,such as environmental factors,self-aging and poisoning.Therefore,the change of sensor sensitivity will result in the fluctuation of sensor responses when the electronic nose exposed to the same gas in different time,which is called the sensor drift.It limited the application of electronic nose.Therefore,the thesis mainly studies the sensor drift compensation algorithm in the electronic nose system to improve the service life of the electronic nose.The sensor drift makes a variation of the stastical distribution between the collected samples previously(data without drift)and the collected samples later(drift data)and causes a decrease in prediction accuracy of pattern recognition.Therefore,the problem of the thesis is how to make the trained model with data without drift be better applied to the drift data.The thesis focuses on the goal of improving the accuracy of pattern recognition.And the data without drift is considered as the source domain and the drift data is regarded as the target domain.From the perspective of using domain adaption to aligning the data distribution,Studies on the drift compensation algorithm will be developed for the two scenarios: a small number of labeled samples in the target domain and unlabeled samples in the target domain.The main research contents are as follows:1.For the scene of the target domain with a small number of labeled samples,In this thesis,a method,namely Estimate Domain-Invariant Prototypes via Adversarial Learning(ALDIP),is put forward for drift compensation.The basic model of the algorithm includes a feature extractor and classifier composed of neural networks,and uses entropy to calculate the similarity between the unlabeled target domain sample features and the estimated prototypes(representatives of each class).In order to extract the discriminative features of the target domain,the conditional entropy of unlabeled target data with respect to the classifier is maximized and minimizes it with respect to the feature extractor.Therefore,the trained classifier with labeled data can be better used on unlabeled target domain data.Finally,experiments show that the algorithm can effectively reduce sensor drift in the electronic nose system.2.For the scence where the target domain has only unlabeled samples,the algorithm named Wasserstein Distance Learning Feature Repretations(WDLFR)is presented in this thesis.The algorithm mainly uses the Wasserstein distance to measure the differences between domains.First,It regards a neural network as a domain discriminator to measure the empirical Wasserstein distance between the source domain and target domain.And the method minimizes Wasserstein distance by optimizing the feature extractor in an adversarial manner.Finally,in order to make the feature repretations discriminative,the supervised imformation is integrated into the process of learning feature repretations.Therefore,the learned feature repretations have a domain-invariant and discriminative characteristics.Experiments show that the proposed method is more competitive than the comparison algorithms.