Study On Online Sensor Drift Compensation Algorithm Of Bionic Olfactory System

Bionic olfactory system,also known as electronic nose system,is a system for gas recognition consisting of gas sensor array,signal preprocessing unit,and pattern recognition algorithm.When test gas enters the system,sensor array generates an electrical signal based on the characteristics of the gas.After being processed by signal preprocessing unit,it is inputed into a pattern recognition algorithm model to achieve gas discrimination.Sensor drift occurs continuously in bionic olfactory systems and is difficult to predict.Drift can cause abnormal output response of gas sensors,make different distribution in collected gas samples and degrade the performance of pre-trained recognition model,affecting the accuracy of gas recognition.Most of drift compensation methods proposed are offline.Although those methods have achieved good recognition results,they are not suitable to deploy in practical applications.In this thesis,the idea of transfer learning is introduced so the sets of sensor output response samples collected in two successive time periods on timeline are using as the source domain and target domain respectively.To overcome the limitations of offline algorithms,this thesis proposed two online drift compensation methods to realize the recognition of target domain samples,enhance the robustness of the recognition model,and achieve the purpose that extend the life of bionic olfactory system effectively.1.Online sensor drift compensation based on subspace distribution adaptation.This method constructs a geodesic flow kernel to embed source domain samples and target domain samples into manifold subspace firstly.Then it introduces conditional distribution adaptation and manifold regularization to reduce the difference of feature distributions between source domain and target domain.Finally,a classifier is constructed by structural risk minimization principle.Because drift is a dynamic process occurring continuously,this method using the geometric properties of the grassmann manifold space to updates the classification model online by introduce the drift samples which have obtained predicted labels after each round of classification into the next classifier training process though source domain reconstruction.2.Online sensor drift compensation with sparse autoencoder.This method completes the training of sparse autoencoder and classifier without using any target domain samples.When the target domain sample arrives,it is inputed into the sparse autoencoder to obtain a reconstructed representation containing the nonlinear structure of the source domain,reducing its deviation from the feature distribution of the source domain firstly.Then feature augmentation is used in this reconstructed representation to maintain its characteristic in the original target domain space.And finally,the augmentated sample is brought into the source domain classifier to complete the classification.Experiments are performed on the two methods proposed in this thesis with the public dataset of gas sensor array drift,and the results verified the effectiveness of both methods.