Study On Sensor Drift Compensation Algorithm Of Electronic Nose System

Electronic nose system,also known as artificial olfactory system,consists of a variety of gas-sensitive sensors.Electronic nose systems are mainly used for monitoring,identifying and analyzing complex gases as well as volatile components,and currently have a wide range of applications in environmental monitoring,food manufacturing,medical diagnostics,and other fields.Due to the hardware limitations of the gas-sensitive sensors,the electronic nose system is always subject to drift,which results in a continuous degradation of the performance of the gas recognition model with time migration.The drift is a continuous process that tends to have a large impact on the distribution of data samples,resulting in large recognition errors in the classifier and recognition model.The drift phenomenon leads to the need for periodic replacement of the electronic nose system,while the complete set of electronic nose system is expensive,and direct replacement of the system will increase the cost significantly.While improving the sensor production process,it is also necessary to use artificial intelligence algorithms for drift compensation,and the current research mainly focuses on the drift compensation algorithm of the electronic nose.In this thesis,we focus on the drift compensation algorithm of the electronic nose system and study how to improve the recognition accuracy of the algorithm for different distribution data when drift is generated.In this thesis,the data without drift is considered as the source domain and the data after drift is generated is considered as the target domain.For whether the target domain contains a small amount of labels,two different drift compensation algorithms are proposed in this thesis,as follows.1.For the scenario that the target domain contains a small number of labels,this thesis proposes the drift compensation algorithm with dynamic distribution alignment of Wasserstein distance with time factor.The algorithm mainly considers the drift degree between different data batches and dynamically adjusts the weight of edge distribution alignment and conditional distribution alignment to increase the recognition accuracy of the model under different data distributions.The adaptive factor of the model mainly considers the inter-domain distribution distance and data acquisition batch time to estimate the drift degree.Finally,the algorithm is experimentally verified to be effective for drift compensation of the electronic nose system.2.For the scenario that the target domain does not contain labels,this thesis proposes the drift compensation algorithm of Wasserstein dynamic adversarial adaptive network.The algorithm is based on the previous algorithm and combined with the idea of generative adversarial network to carry out the research.The network consists of four parts: a feature extractor,a label classifier,a global domain discriminator,and a local domain discriminator.The global domain discriminator achieves alignment of the edge distribution of the source and target domains,while the local domain discriminator achieves a more fine-grained alignment of the conditional distribution,and the contribution of both to the overall network is determined by the adversarial adaptive factor,which is determined by the Wasserstein distance between the source and target domains.In this thesis,the method is validated with an electronic nose sensing array drift dataset and a drift-shifted electronic nose dataset,and the experimental results show that the method has good compensation for the drift phenomenon of the electronic nose system.