Study On Key Issues Of MOS-based E-nose And Its Application

Metal oxide semiconductor(MOS)-based gas sensor is widely used in the MOS-based electronic nose(E-nose)system as a perception unit due to its high sensitivity,quick response,and low-cost.However,some important interference factors could lead to the poor performance(low accuracy and poor generalization)of the E-nose system in real application scenarios,including signal drift of the gas sensor and external environmental interference.Aiming at the key issues of the MOS-based E-nose exposed to the applications,we focused on the theory studies of the issues and the application studies of the E-nose in detection of Chinese tea.Firstly,gas concentration is an important interference factor when the MOS-based E-nose is used in the qualitative analysis of gas.So,a transductive transfer learning method,Maximum Independence of the Concentration Features(MICF)-Iterative Fisher Linear Discriminant(IFLD),was proposed to suppress the signal drift of the MOS-based gas sensor from the perspective of feature extraction.The MICF reduced the difference in data distribution caused by concentration from the perspective of data distribution.And the IFLD was used to extract the deep features of the sensor response signals by reducing the divergence within classes and increasing the divergence among classes,which helped to prevent inconsistent ratios of different types of samples among the domains.The MICF-IFLD improved the consistency of sample data distribution,and improved the accuracy and generalization of the classification model in gas qualitative analysis.Secondly,a semi-supervised domain adaptation model based on Long Short-Term Memory network(DALSTM)was proposed for signal drift compensation in the E-nose applications,in which it was unnecessary to subdivide the factors that caused drift.The robust model could be obtained by leveraging a limited number of labeled data from target domain adaptively.Experimental results showed that the proposed DALSTM model achieved high accuracy in the experiments,thus the difficulty during feature extraction was effectively avoided.Furthermore,we introduced the attention mechanism to improve the generalization performance of the proposed DALSTM model further.Thirdly,for the qualitative and quantitative analysis of gas,a multi-task learning-long short-term memory(MLSTM)recurrent network was proposed for simultaneous gas detection and concentration estimation,in which the two tasks shared underlying features.The network utilized the synergy between both two tasks,thereby enhancing the individual performances.The effectiveness of the framework was verified,which alleviated the concentration estimation error caused by the identification errors to some extent.And Particle Swarm Optimization(PSO)proved a good method for optimizing the hyper-parameters used in the framework.Also,we presented two algorithms for simultaneous recognition of four properties(wine region,grape variety,vintage,and fermentation processes)based on a multitask learning framework for multi-property detection of wine.Finally,three applications of the MOS-based E-nose were studied in tea detection.Some pattern recognition methods were developed with strong robustness,good generalization and high accuracy,which could identify the tea brand,origin,quality and other information accurately.Also,we explored the differences between different types of tea samples by Gas Chromatography-Mass Spectrometer(GC-MS).It was verified that the effectiveness and accuracy of the E-nose in tea detection.The proposed methods and models could provide scientific support for gas analysis in several industries(food detection,environmental protection,and leakage detection of toxic/harmful gas,etc.),and the results of the study could promote the commercial application of the MOS-based E-nose.