Research On Feature Extraction And Optimization Methods For Signals Of Artificial Olfaction And Taste Systems

Artificial bionic systems are electronic devices that mimic biological sensory systems for intelligent operation,electronic noses and electronic tongues are two important categories among them.Both electronic nose and electronic tongue systems are usually built in a similar structure including three parts: sensor array,signal preprocessing unit,and pattern recognition unit.When a system utilizes combined information from electronic nose and electronic tongue for intelligent detections,it will face the problems caused by mixed noises and signal characteristics of two different devices.Therefore,it is necessary to set up the foundation for comprehensive recognition by exploring feature extraction methods of olfactory and taste information separately.Based on signal characteristics,this research includes a parallel research strategy for feature extraction and optimization studies of both electronic nose and electronic tongue.For both gas-sensor drift problem of electronic nose and dimensional disaster problem of Large Amplitude Pulse Voltammetry(LAPV)electronic tongue,appropriate feature extraction mathematical models are established to improve the recognition performance of these two systems.The detailed contents of this research are as follows:(1)Study on electronic nose drift suppression methods by feature extractionA domain adaptive subspace learning method is proposed to address the problem of data distribution difference caused by drift.This method can learn a domain adaptive subspace maximizing the dependency between feature labels and minimizing feature redundancy(DMDMR)to solve the electronic nose based drift problem.The Hilbert-Schmidt independence criterion is used to characterize the relationship between the features and the labels in the source domain,and the data structure of the samples is maintained by the maximum variance criterion.In the experimental part,a electronic nose drift dataset is used to evaluate the performance of the proposed method and analyze the data distribution differences in the learned subspace.The experimental results show that the DMDMR method can effectively reduce the data distribution difference,and obtain the highest average accuracy compared with other state-of-art feature extraction methods in different drift-level scenarios.It can be seen that the proposed method not only effectively solves the data distribution problem caused by drift,but also retains the features containing a lot of classification information.Thus,the proposed method succeeds in drift suppression and recognition improvement of electronic noses.(2)Study on electronic tongue signal enhancement methods by feature extractionDue to the high dimensionality and common mode components(highly correlated signals)in the response signals of LAPV type electronic tongue,it is difficult for machine learning models to obtain effective features,which leads to a decline in accuracy.To this end,this study introduces a pattern recognition framework that is suitable for LAPV electronic tongues.The framework includes a Feature Specificity Enhancement(FSE)method for feature optimization and dimensionality reduction at the feature extraction level;the pattern recognition part directly selects Kernel Extreme Learning Machine(KELM)to improve the recognition speed and flexibility of proposed framework.The experimental part introduces two data sets from the self-made electronic tongue and a public access to evaluate the effectiveness of the framework.The evaluation results show that the proposed FSE method achieves the highest accuracy and the best computational efficiency on both datasets when using KELM for classification.In addition,parameter sensitivity analysis of the proposed FSE method and KELM provides the algorithm parameter setting range of the pattern recognition framework.In this research,the feature extraction and optimization methods of electronic nose and electronic tongue are parallelly studied,the signal processing methods of artificial olfactory and taste system are constructed,which finally sets up a foundation for information fusion of combined artificial olfactory and taste system.