Signal Processing For Gas Sensor Array Based On Relevance Vector Machine

The gas sensor array consists of a plurality of sensors, and the gas sensor use the inherent’cross-sensitivity’feature information for each sensor to measure the odor information. The sensors have different sensitivity, selectivity and repeatability. The sensor array is formed in response to the multi-dimensional pattern space, and the multi-dimensional space is determined by the array contains more information. But now due to the complexity of the smell of the information, relying on the sensor array solely can not smell a comprehensive and accurate analysis of information, so this thesis introduces the relevance vector machine (RVM) for sensor array signal processing.RVM is a new machine learning method, and used for signal processing of gas sensor array. It has good generalization performance, and can predict the probability of type characteristics. This thesis selects five kinds of common rice and Chinese herbal medicine baiwei for example. Use sensor array and pattern recognition technology combined with the method of herbal baiwei shelf-life determination and testing the quality of rice.Before the experimental sample classification starts, the author uses principal component analysis (PCA) for data preprocessing. It not only reduces the computational complexity, but also improves the classification efficiency. To highlight the effectiveness and feasibility of the RVM classification, then the author compare RVM with support vector machine (SVM), neural networks and other methods.The main research of this thesis is carried out as follows:(1) Collecting sensor array signal. This thesis selects five kinds of rice, and herbal baiwei as the research object.The experiment uses an electronic nose to collected experimental samples’ information. With PCA on rice and herbal for feature extraction, and the initial feature vector principal component analysis, the main component of information retained samples, dimensionality reduction to achieve the purpose of reducing the amount of calculation and improve the efficiency of classification.(2) Using RVM to classify experiment samples and determine shelf-life. Use the experimental method to determine the kernel function and kernel parameter RVM classification model and compare the impact of different kernel functions and kernel parameters on classification accuracy to determine the optimal classification model. The classification experiment results show that the Gauss and Ploy3 kernel get higher classification accuracy and less relevance vector number, and relatively shorter time required for classification. Poly3 kernel thus gets the shortest running time, and easy online real-time detection. And when it is extended to multi-RVM classification, the comparison between different kernel functions and kernel parameters get the accuracy of the classification model.(3) Comparing different algorithms classification accuracy. Compare RVM, SVM and neural network classification. The results show that the RVM can effectively overcome the long test time and too many support vectors. Compared with the test time, the BP, RBF, SVM selection method and under the same Gauss kernel, RVM test time is much shorter than BP, RBF, SVM. The other parameters and the mercer kernel condition are not required to change in RVM. The results show that the RVM classification techniques into rice and shelf-life determination can ensure higher precision, and can get much sparse model obtained, and the shorter test time. Compared with the results with other methods, in this thesis the RVM classification is effective for classification and the feasibility is verified. It is suitable for the detection and classification of other samples.