Method For Improving Detection Accuracy Of Electronic Nose System

Electronic nose is a kind of bionic system which simulates biological olfactory system. Compared to traditional detection method, electronic nose has many advantages, such as fast response, real-time monitoring and high efficiency and so on, which has widely used in air quality monitoring, food industry, medical diagnosis and other fields. The qualitative and quantitative analysis of the gas, namely classifying the gas and determining the concentration of the gas, both of them are two major indicators of detection precision in e-nose system In this thesis, research is in the background of air-quality monitoring, and focuses on the study of improving the detection accuracy of electronic nose.Currently, popular traditional methods to improve the detection accuracy of classification are linear discriminant method, back propagation artificial neural network (BP-ANN) and support vector machine (SVM), which are widely used in both qualitative and quantitative analysis of air-quality monitoring device. Each of them has its advantages and disadvantages. BP-ANN is prone to get a local minimum convergence value. SVM has a relatively well performance on classification while it depends on high quality features. In order to improve the recognition accuracy, SVM is usually used in combination with some feature extraction methods. However, whether linear methods like principal component analysis (PCA) and independent component analysis (ICA) or nonlinear methods such as kernel principal component analysis (KPCA) and Kernel entropy component analysis (KECA), some useful information will be missing during the process of feature extraction, and lead to the gas feature cannot be kept completely. Therefore, deep belief network (DBN) was introduced into electronic nose in this thesis. DBN can automatically do feature extraction and detection by studying a large number of data, so that the performance of classification of the system is improved. The restricted Boltzmann machine (RBM) structure is adopted in hidden layer of DBN, so that the drawback of local minimum problem can be overcame as well as the benefit of BP neural network is inherited. To show the superiority of DBN in feature extraction, classification results by different feature extraction methods combined with DBN and SVM are compared. The results show that DBN outperform other methods.Though DBN can improve the detection accuracy used to classification, which couldn't achieve drift compensation caused by temperature and humidity. Drift is a major factor that affects the accuracy of the electronic nose system. Therefore, it is important to suppress and compensate the sensor drift. For the sensor drift problem caused by humidity and temperature, there are studies that show polynomial fitting can effectively remove the drift caused by temperature and humidity. However, polynomial fitting frequently occur over fitting, which results in good performance in training phase but bad performance in testing. To solve this problem, the sensor drift compensation method based on LP regularization is proposed in this thesis, and simulation results show that the effect of after LP regularization the polynomial fitting is obviously better than that without LP regularization.