Qualitative Analysis And Quantitative Estimate Of Mix Gases Based On The Electronic Nose System

The electronic nose, who is a electronic system that developed by simulating animal smell organ, is consisted of chemical sensor array and pattern recognition system. By using part specificity of sensors and pattern recognition technology, it can smell some gases.Pattern recognition is such a learning process that the machine observed the pattern classification environment, separated effective model from the background and make a reasonable decision. As the'brain'of the electronic nose, pattern recognition algorithm will have a direct impact on its ability on detecting gases. Functionally, gas identification mainly includes the qualitative analysis and the quantitative detection. According to these two aspects, this paper designed a simple electronic nose experimental system and a mixed neural network. The network was trained by the data got through a set of perfect experiments and can achieve a high precision on both aspects.Based on the aim of training a neural network with the qualitative analysis and the quantitative estimate ability, this paper launched a series of research, including:(1) The development of sensor array collection system, the system includes six parts, they are the configuration of gases, the sensor array consisted of seven different types of metal oxide semiconductor sensors, the detection circuit of analog signal, AD sampling circuit, and the PC data receiving and saving software developed by our own. (2) The establishment of intelligent identification system, the system includes data processing and neural network training. Data processing refers to pretreatment and feature extraction. And the pretreatment includes baseline reduction and normalization. In order to reduce feature dimension, this paper comparatively used the principal component analysis (PCA), linear discriminant analysis (LDA), and the feature selection method. The neural network training includes the BP neural network training to qualitative identification and Fuzzy ART neural network training to establish the fuzzy estimator for quantitative analysis.Three dimension reduction methods were contrastively used, which were PCA, LDA and feature selection by the ratio of the global variance and local variance. Whether for ammonia, ethanol or their mixture, LDA performed the most ideal on dimension reduction of classification. The input data of the BP neural network that processed by LDA can identify the class of the gas with an accuracy rate of 100%, while processed by other two methods only have 84.3% and 95.4%. Each class of gas should establish a fuzzy concentration estimator. And each estimator was established by the weight vector got by Fuzzy ART learning. Then, the concentrations of experimental gases can be estimated by defuzzification process. Test results showed that it has a correct estimate rate of 100% to the training concentration, and a correct estimate rate of more than 96% to the ones within the scope of training concentrations.This hybrid neural network including BP and fuzzy concentration estimator based on Fuzzy ART neural network is simple, effective and can play its advantage in the quantitative estimation of the mixture gases.