Mixed Gas Recognition And Concentration Estimation Algorithm Based On RVM

As metal oxide semiconductor(MOS) gas sensor has the advantage of high sensitivity for combustible and explosible gas, fast response and low cost, it is widely used in gas detection system. But it cannot determine the type and concentration of gas according to the response of the single sensor because this kind of sensor has a characteristic of cross-sensitivity, when under the conditions of various gases.This thesis mainly focuses on the research of the gas recognition and concentration estamition with sensor array.In order to improve the selectivity of gas sneor, the thesis uses 5 different gas snenors of sensitivity to form a sensor array, build up the experimental system, test and analyz the characteristics of the response signal of the sensor array and collect gas data for the subsequent experiments.As the sensor response is non-linear, this thesis propses a method combined kernel principal component analysis(KPCA) and M-RVM to realize qualitative identification of the mixed gas. The KPCA algorithme can map the non-linear data to high-dimensional feature space by using kernel function. The M-RVM algorithmeis relatively sparsely, has less parameter Settings and identify the gas categoryin the form of probability. It is sutible for analyzing the uncertity of the classification. Experimental results show that the proposed mixed gas identification method can identify with binary gas mixture, the recognition rate is reach up to 99.83%.In order to reduce error in the mixed gas concentration estamition, this thesis proposes the Multivariate relevance vector machine(MVRVM) altorighm for the gas quantitative detection. MVRVM does well in nonlinear and small sample data. It can solve the multiple variables regression problem and output multiple variables simultaneously so that the testing time is reduced and the algorithm is simplified. The average relative error CO gas and CH 4 gas estimation is 5.58% and 5.38%. Compared with the result of single RVM and LS-SVR, this method has a concentration estimation accuracy and d etection time certain advantage, an effective solution to cross-sensitivity characteristics of the sensor is more suitable for real-time detection.