Research On High Efficiency Gas Identification For Electronic Nose System

Metal oxide semiconductor gas sensors have been selected firstly by electronic nose system,since they have simple structure,fast response,long service life,low cost and high sensitivity to combustible gas and organic volatile gas.However,due to the complex mechanism of the sensor response,it is difficult to obtain a priori response function and an accurate mathematical model.When researchers choose signal processing and pattern recognition algorithms,they often rely on experience,lack unified evaluation criteria,which affect the use and development of the electronic nose.Therefore,how to identify gases efficiently is still a problem that needs to be explored and studied in the field of electronic nose.This paper combines method of optimal directions(MOD)or K-singular value decomposition(K-SVD)with sparse representation classification(SRC)algorithm,and proposes 2 more advanced dictionary learning gas recognition algorithms,which are defined as MOD-SRC and KSVD-SRC algorithms.The training sample set is divided into several sub categories according to categories.Only one subset is selected to participate in the computation.In this way,the time needed for computation will be greatly reduced due to the smaller number of coefficients to be solved.The solution of the sparse coefficient is no longer using the very time-consuming l1-norm instead of the dictionary learning method,which solves the high computational complexity caused by minimized l1-norm.The analysis dictionary separates the training phase and the testing stage in the SRC algorithm.Hence the test time has been shortened.We use three algorithms to identify 8 different concentrations of gas,and find that the average test time for SRC is 2.24 s,but the average test time for the KSVD-SRC and MOD-SRC algorithms respectively is about 5.3 ms and 4 ms,which is about 400 times higher than SRC.The recognition rate of the three algorithms is more than 98%.Therefore,two algorithms proposed in this paper not only maintain the recognition accuracy of the original SRC algorithm,but also have a shorter test time.A new model of joint dictionary learning(JDL)algorithm is proposed to reduce the drift and background noises of gas sensors.The analysis dictionary and synthesis dictionary in JDL model are solved by the least squares and the alternating direction multiplier method.The algorithm is an iterative algorithm,and the training model is obtained by the iterative updating of sparse coefficients,analysis dictionary and synthesis dictionary.Finally,the self learning of the JDL training model is used to suppress the sensor drift,and the relation between the old model and the new sample is balanced by a forgetting factor.The algorithm is evaluated by a sample set of three years.Set the forgetting factor is 0.5.First,the JDL prediction model is established by using the training samples of the first three months.When the prediction model begins to fail,the model is corrected by continuous online learning the sample of the new month.The average recognition rate is gradually increased from 59.63%to 87.43%,which effectively inhibits the drift and noise of the sensor.In this paper,carbon monoxide,methane and ethanol under 4 periodic temperature modulation modes,rectangular wave,triangular wave,sawtooth wave and sine wave,are identified.Using Short-time Fourier Transform to extract 11 eigenvectors below the fundamental frequency f0 and select different window functions to identify the gas.It is found that when different window functions and different frequency eigenvectors are selected,the recognition rate is obviously different.For example,for sawtooth wave,when selecting Taylor window and selecting characteristic frequency range from 3.88 to 4.85 mHz,the recognition accuracy is 100%,while the recognition accuracy of triangle wave is much lower than that of the other three kinds of modulation waveforms,the highest recognition accuracy is only 96.88%.Finally,genetic algorithm is applied to optimize the combination of 11 feature vectors,and the best feature subset is selected for gas recognition.It is found that the gas recognition rate has been significantly improved,and the recognition rate of some feature subset has reached 100%.A new adaptive temperature modulation circuit is designed to provide heating voltage for the electronic nose system.The metal oxide gas sensor is combined with a pulse waveform generation circuit.The main part of pulse generating circuit is a multivibrator that contains a 555 timer,the sensitive resistance of MOX gas sensor is added to the charge and discharge circuit,which leads to the pulse waveform be affected by the sensitive resistance.Therefore,the frequency of the output pulse waveform will change with the change of the sensitive resistance.Gabor transform is used to extract the feature vector of the sensor signal.Compared with other data processing algorithms,the proposed algorithm has obvious advantages and achieves higher recognition accuracy,and the recognition accuracy is close to 99%.