Spectrum-based Rapid Gas Identification Research

An Electronic Nose(EN)is a device that mimics the human olfaction to recognize different odors,also called artificial olfactory system.Through advances in sensor design,software innovations and micro circuitry design,recently EN has achieved wide applications in agriculture,medical field and manufacture attracting increasing attentions of many researchers.The basic EN is composed of a sensor array,a feature extraction unit and a pattern recognition unit,wherein the sensor array consists of several cross-selective sensors which has different response patterns for different gases.The core idea of gas identification of EN is to utilize the pattern recognition algorithm to train gas features which extracted from these response patterns.Taking the application of rapid gas detection as a background,this thesis studied the influence of features and pattern recognition algorithms on the recognition,and proposing an optimization on the detection speed and accuracy.In the selection of pattern recognition algorithms,the Support Vector Machine(SVM)now attracts increasing attention in gas classification due to its high performance towards small samples and nonlinearity problems of the dataset.Previously,the probable mismatch between the dataset and the training parameters determined by trial and error or grid search may hinder the exploration of the best result.In this paper,we propose a novel approach to estimate the most suitable training parameters,based on the inbreeding prevention of genetic algorithm(GA)by assigning the training model parameters of SVM as its chromosome.Treating the k-fold cross validation of SVM training as the objective function,our new method makes the population on the whole evolve towards the values that are more appropriate for the dataset.The inbreeding prevention mechanism(IPM)can protect the population from converging over-rapidly before reaching the optimum value.Compared with the standard SVM,the proposed method has greatly improved the prediction accuracy in both training data and testing data.As for the feature extraction,the traditional gas detectors favor toward the steady-state feature as a result of its high recognition rate,which indicates the steady-state feature carrying affluent distinction information.However,the steady-state feature cannot be extracted until the response stepping into the saturation,which usually taking a few tens of seconds for metal-oxide gas sensors.The awkward waiting time extremely limit the application of the odor identification,such as the fast detection of the toxic.Actually,for the substitute of the steady-state feature,the researches on the transient feature which is extracted from the dynamic response have been discussed.The famous method is using Exponential Moving Average(EMA)feature which relying on the compromise between the reaction speed and amplitude,can save about 50%time consuming with keeping a high recognition rate.Inspired by the EMA,we attempt a novel feature extraction based on spectrum,by the technique of integrating linear discriminant analysis(LDA)with Fast Fourier Transform(FFT)making the wait time shrink by 90% referring to the overall transient period(100%).Besides,a frequency filtration trick to improve the robustness against drift is also introduced in this topic.