Research On Ensemble Learning Based On Artificial Olfaction System

Artificial olfactory system,also called electronic nose(E-nose),is made up with a gas sensor array with specificity,signal process unit and machine learning algorithms.E-nose has the ability to identify single or complex smells.Compared with traditional gas analysis technology,E-nose has the advantages of low cost,convenient operation,quick identification,high reliability,strong objectivity etc.With the rapid development of computational science,the theory and technology of E-nose have been gradually improved,and the qualitative or quantitative detection of complex objects has become one of the hottest research fields.In this paper,we select two kinds of typical Chinese herbal medicine,ginseng and dendrobe,as experimental objects.Then we prepare the samples,design experimental procedures,use the home-made E-nose platform to control the operations of sample testing and collect the response signals.Meanwhile,we select the appropriate data pre-treatment approaches,including noise reduction,signal filtering and baseline correction,and then extract the sample features.An optimized two-layer multi-classifiers ensemble learning based on Adaboost.M2 is constructed:In the first layer,the algebraic fusion rules are adopted to adjust weights of outputs from base classifiers,then the outputs are normalized.In the second layer of Adaboost.M2,their outputs normalized are set as the hypotheses in the process of each iteration learning.Meanwhile,the errors of the hypotheses not only modify the weighted distribution of training samples,but regulate the weights of the voting rule in the whole iteration process,in order to continuously strengthen the learning ability of the ensemble model.The diversity measurement is introduced to analyze the similarity and difference of base classifiers' performance,and the performance of ensemble model in typical iteration number is compared,so as to find the optimal combination of base classifiers for ensemble learning.As for ginseng and dendrobe identification,the optimized ensemble model is "SVM-PNN-LDA" and "MLP-PNN-LDA",with the classification accuracy of 91.75%and 87.71%,respectively.The iteration number is optimized as 30,and the mean rule is chosen for the fusion of base classifiers.The optimized two-layer Adaboost.M2 ensemble learning is a flexible tool to make a valid probabilistic and precise prediction for target recognition and not limited to the applications of E-nose.This approach also proposes an idea for ensemble system applications,supplies a feasible solution for online classification.Back to the mammalian olfactory mechanism,the K? olfactory neural network and its mathematic model are investigated.The K? network parameters are configured,and the state variable of network node is solved.The complete network is simulated to demonstrate its bionic performance.Based on the learning mechanism of K? network,the pattern recognition process is as follows:1)configuring the parameters in K?network to learn feature X from the samples;2)using the global adaptive Hebb rule to train the K? network,until the connection weights of the M1 nodes in OB layer are converged;3)taking the amplitude modulated oscillation waveforms of Ml nodes as outputs of K? network,and computing the standard deviation as characteristics;4)using classification rule to leanr the outputs from the K? network.With respect to the identification of ginseng and dendrobe samples,the workflow based on K? network is constructed,including:1)the self-organization unsupervised learning of K? network;2)feature normalization;3)supervised learning prediction model.The prediction model adopts base classifiers and corresponding ensemble model,respectively.Then the performance of various model is compared.As for ginseng and dendrobe samples,the best identification results based on K? network are 88.57%and 86.46%,respectively.