Research On Improved Extreme Learning Machine Based On Denoising Autoencoder And Its Application

The Extreme Learning Machine(ELM)is a kind of machine learning algorithm designed for forward neural network.Compared with other forward neural networks learning algorithms based on gradient optimization,the ELM has the obvious advantages of good generalization performance,fast learning speed and short training time.At present,the ELM has been widely used and achieved good results in disease diagnosis,traffic sign recognition,image evaluation.However,due to the ELM input weights and bias are generated randomly,when higher dimension of sample data,the whole algorithm model is filled with a large number of random parameters,the effects of the parameters on the model of a lot of random limited even no effect,in order to ensure the classification performance,the ELM needs a lot of hidden layer nodes as support,this directly led to the extreme learning machine network structure is complex,which directly leads to the complex network structure of the extreme learning machine and the unstable classification performance.At the same time,when there is noise interference in the sample,the single hidden layer ELM can not express the characteristics of the sample data well,which has a serious impact on the effect of data classification.In addition,the ELM has a lot of application research in the field of handwritten Chinese character recognition because of its excellent characteristics,but it has rarely been involved in recognizing minority characters,especially in Xiangxi square Miao Language recognition application is still in the blank stage.In this paper,the ELM is further researched and discussed in view of the above problems and shortcomings.The main research results are as follows:1.Aiming at the problem that the extreme learning machine requires a large number of hidden layer nodes to ensure its classification performance when the input data dimension is high,a double pseudo-inverse weight determination extreme learning machine(DPELM)was proposed by further analysis and deduction of the principle of the ELM.The output weights of DPELM are determined randomly at first,and the input weights are obtained by pseudo inverse calculation.Then,the output weights are determined by pseudo inverse method again,so that the input weights and output weights are both the best calculated by analytical formula Weight,and finally the growth method is used to determine the number of neurons in the hidden layer of the network.In the performance comparison experiment with traditional ELM,it is found that:(1)The accuracy of the improved algorithm is improved;(2)when the improved algorithm reaches the best classification accuracy,the number of hidden layer neurons is reduced;(3)the stability of the improved algorithm is better.The improved algorithm is applied to the classification and recognition of breast tumors,and its diagnostic accuracy,false negative rate and time-consuming were significantly improved compared with ELM,AFSA-ELM,ELM,LVQ and BP optimized by the improved fish swarm algorithm.Experimental results show that DPELM has the advantages of fast diagnosis,high classification accuracy,and low false negative rate in breast tumor auxiliary diagnosis model.2.Aiming at the problem of the ELM's weak ability to express data features when the data contains noise or outliers,the classification accuracy is low,and the denoising autoencoder(DAE)algorithm is introduced to extract more essential characteristics of the data.The ability is integrated with the improved the ELM,and the double pseudo-inverse ELM(DAE-DPELM)algorithm based on the denoising autoencoder is proposed.The input data is extracted by DAE,and then the extracted features are used as the input data of DPELM for network training.Comparative experiments were conducted on the noisy and no-noise Fashion MNIST,MNIST,Rectangles and Convex datasets.The results show that the DAE-DPELM algorithm has the best overall performance and the number of hidden network nodes for classification is the least.3.The improved algorithm is applied to the recognition of handwriting square Miao Language in Xiangxi,and through comparative experiments with traditional ELM algorithm and DAE-ELM algorithm,it is found through detailed analysis of various experimental results that: the algorithm proposed in this paper has a significant improvement in the recognition accuracy of Xiangxi handwritten square Miao Language than the ELM and DAE-ELM algorithms,and it also fills the gap in this application field of the ELM.In summary,the DPELM algorithm proposed in this article significantly improves the classification accuracy of ELM and the stability of the results without introducing other hyperparameters,and at the same time simplifies the network structure of the algorithm model.In summary,the DPELM algorithm proposed in this article significantly improves the classification accuracy of ELM and the stability of the results without introducing other hyperparameters,and at the same time simplifies the network structure of the algorithm model.The proposed DAE-DPELM algorithm is also significantly better than the existing DAE-ELM algorithm in terms of anti-noise performance and feature extraction ability,and has achieved good results in the recognition of Xiangxi handwriting square Miao Language,which further proves the effectiveness and superiority of the improved algorithm in this paper.