Research On Deep Extreme Learning Machine And Its Application

As a simple and effective single-hidden layer feedforward neural network,extreme learning machine(ELM)only needs to set the number of nodes in the hidden layer of the network and randomly selects input parameters to generate a unique optimal solution.Therefore,it has the advantages of fast learning speed and good generalization performance,which has attracted much attention.However,when the data dimension is high and mixes with a large number of noise variables,ELM cannot achieve the desired effect.For this reason,the deep extreme learning machine(D-ELM)combines the advantages of deep learning to expand the hidden layer to make up for its shortcomings.In order to further improve the performance of deep extreme learning machine,this thesis proposes a deep extreme learning machine model by combining the multi-kernel learning and evolutionary computing methods,that is,constructing multi-kernel function learning machine,and then using evolutionary calculations to optimize model parameters,and the model is applied to well logging gas layer identification.The main research contents and innovations are as follows:(1)Research on deep multi-kernel extreme learning machine.Combining the multi-kernel learning method with the deep extreme learning machine,the deep multi-kernel extreme learning machine(D-MKELM)is proposed,which uses the deep network structure to extract the feature layer by layer from the input data,and the obtained features are mapped and classified in a high-dimensional space through a multi-kernel function.This not only improves the classification accuracy and generalization performance,but also solves the problem of high-dimensional data.In addition,the multi-scale kernel function is used as the multi-kernel function in the deep multi-kernel extreme learning machine,which has the advantage of good flexibility and complete scale selection.The experimental results of seven data sets in UCI database show that the proposed deep multi-kernel extreme learning machine has high classification accuracy.(2)Improvement of flower pollination algorithm.Flower pollination algorithm(FPA)is a better evolutionary calculation method,but it has the disadvantages of being easily trapped into a local optimum and slower convergence speed.To this end,a quantum system and a cloud model are introduced on the basis of FPA,and the cloud quantum flower pollination algorithm(CQFPA)is obtained,which improves the global search ability of individual flowers and accelerates the population convergence to the optimal position.The experimental verification of the classic test function shows that the CQFPA optimization effect is significant,which is superior to the commonly used evolutionary calculation algorithms.(3)Research on parameter optimization of deep multi-kernel extreme learning machine based on flower pollination algorithm.In deep multi-kernel extreme learning machine,parameters such as regularization factors,kernel parameters,and the number of neurons all affect their classification accuracy and generalization performance.To this end,CQFPA is used to optimize all parameters of the deep multi-kernel extreme learning machine,that is,to construct fitness function first,and then use CQFPA for optimization.After testing the UCI dataset,the results show that the deep multi-kernel extreme learning machine based on CQFPA is superior to commonly used classification algorithms in processing multiple classifications and complex big data.(4)Research on application of logging gas layer identification.In order to solve the problem of low accuracy of petroleum logging gas layer recognition,a gas layer recognition system based on deep multi-kernel extreme learning machine is designed,which includes the processes of log data preprocessing,identification model establishment and gas layer identification.Through the actual logging data test of an oil field,the results show that the recognition accuracy of the designed deep multi-kernel extreme learning machine recognition system is as high as 97.03%.Compared with other methods,its recognition accuracy is significantly improved.In addition,with the increase of the number of tests,CQFPA-D-MKELM has the smallest fluctuation in test accuracy,its system stability is the highest,and the area under the ROC curve(AUC)of its comprehensive evaluation diagnostic experiment value reaches 0.9071.The classification method proposed in this thesis has significant effects in gas layer recognition,which provides an effective way for gas layer recognition.