Performance Analysis Of Multiple Kernel Extreme Learning Machine And Its Application In Pulse Recognition

Since 21 century,with the continuous improvement of computer performance and network technology,we are stepping into a high-speed and informational society.In particular,in the field of pattern recognition and data mining,more and more innovative research and methods have been proposed,which has created the prosperity of machine learning algorithms.Traditional machine learning algorithms,such as logical regression algorithm and BP neural network algorithm,are mostly based on the gradient descent method to find the global optimal solution.The gradient descent method usually spends a long training time adjusting a lot of parameters.At the same time,it is easy to fall into the local minimum point,but cannot achieve the expected training accuracy.Extreme learning machine is a novel machine learning algorithm,which solved the problems in traditional algorithm to a certain extent.It randomly selects the input weight and transforms the sample classification problem into a quadratic programming problem.Then,by solving a Moore-Penrose generalized inverse,the samples can be classified accurately.By introducing the kernel method to the extreme learning machine,kernel extreme learning machine is proposed,which reduces the uncertainty caused by random weights,and promotes the training accuracy of extreme learning machine in linear inseparable samples.In this paper,the experimental comparison between the kernel extreme learning machine and other improvements of extreme learning machine is carried out.The results show that the kernel extreme learning machine is superior to other improvements in the classification accuracy and training time.Now there are many feasible kernels,such as polynomial kernel and Gaussian kernel.But there is not deep research of the selection method of kernel functions.In this paper,the selection method of kernel functions in extreme learning machine is analyzed and discussed.Based on previous studies,a new wavelet kernel function is introduced,and its feasibility as a kernel function is proved.Experimentsshow that the performance of different kernel functions on different data sets is uneven,which shows that the application fields of different kernel functions are different.In order to design a more general kernel function,a variety of kernel functions are combined in different combinations,and their classification accuracy is verified by experiments.These combination methods can be divided into six types:linear combination without weights,linear combination with result-dependent weights,linear combination with model-dependent weights,nonlinear combination and the combination based on Adaboost algorithm or meta heuristic algorithms.Experiments show that the combination method based on PSO algorithm performs best on most data sets.In addition,this paper classifies the pulse data of pregnant women and non pregnant women by using the multiple kernel extreme learning machine,and verifies its practical ability.The experiments show that the multiple kernel extreme learning machine can accurately identify the pulse of pregnant women and non pregnant women,and the accuracy rate is more than 95%.