Algorithm Research On Classification And Regression Based On Wavelet Transform And Neural Networks

Time-series data is ubiquitous in our life,and it is easy to find them widely in medicine,aerospace,finance,business,meteorology,industry and other aspects.The classification and regression of time series data are widely used in life and have been applied to finance,entertainment,medicine and industry.Wavelet transform is a method based on multi-resolution analysis,it decomposes the input data into different scale space,and also has very good time-frequency decomposition characteristics in different scale space.Therefore,wavelet transform can extract marked features on various scales and frequency characteristic and has great advantage in dealing with series signals.This paper presents a classification and regression algorithm based on fast wavelet transform.The method that dealing with Time-series data mainly consists of two modules:the multi-scale expression and combination of input data,WTNN module.The first module is used to get the multi-scale expression of input data and the second module is consists of two submodules:feature extraction submodule,classification and regression submodule.It is necessary to point out that both submodules are integrated into a network.In the first submodule,in order to facilitate classification and regression,the wavelet transform is used to extract multi-scale and multi-frequency features of time series signals.In addition,the original wavelet transform need discrete different scales and the basis functions,which is inconvenient.What's more,fast wavelet transform is facilitate to discrete different scales,different frequency and different wavelet basis functions.And fast wavelet transform enable the discretization and the design of basis functions more flexible.In wavelet transform,the shape of the basis function for the wavelet filtering and scale filtering plays an important role.Based on the idea of randomly assigne weights,this paper randomly generated basis function,not only reflact the randomly connection between neurons,but also implement the diversity of the basis functions.And thus can extract markable features in large number of random scale and frequency characteristic.In this paper,in order to realize the multi-scale of input signal,the original time scale of input data was divided into different time scales and then combine these different time scales.Finally these different linear combinations were used as input features.The second submodule,in order to improve the generalization ability of the algorithm,the objective function is minimize the training error and the output weight.In a word,the network can not only effectively deal with timing signal,but also with strong generalization ability and fast learning speed.This paper applies the above algorithm to the solar data and time series data set of UCI.In solar power dataset,the minimum MAE is 24.5588(kW),which is far smaller than contrast methods.In UCI dataset,the accuracy of Drift is 10%higher than compare method and the maximum accuracy of Occupanyis 98.88%.The experimental results show that wavelet based neural network for classification and regression has a higher recognition accuracy or smaller error,the algorithm we put up performs better.