Research On Nolinear Time Series Modeling And Prediction Based On Random Project Neural Networks

Nolinear time series modeling and prediction is an important research direction for data driven control of complex systems.At present,it has been widely used in industrial system failure analysis and prediction,industrial process control optimization,financial market data prediction,the performance testing of aircraft development process,the remaining useful life prediction and other fields.By analyzing the complex system and establishing a corresponding multivariate and nolinear time series forecasting model,people can make a deeper understanding of the internal characteristics of the system and can better achieve system control and decision-making.Random project neural network is a kind of neural network with faster convergence speed,global optimal solution,convenience of learning and other advantages have achieved good performance in the prediction of nonlinear time series.Therefore,for improving the performance of nonlinear time series prediction,this thesis sdudies on two kinds of random project neural network,echo state network and extreme learning machine,optimizes the structure of extreme learning machine,improves the robustness of random project neural network,and establishes two optimized combination models to enhance the model predictive accuracy and robustness.Finally,the improved random mapping neural network is applied to the prediction of NC machining time series.Based on the prediction results and the particle swarm algorithm,the numerical control processing parameters are optimized to fully utilize the CNC machine tools and improve the efficiency of CNC machining.The main research results include the following four aspects:The revised regularization extreme learning machine prediction model framework is proposed.In view of the problem that the number of hidden layer nodes for extreme learning machines is not easy to choose and it is easy to generate redundant information and overfitting problems after high-dimensional space mapping of the time series by the extreme learning machine.This thesis studies on the structure of the extreme learning machine and optimizes the network structure by obtaining the sparse solution of the extreme learning machine output weights.In this paper,a modified regularization extreme machine learning forecasting model framework is proposed by modifying and optimizing the regularization method based on ₁L norm,₁L and ₂L mixed norm.The proposed model framework inherits the ability of variable selection based on ₁L norm and it effectively avoids the problem that the biased estimation of ₁L norm leads to a low prediction accuracy of the model.On the basis of obtaining the model's sparse solution,the network structure is optimized and the prediction accuracy of the model is improved.The robust variational echo state network prediction method is proposed.For real complex system applications,the data is often contaminated by a variety of noise and abnormal points,and the sensitivity of different probability distributions to abnormal points is analyzed.Finally,the Gaussian mixture distribution is selected as the model output likelihood function.The variational inference procedure is utilized to handle the marginal likelihood function of the model output which is analytically intractable for the mixture distribution and the echo state network output weights are achieved.A multivariate and nolinear time series prediction model robust to outliers is proposed.The proposed model not only has strong non-linear approximation ability,but also has strong robustness against outliers.The two multiple random project neural network combinatorial optimization prediction models are proposed.Considering the problem that using a single echo state network or extreme learning machine is difficult to describe the data information adequately,an improved multiple kernel extreme learning machine is proposed based on Adaboost.RT,in addition,the variational optimized multiple sparse echo state network prediction model is achieved by learning the sparse weights of the related ESN and the sparse weights of related basis functions determined by related sample simultaneously.There is no need of selecting the spectral radius and sparsity of ESN by cross validation in the proposed model and only the interval of spectral radius and sparsity are needed to be determined.The prediction results show that the proposed combination optimization models have better adaptability and higher prediction accuracy.The time series from feed system prediction of numerical control machining prediction models based on random mapping neural network are proposed.We apply the models proposed in the second,third and fourth chapters to the prediction of the time series of the CNC machine tool feed system.The performance of feed system time series prediction of models is improved from the three aspects of prediction efficiency,prediction robustness and prediction accuracy,and the effective prediction of the speed response time series of feed system is realized,laying the foundation for subsequent error compensation.Specifically the following aspects are included:firstly,a time-series prediction model of feed system based on modified regularized extreme learning is proposed.Under the premise of ensuring prediction accuracy,the sparse solution of output weights is obtained,which greatly improves the generalization performance of the model;secondly,the time series prediction model of feed system based on robust echo state network is proposed to improve the stability of the prediction model.Again,a time series prediction model of the feed system based on PSO-RVESN is proposed,which further improves the prediction accuracy of the time series of the model prediction feed system.