Research On Wavelet Neural Network And Its Application In PID Controller Parameters Tuning

As production processes become more complex,the research process often encounter the situation that global mathematical model of the controlled system is completely unknown or quite uncertainty,or the controlled process is so complex that it is difficult to utilize a mathematical model to describe and establish an accurate model of the controlled process.The lack of precise mathematical mode leads that the stabilization of the control effect based on the control model cannot be guaranteed.One of the present control problems in the practical application is how to reduce the dependence of the control system on the mathematical model of controlled object effectively while maintaining the control efficiency.PID controller is a model-free controller with simple mode and algorithm.Thus,the class of PID controller dominates in the field of industrial process control over the past several decades.As one of the parameters tuning method of PID,the wavelet neural network is a novel neural network on the basis of wavelet analysis.The wavelet network is combined with the advantages of wavelet analysis and has a strong ability in learning.However,during the update process,wavelet network remains a descent gradient adjusting weights conventional algorithm,and has the problem of emphasis on learning to overcome the error that leads to the weak generalization performance.Based on the wavelet neural network,this thesis combines the algorithm mechanism and controller parameter characteristics,and discusses the parameters tuning method in the following aspects:The merits of randomly generated parameters in weights initialization will affect the number of learning networks,and even lead to the results of non-convergence and other issues.A novel initialization method is proposed based on nonlinear filtering and state estimation.Wavelet network learning and training could be considered as the best estimate for finding the best weight parameters,and the weight matrix and network output values are treated as state quantity and measurement values.The training process is described by state-space description.Combining with the state estimation,A weight initialization method based on state estimation is proposed associated with weights,network input,wavelet function types and theoretical output.The constant step size in wavelet neural network would limit the track velocity of random gradient during the weights iteration updating.By discussing the comparison among wavelet network,least mean square algorithm(LMS)and normalized least mean square algorithm(NLMS),a variable step size algorithm based on improved NLMS algorithm is put forward to improve the convergence of wavelet networks.Firstly,calculate the step function of wavelet derivative function with segmentation process,and make a certain improvement on the basis of the NLMS algorithm in variable step size.The calculation formula of wavelet network step size is derived layer by layer,and the convergence of the algorithm is analyzed.Since kernel method is successfully applied to the nonlinear adaptive filter algorithm and RBF neural network algorithm,but Gaussian kernel function could not generates an ideal complete set of questions based on the subspace in the RBF network and SVM.Thus,this thesis proposed a novel neural network based on wavelet kernel function.At the same time on the basis of wavelet kernel function,by the analysis of parameters and fractional order PID(FOPID)controller and PID controller respectively,the enhanced FOPID is deduced based on the mathematical description and a calculation form controller of FOPID is given.PID controller also proposed wavelet kernel function and FOPID controller parameters tuning method.Some gradient information in the process of controller tuning with neural network is unknown.Thus,a method on calculating the optimal parameters in the unknown objective function is proposed on the basis of stochastic approximation theory.Therefore,this chapter presents a novel FOPID tuning method,that two different parameters can be targeted regulation.When the error changes slightly,it only utilizes neural network for tuning finely.But when the error changes acutely,the fuzzy theory would be introduced for order setting.For the sharp contradictions between the complexity of system and the requirement of control performance,the existing theory and technic based on quantitative mathematical model could not control the complex industrial systems.Thus,this thesis proposes a novel coordinated control strategy and conducts several experiments considering the integrated pressurized water reactor as the control object.The simulation experiments verify the validity and reliability of the coordinated control strategy and the FOPID control based on SPSA kernel wavelet network.