Research Of Auto-tuning Algorithm Based On Improved Gaussian Process Regression

In this paper, we will solve the large amount of computation in thetraditional GPR through two subset approximate methods; put forward anintelligent identification method based on improved Gaussian processregression without setting the model of control object beforeidentification; put forward a machine learning based on improved GPRalgorithm to solve the problem that lambda in controller parameter tuningprocess needs human to choose; put forward a method based onMcLaughlin series expansion and the Pade approximation under H performance index to solve the difficult open-loop unstable controlproblem. The main contributions lie in:(1) The main problem in traditional GPR is large amount ofcomputation and we put forward two methods based on subsetapproximation: one is hybrid algorithm based on distance index whichcombine the merits of traditional random algorithm and greedy algorithm,which can not only avoid the appearance of worst selection happening butalso improve the percentage of best selection happening. The other is analgorithm based on maximum variance index, which can lead to the finalresult contain the most information of original sample set and dismiss thedata redundancy largely. Both of the two algorithms can make aperformance improvement compared with traditional ones.(2) For the traditional identification process, we usually set the modelfor the process that we want to deal with and because of limits in methodbased on control object orders and high noise in actual process, we putforward an intelligent method based on improved GPR algorithm, whichdo not need to choose model and can do it according to the smooth curveby the GPR. The method can largely reduce the degree of humaninvolving and improve the intelligence and automation of control systemas well as save time and soar the efficiencies.(3) For the traditional control theory based on quantitative controllerparameter tuning needs human involving to choose the performancedegree lambda, we propose a method based on GPR machine learningmethod. This method gives the predictive data by learning the acquiredcorrespondence between input and output. And this method can predict the lambda value under the index of IAE with just small data set, whichcan largely reduce the complexity of human regulation and speed up theauto-tuning process.(4) For the difficult open-loop unstable system, propose a newparameter tuning method based on McLaughlin series expansion and thePade approximation under H∞performance index. This method gives anaccurate set point filter and the controller setting has a robustperformance under uncertainty in the process model parameters.Closed-loop responses tuned by the proposed method are compared withexisting methods.