Research On Data-driven Control Method Based On Stochastic Approximation

It has become increasingly difficult to establish the precise mathematical model according to physical and chemical mechanism of the production process and equipment, and subsequently control and give some prediction and evaluation over it. Data-driven control is a new kind of control theory and method directly from the data to the controller design. It will free the controller from the traditional model control mode and effectively solve the problem of dependence on control object model and unmodeled dynamics.From study on a variety of typical data-driven control theories and methods, this dissertation is trying to draw out their common basis theory under the framework of gradient estimation algorithm and controller parameter identification and integrate them into the unified data-driven control system based on the stochastic approximation. This dissertation focuses on the following directions for the in-depth and systematic research and made some achievements.1. Two kinds of typical stochastic approximation algorithm are analyzed: the gradient-based stochastic approximation algorithm RMSA and the gradient-free stochastic approximation algorithm KWSA. Extending the problem of finding minima of functions to control, Simultaneous Perturbation Stochastic Approximation algorithm SPSA are studied in this dissertation, including basic SPSA,one-measurement SPSA、second-order adaptive SPSA (ASPSA) and Modified ASPSA(MASPSA). Studies have found existing SPSA algorithms are not suitable for control problem since SPSA in control problem pay more emphasis on the real-time effectiveness and should balance convergence accuracy and convergence rate in control application. After assessing disadvantage of the basic SPSA and other modified algorithms, an efficient one-sided SPSA method is proposed which uses less measurement of the cost function than conventional SPSA in control and can improve convergence rate and accuracy of optimization procedure.2. A data-driven control method based on SPSA is developed while plant models are unknown. Two key issues in SPSA-based data-driven control are presented and then solved: data utility scheme and control-oriented SPSA method. Data information content and effectiveness is related to control quality in data-driven control, how to choose the measurement data to generate control law is of vital importance. Combining one-sided SPSA and data-driven control method, the data utility scheme based on dynamic historical errors is designed, and the connection weights in Neural Network function approximator are estimated while the system is being controlled by direct adaptive control approach. Therefore the proposed data-driven control method leads to better performance in nonlinear plant control without any plant model information.3. In this dissertation Iterative feedback tuning (IFT) is classified into two divisions: totally model-free IFT algorithm and model-related IFT, both being studied respectively. Emphasized Study is given on the IFT application in the large time-delay system. Combined with Smith prediction and the inner model control structure, a novel data-driven direct control method based on IFT is put forward in this paper. This method puts additional emphasis on anti-delay Focusing on process of large time delay characteristics: a necessary performance criterion which penalizes the predictive error is designed. A novel step size descending algorithm with a rate regulation coefficient and a time-weighted coefficient - variable mask factor are proposed respectively, both under adjustment according to the actual number of iterations to ensure the stability of system. Studies demonstrate that the convergence and stability of the algorithm outperform classical IFT in anti-big-delay application with fewer requirements to the initial parameters.4. Unlike the number of experiments in single IFT iteration is conventionally determined by number of controller parameters, finding the minimum IFT experiment number in the premise of ensuring gradient estimation unbiased is proposed in this paper. Three specially designed closed-loop experiments are carried out to perform the gradient descent optimization of the cost function which is computed totally from I/O data in the close-loop system. Thus the parameters of controller and predictor are autotuned with an unknown plant. At the same time it is effectively reduce the amount of data required for iteration and accelerate controller optimization speed, and is expected to play a key role in online IFT application.5. The data-driven control method based on Virtual reference feedback tuning (VRFT)is also investigated in this dissertation. The paper deeply analyzed the inner equivalence of virtual reference feedback tuning and internal model control scheme, and then VRFT is creatively adopted into the IMC-PID controller tuning. Consequently a novel VRFT-IMC tuning method without using model parameters is presented in this paper. This method combines the advantages of strong robustness in IMC and model-free adaptiveness in data-driven control. Simulation results show that IMC-PID method based on VRFT outperforms the traditional IMC-PID, and is proposed to be used either for an initial tuning of the controller or for auto-tuning in complicated control system.6. Virtual Reference Feedback Tuning (VRFT) is an offline data-driven controller parameters tuning method, and it needs the process model unchanged during tuning. To overcome this problem an online VRFT method was proposed in this paper. An effective filter was introduced to shift time-sequence in offline algorithm in order to obtain real-time computable data. Secondly it presented an online VRFT algorithm based on Recursive Least Square Method (RLSM) with forgetting factor. This method used only real-time data to get optimal controller online with no system information and identification. Simulation results show that the proposed method is self-adaptive and superior to conventional VRFT when dealing with time-variant system.Finally, the research and the innovation of this dissertation are summarized, and further suggestions for the research along the directions in this dissertation are proposed.