Research On Data-driven Auto-learning Feedback Control System

Data-driven feedback control methods have the characteristics of high efficiency, break traditional control methods’ reliance on accurate models, thus meeting the law of development in the Information Age. Data-driven auto-learning methods need to find useful information from the measured data to design feedback controllers. The learning process usually faces the problem about effectiveness of learning methods and convergence of relevant parameters. Since the systems have complex dynamics, it is difficult to identify accurate math models. In this case,using measured data to design feedback control become urgent issues.This thesis mainly studied auto-tuning of PID parameters and approximate dynamic programming techniques. The main work and results are as follows.1. PID auto-tuning methods were analyzed and compared, which rely on open-loop and closed-loop identification respectively. To discuss the requirements and application details, the simulation realized auto-tuning algorithm to six process model benchmarks.The impact of measurement noise on auto-tuning parameters was taken into consideration. Simulation used Butterworth analog filter to avoid the negative influence of measurement noise and achieved good tracking results.2. To deal with unknown linear system, the iterative equations in approximate dynamic programming were converted to linear systems of equations, aiming at reducing the computation. At the same time, Gaussian-Jordan elimination was used to solve the system of linear equations, thus improved the policy iteration algorithm in approximate dynamic programming. The simulation results of centralized system and distributed system both indicated the effectiveness of the proposed method. The iteration numbers deceased while the probing signal was also reduced, which implies smoother state trajectories.3. A Field Oriented Control(FOC) alternating current governor system was chosen as the platform to test the improved approximate dynamic programming algorithm. Input and output data sequences were collected respectively. The experiment began with datapre-processing, followed by two-order state space model identification of FOC system.The identified model was then used as controlled object by learning algorithms to tune feedback gain which will be used in close loop control. The output revolving speed of asynchronous motor with controller parameters from data-driven learning algorithms had smoother response and less overshoot, as compared with other control parameters,especially in low velocity..