Stochastic Gradient Self-Tuning Control Based On Multi-Innovation Parameter Estimation

Traditional identification methods-least squares (LS) and stochastic gradient (SG) algorithms all adopt the methods of single-innovation correcting technique. Multi-innovation identification methods, the extending of single-innovation ones, hold good convergence properties and ability of overcoming bad data, the study of which has both important theories meaning, and latent of applied value. The adaptive control methods on hand aim at the control systems with unknown parameters, which are estimated by least squares or stochastic gradient algorithms, and then design the controller. The main contribution is that the unknown parameters are estimated by using multi-innovation identification methods to study adaptive control methods based on multi-innovation parameter estimation. This thesis study out the subject of stochastic gradient self-tuning control based on multi-innovation parameter estimation. The subject has theoretical significance and application. In checking the foundation of related cultural heritage, the author studied the topic put forward, obtaining the following research result.1. The thesis briefly introduces the basic idea and identification theory about multi-innovation identification methods based on making some surveys on adaptive control methods, then multi-innovation identification is combined with self-tuning control methods for the the linear controlled autoregression (ARX) model with the circumstances of white noise, stochastic gradient projection self-tuning control and stochastic gradient self-tuning control based on multi-innovation parameter estimation algorithms are presented by the usage of multi-innovation stochastic gradient projection algorithms and multi-innovation stochastic gradient algorithms, for more stochastic gradient pole assignment self-tuning control based on multi-innovation parameter estimation algorithms are presented by the usage of multi-innovation stochastic gradient algorithms and pole assignment self-tuning control strategy. The contrasts of computer simulation are studied by using martingale convergence theorem to analyze the stability and convergence of closed-loop systems. It can be seen that the algorithms of stochastic gradient self-tuning control based on multi-innovation parameter estimation are better than the ones of stochastic gradient self-tuning control. And the larger the innovation length is, the faster the control algorithms' convergence speed is, the better the tracking performance is.2. For the controlled autoregression moving average (CARMA) model, stochastic gradient self-tuning control based on multi-innovation parameter estimation algorithms stochastic gradient pole assignment self-tuning control based on multi-innovation parameter estimation algorithms are presented, the global stability of the closed-loop systems and the bounded-ness of output tracking errors are studied. Simulation examples show that the convergence speed of the algorithms is faster by the innovation length increasing.3. Finally identification model of the nonlinear system Hammerstein controlled autoregression (HARX) model is derived, then estimate the parameters of the nonlinear Hammerstein system by the methods of multi-innovation identification, stochastic gradient self-tuning control based on multi-innovation parameter estimation algorithms stochastic gradient pole assignment self-tuning control based on multi-innovation parameter estimation algorithms are presented. The simulation seems that the convergence speed of the self-tuning control based on multi-innovation algorithms for the nonlinear systems is faster than the one of the self-tuning control based on single-innovation algorithms.Theoretical analysis and digital simulation results show that compared with usual stochastic gradient self-tuning control algorithms, the algorithms of stochastic gradient self-tuning control based on multi-innovation parameter estimation have faster convergence speed.