Compared with some existing identification methods such as traditional leastsquares, stochastic gradient algorithm that uses the single-innovation modificationtechnology , the multi-innovation identification method possesses betterconvergent performance. Therefore, it is theoretically and practically important toinvestigate such a kind of identification method. The main results of this thesis areas follows.1. By introducing a weighting matrix into the multi-innovation identificationmethod, this thesis presents weighted multi-innovation identificationgradient identification algorithm. By some numerical simulation examplesit is shown that through selecting proper weighting matrix, weightedmulti-innovation identification methods have faster convergence speedrate and better parameter estimation accuracy than ordinary multi-innovation identification algorithm under the same innovation length. Inaddition, some brief discussions on the effect of the weighting matrix aregiven.2. For time-varying systems, weighted multi-innovation identificationmethods with forgetting factors are presented. The effectiveness of theproposed algorithm is verified by a numerical example.3. For the controlled autoregressive moving average model (CARMA model)with colored noise, in this thesis, the multi-innovation augmentedstochastic gradient identification algorithm is proposed, the advantagesover some existed algorithms are shown by some numerical examples.4. For a more complex system model, dynamic adjusting model (DA model),the thesis proposes multi-innovation generalized augmented stochasticgradient identification. Some numerical examples are given to verify theweighted multi-innovation generalized augmented stochastic gradient identification method has faster convergence rate than the regular multi-innovation generalized augmented stochastic gradient identificationmethod.
|