In this paper, a nonlinear control system incorporating a multi-layered neural network (MNN) with the Self-Tuning Regulator (STR) is designed. The MNN is used to compensate nonlinearity of plant dynamics that is not taken into consideration in the STR design. The control performance is thus improved as compared to the STR alone. Also, the learning time required for convergence and the network size of the MNN can be reduced as compared to conventional MNN based control systems. Because the STR quickly learns the linear part of plant dynamics based upon the least square method. Furthermore, the MNN can be trained on-line with the specialized learning architecture. Because the sensitivity of the system output to the changes of the input is calculated by using plant parameters estimated by the STR. Computer simulations are carried out to show the effectiveness of the proposed method.
SRIF (Square Root Information Filter) is a method of solving for optimum estimate in the least square sense. SRIF finds optimum estimate by updating a square root matrix of an information matrix. In the paper, a new VLSI systolic array for SRIF parameter estimation is proposed. SRIF parameter estimation consists of orthogonal transformation and solving linear equations. The method that solves SRIF parameter estimation by use of Givens rotation and back substitution was proposed. But it is well known that complete pipelining of these two separate algorithms is impossible. To avoid this problem, we propose a new method of using the Faddeeva algorithm instead of back substitution. The Faddeeva algorithm is performed by Gauss elimination. Pipelining Givens rotation and Gauss elimination is now feasible as parallel architecture. Making use of this method, we constitute a high-speed VLSI systolic array for SRIF parameter estimation than a conventional one.
We propose a method of controller reduction, which aims to replace an ideal high-order controller with a reduced-order one so that finite numbers of 1st-and 2nd-order information of the closed-loop system can be preserved. The preserved 1st-and 2nd-order information are those associated with the transfer characteristics of the closed-loop system and its power spectrum characteristics at appropriate complex frequency points, respectively. The proposed method provides the reduced-order controller which yields a frequency-weighted approximation of the closed-loop system considering its stability.
Many communication signals exhibit cyclostationary property. Recently, Gardner and Chen introduced the estimate of the time difference of arrival (TDOA) of a cyclostationary signal at two separate sensors by exploiting the cyclic spectrum of this signal. But their method uses only a single slice from the cyclic spectrum. In this paper, we derive the maximum likehood estimator of the TDOA, using the stationary representation of a cyclostationary signal. We estimate the cyclic spectrum by the multivariate AR model fitting. As this method uses all parts of the cyclic spectrum, the performance of this method is better than the previous ones. This was clarified by the results of the computer simulation.
This paper discusses signal characteristics of a two-dimensional image transformed into frequency domain by using an adaptive digital filter (ADF). System identification using ADF is applied to the image data analysis in each frequency area. ADF determines the system parameters of the image signal. An adaptive Kalman filter which uses the system parameters adaptively obtained from ADF is proposed to restore a noise-contaminated image. The adaptive Kalman filter with the system parameters proposed in this paper contributes the image restoration. A smoothing image by an ε-filter is used to identify the system parameters in order to apply the adaptive Kalman filter. The ε-filter combined with the adaptive Kalman filters is proposed to improve the performance of the noise reduction of the image. Experimental results show that the proposed method has advantages for image restoration.
This paper considers the finite-horizon discrete-time minimax estimation problems related to the H∞ estimation cost. Based on the Lagrange multiplier technique, we show that the minimax estimators are identical to H∞ estimators in both filtering and onestep prediction cases. Moreover, necessary and sufficient conditions for the existence of the minimax estimators are given in terms of an H∞ type Riccati difference equation (RDE) satisfying the positive definiteness of certain matrices. Using the RDE, we compare the minimax estimators with Kalman filter, and investigate the some properties of the minimax estimators. A numerical example is also included to show the applicability of the present minimax estimators.
With the aid of the state variable method, parameters of mechanical structure systems can be identified from the free decay responses. To measure the free decay responses, however, we must use a multi-channel data acquisition system. This paper concerns the problem of precise identification, where the differences in dynamics of the multi-channels of a data acquisition system have been considered. Firstly, we investigate the influence of the differences on the identification results. Secondly, in order to eliminate the influence, we apply the relative dynamic compensation to the identification problem. Finally, we give a simulated example to show its feasibility.