| This work first presents a complete investigation of several time series identification algorithms. Their relationships and similarities are clarified and some selection guides are provided among these algorithms which are essentially of the least-squares (LS) type and only their numerical solutions are different. These algorithms are known as the batch and sequential LS, the covariance and autocorrelation algorithms, the PARCOR (partial correlation) method, the Lattice or the Ladder algorithms, the square-root algorithms and the equation error algorithm. This study further examines their performances in terms of convergence, convergence-rate, computational effort (speed) per iteration, and sensitivity or robustness to computational errors in finite word-length microprocessors.; Then, two microprocessor based systems in which signal processing is applied to the biomedical field are presented: (1) Speech recognition system for paraplegics. (2) Multifunctional control of prosthesis and orthosis systems for above-elbow amputees and partially paralyzed victims (hemiplegics). Both systems are based on the identification of biologically generated signals and the recognition of such patterns with the aid of microprocessor hardware as well as software.; The main goal of this speech recognition system is to provide mobility and some control over a designed environment. Mobility is provided by controlled wheelchairs, while environmental control usually allows the person to activate a phone, typewriter, television and radio or some special devices associated with the disabled person's vocation. The present system employs identification of the parameters of a linear autoregressive (AR) time series model for the speech signals alternatively with a classification algorithm so as to recognize the spoken word. For the rehabilitation applications, especially those associated with mobility control, a single speaker with limited dictionary and isolated words is implemented on a 16-bit microprocessor system.; The second system discriminates among different limb functions by utilizing the complete spectral content of the myoelectrical (ME) signals picked up from one or two surface electrode pairs, on stochastic time series analysis basis, in order to provide multifunctional prosthesis and powered braces control for high level amputees and partially paralyzed persons. The procedure requires a sequential LS identification algorithm which fits a fourth order AR model of the EMG signal and a recognition process based on a Bayesian rule discrimination technique. This system has been tested on more than 12 amputees and a few partially paralyzed victims with total success after only few hours of training.; This work describes significant improvements in the performance, discrimination and speed of these microprocessor based systems in recognizing the spoken command words for the severely disabled and in actuating an artificial limb or a powered brace device for amputees and paralyzed victims. Also, test results of a ten digit vocabulary used to activate a touch-tone phone, and the actuating of four limb functions by means of various changes in the software, hardware and training procedure are given. |
