Fast neural system identification and application to adaptive processing

The design of a fast, accurate, and robust system identification method, and application to adaptive control are the main objectives of this research. This research aim is carried out through black-box modeling system identification using artificial neural networks and an optimal control method. The black-box model implicitly includes the advantage that no physical insight into plant dynamics is available or used in the process of adaptive control design. Applications can include real environments such as rigid-body robot manipulators with unknown friction, payload, backlash, and external disturbances, free-floating robot manipulators with inherently unclear mass-related properties, underwater robotic vehicles with complex hydrodynamic effects, ground vehicles for suspension control with random road conditions, microrobots with viscous effects at low speed, etc. The black-box modeling is a strategy that puts all uncertain dynamics (including even analytically can-be-known dynamics to overcome the limitation of ad-hoc control) into a black-box having only input and output measurements.; System identification based on a black-box model is carried out through a devised recurrent neural network and a suitably modified on-line training algorithm. The algorithm is fast in learning speed, accurate in identification error, and robust with respect to different plants dynamics. Although artificial neural networks (ANNs) have excellent capability, for example function approximation, input/output mapping, massive parallel processing, etc., their learning algorithms have problems of speed, accuracy and robustness generally due to the error back-propagation training strategy. In this thesis, these drawbacks are solved through a novel recurrent neural topology and modified on-line training algorithm. The developed system identification techniques show excellent performance and take into account uncertain dynamics as well as nominal dynamics simultaneously by inputs and outputs via on-line observation. The adaptive control algorithm based on the black-box identification excludes the possible use of previously can-be-known analytic information about a specific plant in order to prevent ad hoc (control) processor working for a specific system only. Therefore, a major goal of this thesis is to design a flexible adaptive control for diverse nonlinear systems and includes the development of powerful neural networks with robust connection weights training algorithms. This research gives specific attention to this aspect.; The neural system identifier, mentioned above, is combined into optimal control techniques for tracking problems of SISO to MIMO systems under the certainty-equivalence principle. Since the conventional combination of model-based identifiers and optimal techniques has been carried out using linear models, the use of nonlinear neural identifiers requires a new approach with optimization methods. The overall control scheme developed in this research is categorized as a multi-variable adaptive self-tuning control with neural identifier. The control method, supported by a neural black-box model, can cope well for nominal and uncertain dynamics with fast tracking speed and wide operating conditions. It is applicable to different nonlinear systems without changing control schemes. For a benchmark plant, the rigid-body, multi-joint, robot manipulator is used to test the concepts because it exhibits highly nonlinear, strongly coupled MIMO, and fast time-varying elements with uncertain nonlinear dynamics for the tracking purpose via pure input/output measurements.