Applications of Neural Adaptive Control Technology

The following sections are included:

• Preface

• Contents

This chapter is a comprehensive description of the work-to-date done on the Neural Adaptive Control Technology (NACT) project, including a guide to the main publications. The project rationale, structure and workplan are given to provide the context of the work. Also, the engineering background is sketched with a brief outline of the industrial problem to which the NACT techniques are being applied (modelling of the longitudinal dynamics and speed control of an experimental vehicle). There are four technical sections. First, the Fourier-Analysis-based neural networks are described which arise from the multi-dimensional (N-D) non-uniform sampling approach; this is a modelling methodology. Then adaptive networks of local models and controllers are presented; the framework covers both modeling and controller design. This is followed by an account of geometric nonlinear control based on neural models; the approach focuses on the control aspects, assuming the availability of a neural model. Finally, some recent advances in recurrent networks based control are described; the emphasis is on theoretical issues related to robust stability. The chapter ends with conclusions, where some indications of progress on experimental results are also given.

This paper gives an overview of feedback control of feedback linearizable nonlinear systems. It describes adaptive and neural control techniques that can cope with model uncertainty.

In this paper, a control aspect of the non-acyclic FMS scheduling problem is considered. Based on a dynamic neural network model derived herein, an adaptive, continuous time neural network controller is constructed. The actual dispatching times are determined from the continuous control input discretization. The controller is capable of driving system production to the required demand and guaranteeing system stability and boundedness of all signals in the closed loop system. Modeling errors and discretization effects are taken into account thus rendering the, controller robust. A case study demonstrates the efficiency of the proposed technique.

It is rigorously shown that inverse-dynamics models can be used to stabilize plants of any order provided that the inverse-dynamic model is used in a mixed mode fashion, in that of a ‘Static and Dynamic’ State-feedback (SDS) mode. When the resulting controller is used for tracking increasing the gain of the dynamic feedback decreases the tracking error. Yet another attractive feature of the SDS scheme is that the inverse-dynamics model can be tuned on-line by any adaptation mechanism without cancelling stability if the conditions of the non-adaptive stability theorem hold at any time instant. Computer simulations of the control of a chaotic bioreactor and a ‘realistic’ robotic manipulator demonstrate the robustness of the approach. It is shown that SDS control will yield zero asymptotic error when controlling the bioreactor using an inverse-dynamics model which when used in a traditional mode would yield intolerably large errors. In the case of the robotic arm simulations the effects of perturbation and sampling frequency are investigated and the SDS control is compared with the non-adaptive computed torque method. A fully self-organizing associative neural network architecture that can be used to approximate the inverse-dynamics in the form of a Position-and-Directionto-Action (PDA) map is also described. Similarities between the basal ganglia – thalamocortical loops and the SDS scheme are discussed and it is argued that the SDS scheme could be viewed as a model of higher order motor functions of these areas.

The goal of integrating sensors into robot motion planning has incited recent research efforts. The Perceptual Control Manifold serves this goal extending the notion of the robot configuration space to include sensor space. In this paper, we develop a framework for sensor-based motion planning of robotic manipulators using the Topology Representing Network algorithm to develop a learned representation of the Perceptual Control Manifold. The topology preserving features of the neural network lend themselves to yield, after learning, a diffusion-based path planning strategy for flexible obstacle avoidance. We demonstrate the capabilities of topology preserving maps using an industrial robot simulator and a pneumatically driven robot arm (SoftArm).

Globally feed forward locally recurrent neural networks possess a high potential for real-world dynamic and nonlinear process identification. This chapter describes the basic model of the dynamic multi-layer perceptron network (DMLP). The network consists of perceptrons, each of which embodies a long-term memory. Both the topology and the parameter estimation procedure of DMLP are introduced. The second part of the chapter addresses the application of the DMLP to the experimental identification of a multi-input, nonlinear model for a heat exchanger used in batch reactors. Based on these models a predictive control scheme is designed in order to control the water outlet temperature of an industrial scale thermal pilot plant which is characterized by a significant static nonlinearity and dead times depending on the operating set point.

This paper extends earlier results on adaptive control and estimation of nonlinear systems using Gaussian radial basis function networks to the on-line generation of irregular sampled networks. The network is utilised in stable adaptive tracking control designs employing a parallel combination of a neural network, a PDcontroller and a sliding-mode component.¹ The Gaussian basis functions are used to determine the required functional dependence in real-time, by approximating the unknown nonlinear functions in the control law. To combat the computational complexity faced in higher dimensions by reducing the required network size, a method is suggested that allows adaptive adjustment of the shape and location of the individual Gaussian basis functions. By adjusting the basis functions according to local frequency characteristics much smaller networks are needed, resulting in a possible solution to high dimensional problems. The development of theory is focused on n joint robotic manipulators, but all deductions are applicable to a much wider range of dynamical systems. Simulations and experimental evaluation reflect that very promising results are achieved in reducing the required network size, while still maintaining superior control performance.

Increased performance and cost reduction are important reasons for the use of new control strategies in the field of industrial process automation. The approach of describing technical processes with artificial neural networks forms an integral part of the fast multiprocessor system Logidyn D2. The use of this technique offers a number of advantages, particularly in conjunction with the modeling of systems which are difficult to describe by analytical means, as well as non-linear and time-variant systems.

Some of the basic preconditions for the following aspects of the use of this technique will be described: user acceptance, functional safety, as well as simplicity of design, engineering, commissioning and maintenance.

Today's control of Four-Quadrant Converters (4QCs) used for power supply in modem AC rail vehicles rely on linear models even though the system exhibits nonlinear behaviour. One nonlinearity is caused by the 4QC's blanking time which is required for safe operation. This effect and other nonlinear phenomena deteriorates the shape of the line current which is requested to be sinusoidal. Limitations in precise analytical modelling of the nonlinear phenomena suggests the application of a neural network to improve the line current. It is shown, that some of the nonlinear effects belong to a certain class of disturbances. For this class of disturbances, a novel control strategy,called vertical control, is presented, in which a neural network may be integrated. Simulation results are presented and the usage of neural networks in 4QC control is discussed.

The following sections are included:

• Index