Jeffrey J. Garside

Modeling torque in a switched reluctance motor for adaptive control purposes using self-organizing neural networks

Training paradigms for topology-preserving Kohonen neural networks are introduced for the purpose of identifying and controlling nonlinear systems. A procedure for locking neuron weights at specific locations in a region is presented. It exploits prior knowledge about the system of interest. As a result, superior representations of an arbitrary multivariable nonlinear mapping can be achieved. In addition, the common problem of twisted meshes in these neural networks is eliminated. The strategy introduced for preferentially training these networks at region boundaries overcomes the limitation of boundary contraction. As an example, a one-dimensional neural network is used to approximate a nonlinear function, although in general an n-dimensional mapping can be used to approximate an m-dimensional system for n<or=m. As a practical implementation, the modeling of the theoretical torque of a switched reluctance motor (SRM) as a function of position and current is presented. The topological torque representation is suitable for adaptive control of SRMs in high-performance applications.<<ETX>>

An algebraic approach to prediction of the performance of PM electric machines during sustained fault conditions

A computer aided method for predicting the performance of three-phase permanent magnet machines during normal and sustained fault conditions is presented. An approach based on Fourier series signal representations of the machine parameters and waveforms is used to determine the sustained fault currents and voltages. Using a model in the abc reference frame, the unknown but periodic currents and/or voltages are modeled as Fourier series. With the constant velocity assumption, the derivatives of the inductances can be calculated, and derivatives of the currents can be calculated as functions of the unknown current Fourier series coefficients. Thus the model equations become algebraic and hence solvable by algebraic means.

Kalman based artificial neural network training algorithms for nonlinear system identification

The utility of artificial neural networks (ANNs) in nonlinear system identification and control is intimately linked with the ability to parameterize the ANN structure on the basis experimental observations. Four existing training algorithms are reviewed under a parameter estimation framework, and the method of target state backpropagation previously proposed by the authors is extended. The new algorithm follows a different approach to the generation of error signals in embedded layers by backpropagating target or desired states rather than partial derivatives. The target states are used in conjunction with a linear Kalman based update algorithm, and transients associated with initial conditions are eliminated through a time-varying method of covariance modification. Comparisons of the five algorithms are made through a system identification problem, and the error convergence associated with each algorithm versus actual training time is presented. The results demonstrate an increased rate of convergence in comparison with backpropagation.<<ETX>>

Nonlinear estimation of torque in switched reluctance motors using grid locking and preferential training techniques on self-organizing neural networks

The torque of a switched reluctance motor (SRM) can be estimated using a topology-preserving self-organizing neural network map. Since self-organizing maps tend to contract at region boundaries, a procedure for locking neuron weights at specific locations in a region is presented. A strategy for preferentially training neuron weights on the region boundaries is introduced. As an example of these training techniques, a one-dimensional neural network will approximate a nonlinear function. In general an n-dimension mapping can be used to approximate an m-dimensional system for n<or=m. As a practical implementation of this technique, the modeling of the theoretical torque of a SRM as a function of position and current is presented. A two-dimensional neural network estimates a three-dimensional highly nonlinear surface.<<ETX>>

A parameter estimation approach to artificial neural network weight selection for nonlinear system identification

A unified framework for artificial neural network (ANN) training algorithms applied to nonlinear system identification based on considering weight selection as a parameter estimation problem is presented. Three existing ANN training strategies are reviewed within this framework, including gradient-descent backpropagation, the extended Kalman algorithm, and the recursive-least-squares method. A strikingly different approach to error backpropagation is presented, resulting in the development of a novel method of backward signal propagation and target state generation for embedded layers. The technique is suitable for implementation with a linear Kalman-based update algorithm and is applied with a unique method of covariance modification for the elimination of transients associated with initial conditions. Experimental nonlinear identification results demonstrate a greatly increased rate of convergence in comparison with backpropagation. The new algorithm displayed similar rates of parameter convergence and a decreased computational overhead compared with the extended Kalman algorithm.<<ETX>>

Estimation of artificial neural network parameters for nonlinear system identification

A unified framework for representing ANN (artificial neural network) training algorithms is developed by considering weight selection as a parameter estimation problem. Three existing ANN training strategies are reviewed within this framework, i.e., gradient-descent backpropagation, the extended Kalman algorithm, and the recursive least squares method. A strikingly different approach to error backpropagation is presented, resulting in the development of a novel method of backward signal propagation and target state generation for embedded layers. The proposed technique is suitable for implementation with a linear-Kalman based update algorithm and is applied with a time-varying method of covariance modification for the elimination of transients associated with initial conditions. Results from a nonlinear identification experiment demonstrate an increased rate of convergence in comparison with backpropagation. The new algorithm displayed similar rates of parameter convergence and a decreased computational overhead compared to the extended Kalman algorithm.<<ETX>>

Identification of switched reluctance motor states using application specific artificial neural networks

In this paper a novel artificial neural network architecture suitable to identify the states of a switched reluctance motor is developed. This architecture incorporates the a priori knowledge about the motor directly into the structure of a feedforward artificial neural network. A method for backpropagating the error is presented with emphasis given to the specifically developed application specific layers for the switched reluctance motor. The switched reluctance motor model is given. A summary of the integration of the motor model into the ANN is presented. Simulation results show increased convergence rates as well as superior overall identification of the motor states.

On training artificial neural networks to identify periodic functions

Training an artificial neural network (ANN) to represent some type of plant, system, or general algebraic function is relatively straightforward and many methods exist. However, most of these methods and ANN architectures do not take into account any a priori knowledge that is often known for the problem of interest. In this paper, a priori knowledge is utilized to encourage the output of the ANN to be a periodic function of the input. Methods are investigated and compared with the general gradient-based (backpropagation) algorithm. Results show improved convergence characteristics and, in some cases, the guarantee of output periodicity with respect to the input. Generalizations can be made to higher dimensional spaces, but these may or may not be reasonable in terms of computational time and effort.<<ETX>>

Switched reluctance motor control with artificial neural networks

This paper presents a new control scheme for switched reluctance motor drives based on artificial neural networks (ANN). The ANNs are trained to generate drive circuitry phase current references for velocity reference tracking. A new, application specific ANN architecture is used to improve modeling accuracy. The control ANNs are trained using data from a state space model. The control scheme characteristics are then presented via two case studies. Firstly, a constant velocity control is simulated and a comparison with previously measured results is presented. A velocity reference tracking case study is then presented.

Using self-organizing artificial neural networks for solving uncertain dynamic nonlinear system identification and function modeling problems

The authors describe novel implementations of the KNN (Kohonen topology-preserving self-organizing neural network) structure as it is applied to nonlinear functions, control system identifications, and switched reluctance motor torque modelings. Specifically, they examine novel training paradigms, including a procedure for initializing and resetting neuron weights, incorporating prior knowledge into a KNN, and preferentially training specific areas of a KNN. Several functions are modeled as examples of the implementation and properties of this technique. Also, the KNN is used to model a nonlinear mapping embedded in a series-parallel control identifier. Finally, a 2-D KNN is used to successfully estimate the torque in a switched reluctance motor.<<ETX>>