Modeling dynamic engineering processes using radial-Gaussian neural networks

Abstract

The paper proposes and evaluates an artificial neural network based method of modeling the dynamic behavior of continua. The technique is applicable to situations where the differential equations governing the behavior of a system are nonlinear and poorly understood, and the data available for training is noisy. A method of modeling the unknown component of governing differential equations using neural network technology, is first described. This includes a method for averaging out localized errors in the neural network function that results from noise in the training data. A description is then given of a radial-Gaussian neural network architecture and training algorithm adopted for this application. The construction of a complete simulation model of a specific system from the trained neural networks is demonstrated. The performance of the proposed approach is assessed in a series of experiments simulating the nonlinear thermal behavior of a translucent solid material. The system is proven to perform most effectively using the proposed error averaging technique, and to be capable of providing an accurate simulation of a system's behavior sustained over many thousands of simulation time steps.