STABILIZATION OF THERMAL NEUROCONTROLLERS

This work deals with the stabilization of neurocontrollers used in thermal applications. The control system can be reduced to an iterative, nonlinear map in time, and its linearization enables a stability analysis. For simple neural networks with few neurons, the eigenvalues can be analytically calculated in terms of the synaptic weights and biases. However, unless care is taken, usual training methods can drive the network to weights and biases such that the corresponding control system is unstable. A modified backpropagation training method is developed here to simultaneously minimize the target error and increase the dynamic stability of the system. Numerical computations are used to analyze the stability of realistic neural networks and their corresponding control systems. The techniques developed are used on an experimental heat-exchanger facility where the stability results are tested and validated.