Fault tolerant neural networks for control systems

The authors review some techniques that have been used to design fault tolerant neural networks. One technique involves the injection of possible faults into the network during the training phase. This technique is called the fault injection, or immunization, technique, because of the resulting improvement in the immunity of the network to these faults. Another technique involves redundant computations by groups of neurons in a network. This technique is called the redundancy technique. The redundant computations are usually performed in parallel with nominal computations. Simulation studies on the application of the immunization technique to the design of a fault tolerant neurocontroller are presented. The simulation results show that the fault tolerance of a neurocontroller can be improved when training incorporates the immunization technique.