Continual on-line training of neural networks with applications to electric machine fault diagnostics

An online training algorithm is proposed for neural network (NN) based electric machine fault detection schemes. The algorithm obviates the need for large data memory and long training time, a limitation of most AI-based diagnostic methods for commercial applications, and in addition, does not require training prior to commissioning. Experimental results are provided for an induction machine stator winding turn-fault detection scheme that uses a feedforward NN to compensate for machine and instrumentation nonidealities, to illustrate the feasibility of the new training algorithm for real-time implementation.