Neural network multi-model based method of fault diagnostics of actuators

This paper introduces an artificial neural network based technique which is capable of distinguishing among different types of faulty states of the analysed system and generating signals to alarm the user about the failures in the system. The developed method can detect, separate and identify faults in the system. Large datasets were generated to train the separator networks. A novel active learning method was developed to speed up the training process of separator network. To find the weakness of the separator's mathematical structure, a complex test process was used where the size of the different faults was varied and the actual performance of the structure was examined. The examination had two parts: a) the appearance and termination of the faults were tested; b) the estimation of the fault size was verified. The separator technique requires mathematical models of the analysed system. In this case, the models were also based on feedforward neural networks with tapped delay line. The developed technique was tested on a traditional vehicle starter motor.