Increasing the Robustness of Fault Detection for Induction Motors based on Neural Networks and the Winding Function Method

This paper presents how a feed forward neural network can be trained on simulated data with high robustness for practical fault detection in induction motors. The basic methodology for the fault detection is a combination of model- and machine-learning-based approaches. The applied framework consists of a feed forward neural network and a multiple coupled circuit model based on the winding function method. The dataset of stator currents in healthy and faulty states simulated by the model allows a neural network-based classification of different faults without the need to measure currents under real fault conditions. However, this approach suffers from the difficulty of transferring the extracted fault characteristics from the simulated data to the measured stator currents. Therefore, the effect of ensemble learning on increasing the robustness of fault detection is investigated in detail. In addition, an analysis of the influence of the hyperparameters of the neural network on the transferability of the extracted fault characteristics from the simulated stator currents is carried out. As a result, it is found that both techniques increase the robustness of the methodology for fault detection.