Neural Network Used to Stator Winding Interturn Short-Circuit Fault Detection in an Induction Motor Driven by Frequency Converter

This work is the application of a Multilayer Perceptron Artificial Neural Network (MLP ANN) to detect early interturn short-circuit faults in a three-phase converter-fed induction motor. The quantity used to analyze the problem is the stator current or, more specifically, the harmonic content of its frequency spectrum, also called current signature. The analysis through the current signature is a non-invasive method and may be embedded in the frequency converter, what is a great advantage. The dataset used for training and validating the ANN is obtained using a test bench that allows applying different levels of interturn short-circuits in the machine. It is observed that the fault motor dataset and healthy motor dataset are difficult to separate, which demands a large computational effort to choose a proper MLP topology. The MLP is trained by two different algorithms (the classical error Back propagation - BP - and an adaptation of the newer Extreme Learning Machine - ELM) and the results are thoroughly explored, including after the application of a pruning method called CAPE. Then it is slightly compared with the results of a Self-Organized Map ANN [1] obtained by using the same dataset.