Machine Learning-Based Severity Assessment and Incipient Turn-to-Turn Fault Detection in Induction Motors

A two-layer soft voting ensemble machine learning based algorithm for the diagnosis of turn-to-turn faults in the stator winding of three-phase induction motors is presented. The suggested approach offers severity assessment and faulty phase identification considering recurrence qualification analysis features, extracted from recurrence plot images generated using the Max-Min difference technique from raw signals. Thereafter, the proposed model is implemented with eight machine learning classifiers that undergo training with extracted features utilizing a 10-fold cross-validation technique. Subsequently, predictions of each layer are aggregated through soft voting. Datasets required for training and validation are gathered from a laboratory-based experimental hardware setup of induction motor, covering various turn-to-turn fault severity considering multiple loading and fault resistance. Performance of the proposed algorithm is verified by considering various performance metrics. Comparative results demonstrate that the proposed classifier outperforms individual machine learning classifiers for turn-to-turn fault diagnosis along with severity and faulty phase detection, which in turn assists in reducing downtime and maintenance costs.