Machine Learning Approach with Multiple Feature Selection Techniques to Diagnose the Inter-Turn Winding Faults in Induction Motor

Abstract

This paper presents an innovative and effective approach for detecting, analysing and classifying stator winding faults in induction motor using the motor current signature analysis (MCSA), combined with machine learning models. Stator inter-turn winding faults are a critical issue affecting the reliability of induction motors, which require accurate fault detection to maintain motor performance and prevent failures. This approach employs advanced signal processing techniques for fault identification, including signal envelope identification analysis, Park’s vector magnitude analysis and zero-crossing time detection (ZCTD), to extract deep features from stator current under both healthy and faulty motor conditions. The motor fault features are computed using statistical feature analysis methods from recorded current signals. The most appropriate feature subsets are identified using feature selection techniques known as Fisher Score, minimum redundancy maximum relevance (m-RMR) and Relief. In the classification stage, conventional machine learning models like k-nearest neighbours (k-NN), logistic regression (LR), random forest (RF) and support vector machine (SVM) are applied to these selected features to efficiently classify the healthy and faulty states of induction motor. To validate the proposed methodology, an experimental study is conducted in the laboratory to record stator current data from both healthy and multiple fault phases of the induction motor under varying load conditions. Hence, this paper presents a promising solution for accurate fault detection and classification of stator winding faults, reducing the need for extensive manpower and sensor usage while enhancing the reliability of predictive maintenance schemes.