Optimal Robust Time-Domain Feature-Based Bearing Fault and Stator Fault Diagnosis

Abstract

In machine learning, the extraction of features is necessary for intelligent motor fault diagnosis. In industrial applications, it is necessary to identify the optimal number of features to differentiate various types of fault characteristics with less computational complexity and cost. However, motor fault diagnosis for real-time applications has challenges in capturing characteristics due to variations in speed, load, force, run-to-failure state as well as the type of the motor and its parts. The deep learning techniques that automatically extract features and perform classification have algorithmic complexity. In this work, the authors address these challenges by: 1) selecting and ensembling optimal time-domain features that are capable of identifying motor faults using current signals of the permanent magnet synchronous motor (PMSM) in bearing; and 2) investigating the feature ensemble constituting optimal features for robust fault diagnosis in the PMSM bearing as well as the stator and bearing of squirrel cage induction motor (SCIM) for various conditions. The optimal features mean absolute value, simple sign integral, and waveform length yields 99.8% and 100% for bearing fault and stator fault diagnosis, respectively, in PMSM. These features show 100% accuracy for identification of fault in SCIM and 98.2% accuracy in the run-to-failure state.