Machine Learning Models for Fault Classification in Industrial Motors: A Comparative Analysis

Undetected failures can often lead to complete deterioration and breakdown of a motor in the long term. This study explores the complex process of identifying and predicting faults in three-phase induction machines, which is an essential part of many industrial environments. We induce varying percentages of stator and rotor faults in an industrial motor and record the three-phase current measurements at fixed combinations. Upon pre-processing the data obtained, we apply six machine learning (ML) algorithms to predict the fault type based on the current measurements. We find that the random forests model outperforms the other models in terms of error metrics-accuracy, precision, and F1 score. Moreover, we also determine the predictability of each of the fault types and hypothesize that certain fault types are more predictable than others due to lesser variability in their raw data. We conclude by identifying a need to set up a multi-stage ML-model pipeline that would first filter out no-fault scenarios and effectively predict the fault types more comprehensively.