Identification of Optimal Time Domain Features for Machine Learning based Fault Classification

Implementing machine learning algorithms on industry-specific embedded platforms poses challenges due to restricted resources like bandwidth, memory, and resolution, among others. Therefore, it is essential to optimize the employed features, tree depth, learning rate, code, etc. This study aims to detect distributed bearing faults in AC drives by utilizing the optimal feature selection in fault diagnosis. This approach is innovative in its comprehensive integration of six distinct types of signals: three vibration signals, stray magnetic flux signals, and two phase-current signals. Using a data acquisition board and a dynamometer setup, 16-bit sensor data is gathered from two induction motors. This encompasses 50 operational points, spanning 10 distinct speed levels and 5 torque levels. A total of 42 time-domain features (TDFs), frequently employed in industrial applications, are chosen, and their significance is assessed within a machine learning (ML) framework. To ensure the reliability of the feature selection process and the fault-detection algorithm, comprehensive tests are conducted for validation purposes. A 10-fold cross-validation with Random Forest (RF) classifiers compares various settings and time-domain features, revealing that the vibration sensor has the greatest impact on classification accuracy. The classification outcomes are compelling, showcasing remarkable potential in detecting distributed bearing faults with only 32 decision trees and 10 features. A key finding of this study is identifying the most effective set of time-domain features in the classification of distributed bearing faults.