Current-Driven Deep Learning for Enhanced Motor Bearing Prognostics

Abstract

An innovative deep-learning framework embedding residual and channel attention blocks is introduced, tailored explicitly to assess the remaining lifespan of cooling motor bearings affected by distributed faults, utilizing only 3-phase current signals. Cooling fan motors are integral to the stability of power electronics systems and data centers. Using an accelerometer to monitor bearings is not very practical for small motors and is costly. Hence different from other studies employing vibration signals for high-power machines, this research prioritizes motor current signals. Aging-related defects are clearly visible on small motor currents, unlike the large ones since the rated torque is low and the modulated torque disturbance caused by bearing defects is more noticeable on the phase currents. The experimental approach involves subjecting motors to diverse testing conditions, spanning from normal operation to failure, to ascertain their RUL before potential critical issues emerge. Seven configurations, encompassing permanent magnet synchronous motors and brushless direct current motors with varying power ratings, undergo rigorous experimental testing from normal to failure states. The resultant diverse dataset forms the basis for developing a robust distributed bearing fault detection algorithm. The proposed deep learning architecture demonstrates notable performance with a train accuracy of 97.10% and a test accuracy of 95.94%. This suggests a high level of effectiveness and generalization capability in accurately predicting the RUL in cooling fan motors using current signals.