A robust deep learning system for motor bearing fault detection: leveraging multiple learning strategies and a novel double loss function

Abstract

Motor bearing fault detection (MBFD) is vital for ensuring the operational safety, reliability, and efficiency of industrial machinery. Because rolling bearing failures account for a massive 40% to 70% of breakdowns in electromechanical drive systems, detecting early-stage faults is essential to prevent catastrophic system failures, minimize costly downtime, and reduce overall maintenance expenses. Traditional diagnostic approaches heavily rely on hand-crafted feature engineering from time- and frequency-domain vibration signals, which often fail to generalize under varying operational conditions or handle limited labeled datasets. To address these limitations, this paper introduces an advanced, hybrid deep learning framework called Robust-MBFD. Instead of relying on a single training paradigm, the proposed system integrates a spectrum of multiple learning strategies—incorporating supervised, semi-supervised, and unsupervised learning—to maximize the extraction of deep feature representations from raw vibration data. A cornerstone of the architecture is the introduction of a novel double loss function designed to refine feature boundaries. This framework effectively bridges the gap between deep learning-based feature extraction and the interpretability of traditional machine learning classifiers, presenting a highly adaptable and comprehensive solution for predictive maintenance in industrial settings.