Adaptation of Machine Learning Models across Domains for Bearing Fault Detection

Abstract

Bearing fault detection is critical for the maintenance and reliability of rotating machinery across industries. Traditional machine learning models have demonstrated significant success when trained and tested on data from the same domain. However, their performance often deteriorates sharply when applied across different domains due to distributional shifts caused by varying operating conditions, sensor types, or system configurations. This paper explores the challenge of domain adaptation for bearing fault detection and investigates the performance of machine learning models enhanced with domain adaptation techniques. Specifically, we examine the impact of methods like Correlation Alignment (CORAL), Transfer Component Analysis (TCA), and Domain-Adversarial Neural Networks (DANN) on improving cross-domain generalization. We conduct extensive experiments using publicly available bearing datasets, simulating realistic shifts in operational domains. Our results reveal that domain adaptation not only improves detection performance across unseen domains but also preserves interpretability, crucial for industrial deployment. This work paves the way for robust, generalizable fault detection systems capable of handling real-world variability without requiring costly retraining.