MULTIMODAL BEARING FAULT CLASSIFICATION UNDER VARIABLE CONDITIONS: A PHYSICS-INFORMED 1D CNN WITH TRANSFER LEARNING

Abstract

Bearings play an integral role in ensuring the reliability and efficiency of rotating machinery by reducing friction and handling critical loads. Bearing failures, which account for the majority of mechanical faults, highlight the imperative need for reliable condition monitoring and fault classification. This study proposes a multimodal bearing fault classification approach that leverages vibration and two motor phase current signals within a one-dimensional convolutional neural network (1D CNN) framework. The proposed method fuses feature from multiple signals to improve the accuracy of fault classification and systematically investigates the impact of regularization strategies—including L1, L2, and Dropout—on model performance and overfitting. The methodology is extensively evaluated on the Paderborn University (PU) Bearing Dataset under multiple operating conditions, allowing for a comprehensive analysis of the framework’s adaptability and effectiveness. Additionally, the framework is enriched by integrating a physics-informed branch utilizing vibration signal envelope spectra to extract characteristic fault frequencies, in particular, the ball pass frequency of the outer race (BPFO) and inner race (BPFI). Furthermore, a physics-informed loss function is incorporated to enforce physically consistent predictions, enabling the model to leverage domain knowledge for enhanced learning and interpretation. For adaptability across varying operational environments, three transfer learning strategies are introduced. The integration of domain knowledge through physics-informed features, in combination with transfer learning strategies, enables the framework to more accurately capture underlying faults and adapt to a range of operational conditions. Overall, this multimodal 1D CNN framework augmented with late fusion, advanced regularization, physics-informed element, and transfer learning strategies, lays a strong foundation for accurate, adaptable, and interpretable bearing fault classification. The proposed approach is robust to changes in operating conditions and offers practical benefits for industrial environments for reliable fault diagnosis and condition monitoring of critical machineries.