A novel multi-sensor multi-granularity attention fusion method for bearing fault diagnosis

Abstract Accurate bearing fault diagnosis is essential for rotating machinery reliability but is frequently challenged by environmental noise. In this paper, a novel framework that robustly identifies bearing faults under challenging noise conditions is proposed by integrating multi-sensor data vibration, acoustics, and temperature with advanced deep learning techniques, where single-sensor and non-adaptive methods often fail. To simulate realistic operational environments, impulse noise and Gaussian white noise are gradually added to the sensor data. Then introduces a Dynamic Noise Resilient Feature Mining (DNRFM) module, which employs an adaptive masking strategy to effectively attenuate these noise types while preserving essential fault-related features. Following noise reduction, the denoised data are fed into a fine-tuned ConvNeXt model for deep feature extraction. To further enhance fault discrimination, a Multi-Granularity Attention Fusion (MGAF) mechanism is developed, which synergistically integrates features from the three sensors at multiple levels of abstraction, thereby capturing both fine-grained and holistic patterns indicative of various fault types. Experimental evaluations on two public datasets (University of Ottawa and KAIST) demonstrate that our approach outperforms conventional methods. The proposed framework achieved diagnostic accuracies of 98.05% and 100%, respectively, under severe noise conditions (-10 dB), showcasing superior robustness and generalization capability. The proposed framework improves the reliability of bearing fault diagnosis and offers a flexible approach to multi-sensor data fusion, particularly suited for industrial environments with unavoidable noise.