A Noise‐Aware Cross‐Modal Learning Approach for Bearing Fault Diagnosis Under Varying Operating Conditions

ABSTRACT Rolling bearings play a vital role in rotating machinery, yet their fault diagnosis under real industrial conditions is challenged by high noise levels, limited labeled data, and class imbalance. To address these issues, this paper proposes a cross‐modal fusion network (CMFN) for few‐shot bearing diagnosis, designed to exploit the complementarity of one‐dimensional vibration sequences and two‐dimensional time–frequency spectrograms. The signal branch employs a noise‐aware pyramid convolutional extractor to capture multi‐scale temporal patterns and incorporates sample‐wise reweighting for enhanced robustness. Meanwhile, the image branch integrates a noise‐to‐vector enhanced efficient channel attention (N2V–ECA) module, leveraging noise priors derived from the signal branch to reduce channel bias and suppress noise‐dominated responses. A feature‐wise linear modulation (FiLM) fusion module further conditions image features on signal semantics, enabling fine‐grained cross‐modal alignment with lightweight computation. Extensive experiments with rigorous statistical validation (including fivefold cross‐validation) were conducted on the Jiangnan University and Harbin Institute of Technology bearing datasets under strict few‐shot settings (30 training samples per class). CMFN consistently outperformed state‐of‐the‐art single‐ and dual‐modal baselines, including recent heavy‐duty Transformer architectures. It achieved an outstanding average test accuracy of 99.0% with an extremely lightweight footprint of only 0.86M parameters and a 2.92 ms inference latency. Furthermore, it demonstrated exceptional noise resilience (maintaining 79.0%–97.0% robust accuracy across −5 to 3 dB SNR) and obtained 97.2% average accuracy on a heterogeneous unseen aero‐engine dataset without architectural adaptation. These results thoroughly demonstrate its superiority as an integrated, highly generalizable framework for industrial edge deployment.