MSF-DFormer: A Multisensor Multiscale Fusion Network With Deformable Transformer for Fault Diagnosis Under Complex Working Conditions With Limited Samples

Bearings are critical components in rotating machinery, and their failures could result in significant downtime and safety risks. Intelligent fault diagnosis could encounter considerable challenges in noisy environments and under limited labeled data conditions. A novel multisensor multiscale fusion network with a deformable transformer (MSF-DFormer) is proposed in this article, which incorporates multiscale feature extraction, spatial–channel fusion, cross-layer dense enhancement, and a deformable attention mechanism to enhance feature representation across different sensor modalities. Each raw sensor signal is processed through independent cascaded multiscale depthwise separable convolutional blocks (MDSCB), enabling comprehensive multiscale feature extraction. The resulting features are refined via a spatial–channel reorganization module (SCRM) and preliminarily fused using a sensor feature fusion module (SFFM). Besides, further integration of feature representations from different stages is performed through an adaptive dense feature fusion (ADFF) module. Finally, a deformable offset transformer (DOT) module is employed to integrate outputs from multiple stages by dynamically adjusting offsets to capture long-range dependencies and fine-grained local patterns. The experimental results demonstrate that the proposed method achieves an accuracy of 99.78% on the small-sample Paderborn University (PU) dataset, which contains both vibration and current signals, exhibiting strong diagnostic capability under noisy conditions. On the multisensor dataset from the University of Ottawa, which includes both vibration and acoustic signals, the method outperforms several mainstream models, showing superior accuracy, robustness, and generalization performance. These findings confirm that MSF-DFormer performs well in fusing vibration and current signals, as well as in cross-modal fusion scenarios involving vibration and acoustic data, highlighting its versatility and effectiveness across diverse multisensor fusion tasks.