Answers are grounded in this paper and cite their sources.
In practical industrial scenarios, bearings predominantly operate under normal conditions, yielding extremely limited data when faults occur. Consequently, intelligent fault diagnosis methods face a severe small sample problem during application. Furthermore, deploying deep learning-based diagnostic models on resource-constrained devices requires lightweight networks; however, due to their limited parameter capacity, these networks are highly susceptible to overfitting under small sample conditions, thereby degrading generalization capability. To address these challenges, this paper proposes a novel diagnostic framework named PAG-MHA-MobileNet. First, a Physics-Aware Generative (PAG) module is designed. By introducing a Frequency Consistency Loss as a physical constraint, this module compels the generator to preserve the power spectral density characteristics of the signal while restoring image textures, thereby generating augmented samples with high physical fidelity. Second, an improved MHA-MobileNet classifier is proposed. Building upon the MobileNetV3 architecture, it integrates a MultiScale Stem (MS-Stem) module to capture broadband features, employs a Hybrid Attention mechanism to dynamically enhance feature representations across different hierarchical levels, and designs a Dual-Pooling Head to retain transient fault impulses. Extensive experiments were conducted on the Case Western Reserve University (CWRU) bearing dataset, a self-constructed bearing testbed, and the Southeast University (SEU) bearing dataset. The experimental results demonstrate that PAG-MHA-MobileNet achieves superior diagnostic accuracy over existing state-of-the-art methods under small-sample conditions, attaining 99.50%, 98.20%, and 97.80% accuracy on the CWRU, self-constructed, and SEU datasets, respectively. Meanwhile, the proposed model maintains an exceptionally low computational complexity of 78.400M FLOPs and 1.737M parameters, striking a favorable balance between lightweight design and recognition performance. This provides efficient algorithmic support and a theoretical reference for intelligent fault diagnosis in resource-constrained industrial scenarios.
No takes yet. Share an insight, caveat, or question.
zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: