A Multi-Aspect Geometric Structure Learning and Discrimination Framework for Few-Shot Fault Diagnosis of Rolling Bearings

Identifying rolling bearing faults under few-shot conditions remains difficult because fault samples are scarce, class-space distributions are unstable, and inter-class boundaries may become ambiguous. This paper proposes a fault diagnosis model based on the Multi-Aspect Geometric Structure Learning and Discrimination Framework (MAGS-LDF). First, one-dimensional vibration signals are mapped into three two-dimensional representations, namely angle field (AF), band-energy (BE), and time-frequency (TF) images, to describe fault information from temporal correlation, frequency–band energy distribution, and time-frequency response perspectives. Second, a dual-branch feature extraction network is designed to extract fusion features and channel features. For each fault class, channel features are aggregated into channel centers, which are further fused to obtain a public center. Moreover, regular polytope anchor centers aligned with the public-center distribution are introduced to impose geometric constraints, encouraging intra-class compactness and inter-class separation. Finally, channel distances and public distances are jointly used to construct a multi-scale distance discrimination mechanism for few-shot fault classification. Experiments on CWRU and SEU show that MAGS-LDF outperforms the best comparison method by 2.34%, 5.37%, and 5.34% on CWRU and by 3.55%, 5.10%, and 3.46% on SEU under the three-shot, five-shot, and 10-shot settings, respectively.