Multi-modal dynamic evolution and hypergraph information bottleneck network for small-sample bearing fault diagnosis

Bearing fault diagnosis in practical monitoring is challenged by limited labelled samples and multi-modal sensor data. Existing deep learning methods overfit and struggle to capture higher-order cross-sensor dependencies under small-sample conditions. To address these issues, this paper proposes a liquid hypergraph information bottleneck (LHIB) framework for multi-modal dynamic evolution and feature fusion in small-sample bearing fault diagnosis. LHIB models each sensor channel as a node and learns continuous-time co-evolution of node embeddings and hyperedge states via coupled ordinary differential equations with adaptive liquid time constants. Unlike static hypergraph constructions that rely on predefined topological priors, the liquid evolution mechanism enables diagnostically meaningful higher-order sensor collaborations from minimal initialisation. To improve generalisation, an attention-enhanced hypergraph information bottleneck (AE‑HIB) aggregates the evolved node states and hyperedge states through multi-head attention pooling, structure-aware fusion, and variational compression into a compact latent variable. This bottleneck suppresses task-irrelevant noise at the feature and higher-order structural levels while preserving label-discriminative information. Experiments on two public datasets demonstrate that LHIB achieves 98.30%±1.32% accuracy on the PU dataset and 95.00%±1.34% on the BJTU-RAO dataset, outperforming state-of-the-art methods by margins of 1.6%–8.7% in the most challenging scenarios. These results suggest LHIB is efficient for practical multi-sensor bearing monitoring.