Evolutionary hypergraph: A Kalman-guided spatiotemporal dynamic hypergraph neural network for multi-sensor rolling bearing fault diagnosis

Modern multi-sensor systems generate high-dimensional, noise-contaminated, and dynamically evolving multivariate time-series data, in which complex spatiotemporal dependencies play a key role in accurate fault diagnosis. However, effectively modeling these high-order correlations remains a significant challenge. To address this issue, this paper proposes a Kalman-guided spatiotemporal dynamic hypergraph neural network(KSD-HGNN), which is designed to comprehensively characterize high-order relationships and dynamic evolution among multiple sensors. The framework first employs a learnable band-pass filter bank for adaptive frequency-domain feature extraction, and then constructs and optimizes a dynamic spatiotemporal hypergraph to jointly encode spatial adjacency and cross-step temporal dependencies. Furthermore, by integrating Kalman filtering into hypergraph convolution, the model performs temporal smoothing and state estimation over hyperedge sequences, yielding more stable spatiotemporal representations. Finally, the state dynamics learned by the Kalman-guided branch and the structural dynamics captured by the adaptive topology branch are fused via a gating mechanism, yielding a more stable and comprehensive system representation. Experimental results demonstrate that KSD-HGNN consistently outperforms existing methods under varying operating and noise conditions, significantly improving multi-sensor bearing fault diagnosis performance.