关键词: 空间机器人, 关节轴承, 故障诊断, 信号预测, 语义建模

Abstract: To address the problems that early degradation characteristics of space robot joint bearings are weak under extreme environments, nonlinear dynamics are difficult to characterize accurately in Euclidean space, and continuous vibration signals are not readily compatible with large language model reasoning, a context-aware manifold geometric semantic modeling framework for fault diagnosis is proposed. First, a dual-discriminator generative mechanism integrating data authenticity and physical consistency constraints is constructed to generate manifold samples satisfying dynamic constraints. Hankel phase-space reconstruction and covariance mapping are employed to embed vibration signals into a Riemannian symmetric positive definite manifold, while the affine-invariant Riemannian metric is introduced to characterize the geometric degradation trajectories during the degradation process. Subsequently, a Riemannian discrete representation learning method is utilized to semantically encode manifold trajectories, and frequent pattern mining together with remapping strategies are incorporated to transform the sequences into statistical structures analogous to natural language. Finally, the discrete semantic sequences are fed into a large language model for temporal reasoning, enabling bearing degradation state prediction and fault identification. Experimental results demonstrate that the proposed method effectively enhances early degradation sensitivity and achieves superior diagnostic performance under complex operating conditions, reaching an accuracy of 96.28% and realizing unified modeling from physical signals to semantic reasoning.

Key words: Space robots, joint bearings, fault diagnosis, signal prediction, semantic modeling