STRFormer: A spatial topological relationship-guided multi-modal variational fusion network for intelligent health state diagnosis of the manipulator

The precision and reliability of advanced precision machinery equipment hinge critically on the health state of their embedded precise manipulators. In real-world industrial scenarios, the manipulators’ intricate coupled topological structures and the complex operating environment invariably introduce diverse interference and uncertainty factors into the monitored performance data. This significantly complicates high-quality feature extraction and effective information fusion. To address this issue, a novel spatial topological relationship-guided multi-modal variational fusion network, termed STRFormer, is developed and employed for the intelligent health state diagnosis task of manipulators. Specifically, the intrinsic properties characterizing the health state of the manipulator embedded in each independent modal information are initially extracted and strengthened through the constructed integrated scale-wise self-attention (ISSA) module, which takes into account both intra- and inter-scale joint information. Subsequently, from the perspective of the manipulator’s kinematic chain branches, the modal information affiliated with the corresponding kinematic chain branches is further fused through the established hybrid cross-modal variational fusion (HCVF) module, in which the variational inference mechanism is employed to suppress the interference of uncertainty factors. Furthermore, the spatial topological relationships between the individual kinematic chain branches and the manipulator are abstracted into a graph structure, and the ultimate system-level information fusion and task matching are achieved based on this. Through a typical 3-PRR planar parallel manipulator in multi-scenario cases, the feasibility, superiority, robustness, and anti-interference capability of the proposed fusion method are comprehensively verified and discussed.