摘要详情

ID / 提交时间

100 / 2025-05-15 06:39:40

标题

Inverse Physics-Constrained Learning for Robust Fault Diagnosis in Robotic Joint Transmission Systems

关键字

Physics-guided fault diagnosis, Robotic joint transmission, Residual-constrained learning, Latent degradation modeling, Multi-modal sensor

主题及专题

Special Sessions > SSL 21 Key Technologies for the Operation and Maintenance of Industrial Robots

状态

全文待审

作者

Qiao Han / South China University of Technology；Institute for Super Robotics（Huangpu）

Jing Liu / South China University of Technology

Yuan Zheng / South China University of Technology

Hao Lan / South China University of Technology

Xu Tan / South China University of Technology

Weihua Li / South China University of Technology

摘要

Accurate and interpretable fault diagnosis for robotic joint transmission systems remains a key bottleneck in achieving reliable industrial automation. These systems often degrade through complex time-varying dynamics—such as stiffness loss and damping fluctuations—that challenge conventional black-box models, especially when signal noise and structural ambiguity are involved. To overcome this, we propose an inverse Physics-Constrained Learning (iPCL) framework that infers latent mechanical parameters from vibration signals using a simplified dynamic model. A residual physics constraint is introduced to align estimated responses with system dynamics, offering physically consistent supervision without relying on hard-to-measure forces. In parallel, motor current signals are leveraged as auxiliary inputs to enhance class separability, decoupled from the physics pathway to preserve interpretability. The fused representation significantly improves both classification accuracy and physical insight. Experimental results on real-world joint datasets demonstrate that iPCL consistently outperforms traditional signal-driven and deep learning baselines. This work establishes a scalable and physically grounded diagnostic paradigm for intelligent health monitoring in industrial robotic transmissions.