Fault Detection for Quadruped Robots Based on Dynamic Decision-Making and Layered Elastic Weight Consolidation

Quadruped robots operating in dynamic, unstructured environments demand robust fault detection to ensure safety and autonomy. Traditional model-based approaches, while interpretable, often fail to capture the complex nonlinear dynamics of legged locomotion due to unmodeled interactions and environmental uncertainties. Data-driven methods, particularly deep learning-based techniques, have shown promise in extracting fault patterns from high-dimensional sensor data. However, existing methods often struggle with complex spatio-temporal fault patterns, limited adaptability, and catastrophic forgetting when learning new faults. To address these challenges, we propose a Dynamic Spatio-Temporal Learning framework for Fault Diagnosis in Quadruped Robots, integrating Heterogeneous Graph-Augmented Spatiotemporal Convolutional Network (HG-STCN) and Deep Reinforcement Learning. The HG-STCN explicitly models the robot's topological structure for accurate spatio-temporal fault diagnosis, while a DQN dynamically optimizes computational resource allocation to balance real-time performance and reliability. In addition, Layered Elastic Weight Consolidation) enables incremental learning, allowing the system to adapt to new faults without forgetting prior knowledge. Experimental results demonstrate superior detection accuracy, robustness, and adaptability compared to baseline methods. This work advances the reliable, long-term autonomy of quadruped robots in challenging environments.