Physics-Informed Ensemble Learning for Hierarchical Fault Diagnosis in Quadruped Robots

• Propose a physics-informed ensemble model for hierarchical fault diagnosis in quadrupedal robots. • Introduce a domain-guided feature engineering pipeline that enhances fault discriminability. • Achieve the state-of-the-art performance with 0.958 joint accuracy across detection and classification. • Deliver robust multi-class diagnosis for seven distinct fault types. Reliable fault detection and diagnosis are essential for the safe and autonomous operation of quadruped robots in real-world environments. While deep learning methods have been explored for this task, they often suffer from limited interpretability, high computational demands, and sensitivity to class imbalance. In this work, we propose a physics-informed, hierarchical machine learning ensemble framework for fault detection and classification in quadrupedal robots. Our approach combines sliding-window statistical features with domain-derived metrics and fault-specific signatures to construct a discriminative and interpretable representation. The proposed diagnostic pipeline operates in two stages: first, a binary detector identifies the presence of a fault; second, a multiclass classifier identifies one of seven physically grounded failure modes, both trained exclusively on relevant subsets of the data. Evaluated on a large-scale dataset of 10,000 simulated locomotion trials with precisely injected faults, our soft-voting ensemble of gradient-boosted trees achieves a joint hierarchical accuracy of 0.958 and a weighted F1-score of 0.952, with statistically significant improvement over all baselines ( p < 0.01). Furthermore, an interpretability analysis confirms that contact forces and joint torque dynamics are the most influential features, aligning with mechanical intuition. Extensive experiments and comparisons with various methods demonstrate that our domain-guided feature engineering combined with ensemble learning provides a robust, efficient, and trustworthy alternative to end-to-end deep learning for safety-critical robotic fault diagnosis. Our source code is available at https://anonymous.4open.science/r/FD-6613 .