An Uncertainty-Adaptive Robust Graph-Based Belief Rule Base for Interpretable Four-State Fault Diagnosis of Industrial Robots

Industrial robot severity diagnosis must combine evidence from coupled subsystems while remaining stable under fluctuating measurements. We develop an uncertainty-adaptive robust graph-based belief rule base (RG-BRB) that maps monitored attributes to transmission, servo-control, and motion-execution nodes. A row-stochastic graph propagates node evidence, while local fluctuation adjusts antecedent widths and attribute reliability before rule activation. Normalized evidential reasoning yields a belief distribution over normal, slight-degradation, severe-degradation, and failure-related states. An interpretability-constrained adaptive covariance search estimates graph and rule parameters under ordering and normalization constraints. The evaluation uses one record of 2000 observations, with a fixed 600-observation test partition for the primary comparison. Across 10 random-seed runs on that partition, RG-BRB achieved 91.55% mean Accuracy and 0.9115 mean Macro-F1. Condition-shift, noise, calibration, and ablation analyses on the same record show how subsystem-constrained propagation and reliability-aware matching contribute to robust four-state diagnosis.