Condition-Adaptive Hybrid Anomaly Detection for Machine-Tending Applications

Abstract

Retrofit condition monitoring of industrial manipulators should distinguish actual mechanical anomalies from signal changes produced by payload, speed, program phase, and transient motion. This study presents a hybrid detector for a six-axis machine-tending robot using a forearm-mounted inertial measurement unit and an auditable two-stage decision architecture. Engineered descriptors support two branches: a Random Forest estimates similarity to reviewed abnormal patterns, while a PCA representation measures context-compatible geometric novelty. The branch scores are combined via a linear fusion coefficient selected on grouped validation runs. Operating context selects a pre-validated, controlled context-compatible PCA reference and modifies the final decision through a bounded threshold correction; it does not update the nominal model online. An H-of-K persistence rule converts repeated window-level exceedances into event-level maintenance evidence. All data-dependent transformations are fitted after complete physical acquisition runs have been assigned to training, validation, or held-out testing. In the run-grouped archive, the complete adaptive hybrid achieved 97.1 ± 0.8% accuracy and an F1-score of 0.96 ± 0.01 across 18 held-out runs. The resulting framework prioritizes leakage-controlled validation, constrained adaptation, and computationally modest retrofit deployment; further developments will enable online adaptation.