Control-Aware Predictive Maintenance of Industrial Robot Motors Using Multi-Sensor Fusion and FDIR Integration

Abstract

This paper presents a deployable multi-sensor anomaly detection pipeline for predictive maintenance of robot joint motors using synchronized temperature, voltage, and encoder position measurements. The proposed workflow performs preprocessing and temporal alignment, constructs lightweight temporally informed features (rolling statistics), and applies feature-level fusion prior to classification. Because confirmed fault annotations are often unavailable in practice, we generate proxy anomaly labels using interquartile range (IQR) fences on each sensor channel and fuse flags with an OR rule, yielding an anomaly prevalence of 26.12%. To reduce leakage from temporally correlated time-series data, we evaluate generalization using a session-based split across eight sessions and six motors. We compare three model classes: Random Forest, XGBoost, and an LSTM sequence model. We further integrate model outputs into a fault detection, isolation, and recovery (FDIR) framework that maps anomaly evidence to residual checks and staged recovery actions suitable for industrial supervisory control. The resulting implementation supports real-time deployment via a REST interface with a median single-prediction latency of 42 ms on a standard CPU platform.