Physics-Aided RUL Prediction of Pick-and-Place Robotic Arm Using Dynamic Simulation and GMM-Based Health Indicator

Abstract

Pick-and-place robotic arms driven by closed-loop servo motors operate under repetitive trapezoidal motion profiles (acceleration–constant speed–deceleration–stop), which yield non-stationary and multi-modal signal distributions and can undermine conventional distance-based anomaly detection and health indicators. This paper proposes a physics-aided prognostics and health management (PHM) framework for a one-degree-of-freedom servo-driven robotic joint. Based on failure mode and effects analysis (FMEA), bearing degradation (represented as increased friction torque) and shaft misalignment (represented as an eccentric disturbance torque) are selected as critical degradation mechanisms. To address the scarcity of industrial run-to-failure data, a high-fidelity Modelica dynamic model of a motor–shaft/bearing–compliance–load drivetrain is developed to synthesize progressive degradation scenarios. We first show that principal component analysis (PCA) visualization and a Mahalanobis-distance-based indicator can be unreliable because healthy data form multiple clusters induced by the servo operating phases. To explicitly capture healthy multi-modality, we train a Gaussian mixture model (GMM) using healthy modes and define a probabilistic health indicator as the negative log-likelihood (NLL) under the learned healthy distribution; the indicator shows an increasing trend across the designed degradation scenarios with degradation progression. The health indicator is then normalized between healthy and failure thresholds to construct a normalized remaining useful life (Normalized-RUL) index that provides a relative life measure without requiring ground-truth failure-time labels. Finally, continuous Normalized-RUL trajectories are estimated using a continuous wavelet transform–long short-term memory (CWT–LSTM) model that learns time–frequency degradation patterns and their temporal evolution. In the most severe simulated condition, the predicted remaining life decreases to approximately 40% of the nominal life, demonstrating the potential of the proposed framework for predictive maintenance of servo-driven robot joints.