Uncertainty-aware sensorless anomaly detection using a reliable indicator from position-guided multi-step deep decomposition network

Condition monitoring (CM) and predictive maintenance (PdM) are essential for ensuring reliability and efficiency in intelligent manufacturing. While bearings and gears have been extensively studied, roller chains have received limited attention, primarily due to the difficulty of installing contact accelerometers on moving chains and the high cost of deploying sensors across long spans. Consequently, effective CM techniques for roller chains remain an open challenge. This study introduces a sensorless CM framework that relies solely on motor driver signals, eliminating the need for additional physical sensors. Motor torque and position data are acquired and processed using a Position-guided Multi-Step Deep Decomposition Network (PMSDDN), a lightweight neural architecture designed to construct a reliable health indicator (HI). PMSDDN segments the torque signal based on the motor position, decomposes each segment into low- and high-frequency components, and predicts them independently through efficient linear modules. This decomposition reduces noise and fluctuations, producing a smooth degradation trend that enhances interpretability. Compared with traditional HIs and state-of-the-art deep learning models, PMSDDN delivers higher effectiveness and computational efficiency. For anomaly detection, a Weighted K-Nearest Neighbors (WKNN) method is developed, combining three different statistical measures with uncertainty quantification to improve robustness and sensitivity to early degradation. The framework is validated on nine roller chains under three operating conditions. Results confirm superior performance in both health assessment and anomaly detection, highlighting its potential as a practical and scalable CM solution for roller chain systems in modern manufacturing environments.