This paper presents a comprehensive predictive maintenance system for industrial robot motors utilizing multi-sensor fusion and machine learning techniques. The proposed system analyzes 84,942 real-time sensor measurementsfrom six motors across eight test sessions, integrating temperature, voltage, and position data to detect operational anomalies. We implement and compare three machine learning approaches: Random Forest (RF), XGBoost, and Long Short-TermMemory (LSTM) networks. Using proper session-based data splitting to prevent leakage, RF achieves an AUC score of 0.871 with corresponding precision-recall AUC of 0.824 and F1-score of 0.813. The system processes a dataset with 26.12% anomaly prevalence (IQR-rule labels), with position sensors providing the strongest predictive signal. Our feature engineering pipeline incorporates rolling statistics and temporal patterns, improving prediction accuracy by 15% over baseline models. The developed web API enables real-time deployment with 42ms single-predictionlatency, making it suitable for industrial IoT applications. Experimental results couldreduce unplanned downtime by 30—45% under typical PdM adoption scenarios (assumptions detailed in §V-D). This work contributes to the field by providing a scalable, production-ready framework for multi-sensor anomaly detection in robotic systems.