Anomaly Detection for Condition Monitoring in Robot Systems

Abstract: This thesis addresses the detection of wear patterns in robot joints as an indication of the increase in wear level or impending failures. The main challenges include identifying key wear features, developing efficient anomaly detection algorithms, ensuring generalization across different joints and operating conditions, and enabling real-time monitoring. The research questions formulated for this study focus on the effective characterization of torque signals, the selection of algorithms for distinguishing normal wear patterns from abnormalities, and the improvement of system robustness against noise. Experiments were conducted on torque measurements provided by ABB Robotics. The research contributions include developing a robust condition monitoring system using machine learning techniques, evaluating four machine learning models for anomaly detection, calculating anomaly thresholds, incorporating trend analysis for enhanced robustness, and comparative analysis of algorithm combinations. The findings provide insights into the strengths of the algorithms, and the effectiveness of incorporating trend analysis, and contribute to advancements in the field of anomaly detection.