About this Abstract |
| Meeting |
2025 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2025)
|
| Symposium
|
2025 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2025)
|
| Presentation Title |
Reinforcement Learning-Based Digital Twin for Predictive Maintenance and Anomaly Detection in UR5 Robots |
| Author(s) |
Salma MESSAOUDI, Ahmed Bendaouia, Esequiel Garcia, Md Shahriar Forhad, Kenneth Duran, El Hassan Abdelwahed, Jianzhi Li |
| On-Site Speaker (Planned) |
Salma MESSAOUDI |
| Abstract Scope |
Traditional anomaly detection methods in industrial robotics often rely on supervised learning, which struggles with dynamic environments and requires extensive labeled data. In this study, we propose a reinforcement learning (RL) framework for real-time, unsupervised anomaly detection in a UR5 robotic arm. Using high-dimensional data for different payloads and operational conditions, we train four separate Deep Q-Network (DQN) agents, each dedicated to a specific anomaly type: general, position, velocity, and current. Unlike Q-learning, which is ineffective with large state spaces, DQN leverages deep neural networks to model complex robot behavior. Our experiments show promising results, achieving up to 93.2% accuracy, with the best performance observed in position anomaly detection (F1-score of 0.89). This study highlights the scalability and adaptability of RL-based models for autonomous robot monitoring, and paves the way for future integration with temporal-aware architectures and real-time Digital Twin environments. |
| Proceedings Inclusion? |
Planned: Post-meeting proceedings |