Fault Detection and Diagnosis (FDD) is crucial for ensuring safe, reliable, and energy-efficient operation of collaborative robots, especially with the growth of Industry 4.0. Industrial robots are nonlinear, complex, and dynamic, making traditional threshold- and rule-based FDD methods inadequate for accurate fault detection. Deep learning approaches address this by enabling autonomous, data-driven analysis through learning hierarchical patterns from large sensory datasets. This paper reviews recent deep learning techniques for FDD in IIoT-based robotic systems, categorizing them by architecture: LSTM and CNN models for time-series fault analysis, autoencoders (AEs) and variational autoencoders (VAEs) for anomaly detection, and hybrid models for multi-sensor data integration. It also highlights the role of IoT infrastructure in real-time data acquisition, fault communication, and predictive maintenance via edge, fog, and cloud layers. Additionally, evaluation metrics, benchmark datasets, and performance comparisons are discussed. However, key limitations include lack of real-time deployability, poor generalization to unseen faults, limited interpretability, and class imbalance. The paper concludes with future directions such as federated and edge learning, self-healing robotic systems, transfer learning, and integration of Explainable AI (XAI) to develop scalable and fault-tolerant cobot systems.
Banu et al. (2026) studied this question.