Research on Robot Fault Diagnosis Based on a Digital Twin and a CNN-LSTM Hybrid Model
DOI:
https://doi.org/10.70693/1z4nba27Keywords:
predictive maintenance, fault diagnosis, digital twin, CNN-LSTM, deep learning, roboticsAbstract
Deep learning models create tremendous opportunities for data-driven fault diagnosis, but their training requires large amounts of labeled fault data, and most models rely on component-level monitoring data. This paper proposes using digital twins to support the development of fault diagnosis models, thereby reducing the amount of fault data required for training and enabling component-level fault diagnosis based on system-level condition monitoring data. Building on the original LSTM model presented in the paper, this study constructs a CNN-LSTM hybrid model that first uses a convolutional neural network to extract local temporal features and then uses LSTM to capture long-term dependencies. Experiments were conducted on the Hiwonder ArmPi FPV robotic system. The results show that the CNN-LSTM hybrid model achieved a test accuracy of 71.11%, an improvement of approximately 9.5 percentage points over the original LSTM model's 61.56%, validating the effectiveness of the hybrid model for fault diagnosis using system-level monitoring data. Meanwhile, the experimental results also reveal that the simulation-to-reality gap remains a major bottleneck limiting diagnostic performance.