Anomaly detection for industrial robots based on temporal graph neural networks and differential privacy

Abstract To investigate the effectiveness of privacy-preserving anomaly detection in real-world industrial robot systems, this study evaluates a federated learning framework with differential privacy by deploying it on multi-source operational data collected from practical manufacturing environments. This mechanism constructs a Temporal Graph Neural Network (T-GNN) that integrates dynamic graph convolution and temporal attention to jointly encode the physical connections and statistical correlations of multiple joints in a robot, effectively modeling the spatiotemporal coupling relationship of multi-source heterogeneous sensor data. It also introduces a Noise Sensitivity Compensation (NSC) module and an adaptive privacy budget decay strategy to dynamically balance privacy protection and model learning capabilities under the ( ε , δ )-differential privacy constraint. Experiments based on real KUKA robot datasets from three automotive manufacturing plants show that when ε increases to 1.5, the method achieves an F1-score of 0.88, significantly outperforming benchmark methods such as FedAvg (Federated Averaging); the number of communication rounds is reduced from 42 to 26, decreasing the communication required for convergence by 38.1%; the Mean Time Between Failures (MTBF) is improved by 23.5%; peak memory usage is 2650 MB, and inference latency is 11.7 ms, with manageable increases in resource consumption. This study breaks through the bottleneck of the negative correlation between detection accuracy and privacy protection under differential privacy, and provides a new paradigm of predictive maintenance that is high-precision, highly privacy-oriented, and edge-feasible for safe collaboration of industrial robots across factories.