RJADNet: A Structure-Aware Multijoint Network With Topology-Constrained Aggregation for Industrial Robot Anomaly Detection With Compositional Generalization Under Imperfect Sensing

Multijoint anomaly detection in industrial robots remains challenging when sensing conditions are corrupted by environmental noise and sensor fault, especially when higher-order compound anomalies are absent from training. This article addresses this problem from the perspective of compositional generalization, where models are trained on simple anomaly compositions but tested on unseen higher-order ones. To this end, we propose Robot Joint Anomaly Detection Network (RJADNet), a structure-aware multijoint network with topology-constrained aggregation for industrial robot anomaly detection under imperfect sensing. RJADNet first employs a Robot Single-Joint Feature Extraction backbone based on Trainable Discrete Wavelet Convolution blocks to preserve joint-private anomaly cues, and then applies an Inter-Joint Graph Structure-Aware neck for joint-wise reweighting, topology-constrained inter-joint aggregation, and temporal feature aggregation before classification. Experiments on the Perfect Sensing Dataset and Imperfect Sensing Dataset show that RJADNet delivers strong performance on both seen simple compositions and unseen higher-order compound anomalies compared with representative baselines. Sensitivity analysis, ablation studies, and visualization results further verify the effectiveness of the proposed design.