An unsupervised micro-anomaly detection method using U-KAN with physical residual decoupling for industrial robots

Abstract Industrial robot grasping faces several key challenges, such as subtle anomaly features, scarce abnormal samples, significant domain shifts in training trajectories, and constrained edge computing capacities. To address these issues, this study introduces an unsupervised anomaly detection method, namely the physical residual-decoupled U-shaped Kolmogorov–Arnold network (PR-UKAN). Unlike a simple combination of physical modeling, U-shaped reconstruction, and KAN-based nonlinear fitting, PR-UKAN is motivated by a residual-decoupled semantic-conflict interpretation. Specifically, physical residual decoupling based on robot dynamics first removes task-level macroscopic motion interference and converts weak disturbances into physically meaningful residual features; cascaded smoothing and differential enhancement then expose transient micro-anomaly components that are easily submerged in raw sensor signals; finally, the U-KAN autoencoder preserves local high-frequency abnormal details through skip connections while the Efficient-KAN bottleneck constrains the reconstruction to the normal global manifold learned from anomaly-free samples. The observed branch behavior is consistent with the interpretation that a mismatch between local abnormal details and the normal bottleneck representation may increase reconstruction errors for out-of-distribution micro-anomalies. On the voraus-AD dataset, PR-UKAN achieved a mean AUROC of 93.7 ± 0.4%, compared with 93.1 ± 5.8% for MVT-FLOW under the common evaluation protocol. This descriptive difference did not establish statistical superiority. The measured CPU forward-reconstruction time was 4.19 ± 0.68 ms for one preprocessed 1 × 36 × 1168 sequence. Because preprocessing and anomaly scoring were excluded, this measurement did not represent end-to-end detection latency. These results indicate competitive detection performance under the evaluated protocol, while real-time online deployment remains to be validated.