Multi-Source Sensor Fusion and Dynamic Accuracy Retention Control for Collaborative Robots in Digital Manufacturing

Abstract

Collaborative robots deployed in digital manufacturing environments must sustain high positioning accuracy despite payload variation, external interference, sensor noise, and communication delay. Existing approaches frequently concatenate asynchronous sensor streams, apply fixed fusion weights, or decouple abnormal-state prediction from bounded control updating, thereby limiting their practical reliability. This study proposes a confidence-aware multi-source sensor fusion and safety-bounded residual control framework designed to address these limitations. The proposed method integrates delay-aware synchronization, modality-specific temporal convolutional encoders, reliability-gated fusion, dual residual and anomaly prediction heads, and projected online controller adaptation. Supervised in-domain evaluation was conducted on the AURSAD dataset, while cross-domain transfer assessment was performed using the official voraus-AD test set through shared joint and electrical channels. Closed-loop tracking performance was further validated within a UR3e digital-twin environment. Across five independent runs, the proposed method achieved Macro-F1 scores of 88.4 ± 1.3% on AURSAD and 86.2 ± 1.5% in the transfer evaluation, surpassing Transformer-based model predictive control by 0.8 and 1.3 percentage points, respectively. The framework also attained a precision retention rate of 89.7 ± 1.6%, exceeding the strongest baseline by 3.6 percentage points. Cartesian root mean square error was 1.12 ± 0.08 mm, comparable to the 1.09 ± 0.08 mm of Transformer-based model predictive control with no statistically significant difference. These results demonstrate that the proposed framework primarily enhances interference tolerance, transfer stability, and bounded accuracy retention in collaborative robotic systems.