A Spatter-Aware, Data-Driven Closed-Loop Framework for Real-Time Monitoring and Adaptive Control in Wire Arc Additive Manufacturing

Real-time monitoring and adaptive parameter tuning in Wire Arc Additive Manufacturing (WAAM) are critical to ensuring stable process behavior and consistent product quality. This study proposes a low-cost, real-time closed-loop framework that integrates spatter detection, parameter control and quality prediction. CCD images acquired during welding are first preprocessed through glare suppression, normalization, and region-of-interest cropping, and then fed into a modified YOLOv11 network incorporating CBAM attention and CIoU-based localization, enabling robust detection of tiny spatter under intense arc radiation and fume interference. Frame-wise detection results are transformed into a unique-count time series, from which windowed statistical features (mean, extrema, and trend) are extracted and fused with electrical signal measurements. A random forest regression (RFR) module is subsequently employed to recommend micro-adjustments of welding current and voltage when an increasing spatter-risk trend is identified. An interactive visualization interface is developed to integrate online detection results, spatter trend analysis, parameter status cards, and layer-wise quality heatmaps, thereby facilitating in-process intervention. On the validation dataset, the proposed detector achieves a precision of 92%, a recall of 89% (F1 score = 0.90), and real-time inference performance of approximately 138 FPS at a resolution of 640 × 640. Ablation studies demonstrate that CBAM improves mAP@0.5 from 78.2% to 80.3% with negligible computational overhead, while CIoU further enhances localization accuracy to an mAP@0.5 of 94.5 under an identical training configuration. Compared with YOLOv5 and Faster R-CNN baselines, the proposed approach delivers superior precision–recall performance and more stable training convergence. Moreover, the spatter-driven parameter recommendation strategy enables effective online process stabilization, making the framework suitable for industrial-scale data streams. Overall, this work provides a practical and scalable pathway toward consistent WAAM through detection-guided parameter control and quality visualization.