Towards edge-deployable non-destructive testing: A hardware-aware lightweight framework for real-time phased array ultrasonic testing of girth welds

Phased Array Ultrasonic Testing (PAUT) for pipeline girth welds is transitioning to automated diagnostics. However, integrating deep learning into portable PAUT instruments is bottlenecked by conflicts between precision and limited computational resources. This paper proposes You Only Look Once-Ghost Large Lightweight Loss Prune (YOLO-GLLLP), an efficient framework for rapid detection of weld defects. To capture unique acoustic spatial characteristics while maintaining real-time performance, we introduce a lightweight backbone to minimize the computational footprint. A large separable depthwise convolution attention mechanism expands the receptive field, enhancing sensitivity to subtle defects. Furthermore, a shared depthwise convolution-based detection head mitigates parameter redundancy during multi-scale feature fusion. To ensure robust convergence and localization accuracy, advanced bounding box regression loss functions are implemented. Finally, a structured model pruning algorithm is applied to further compress the network for hardware deployment. Experimental results on a pipeline PAUT dataset demonstrate that compared to the baseline You Only Look Once11 nano (YOLO11n), YOLO-GLLLP improves mean Average Precision at 50% Intersection over Union (mAP50) by 2.6%, while reducing parameters by 57.7% and Floating Point Operations (FLOPs) by 49.2%. Frames Per Second (FPS) is concurrently enhanced by 25.2%. To validate edge-hardware feasibility, the optimized model is deployed on a low-power edge Intel Core i5 using the OpenVINO toolkit, achieving a peak throughput of 47.03 FPS and a sustained throughput of 30.53 FPS under continuous thermal load. This work provides a computationally efficient framework, offering a promising solution toward the practical implementation of online, automated pipeline inspection systems.