YOLO-PSA: A Digital Twin-Driven Robotic Part Detection Method with Pinwheel Convolution and Single-Head Attention for Industrial Manufacturing

In this paper, we propose a robot part category detection method based on the synergy of digital twin and simulation software to achieve efficient and accurate industrial part recognition by improving the YOLOv11n framework. To address the small target detection challenges, a windmill-like convolution module (PConv) is designed to enhance the local feature extraction capability by using an asymmetric filling strategy; a single-head self-attention mechanism (SHSA) is proposed to reduce the computational redundancy through channel segmentation; and an in-scale feature interaction module (AIFI) is integrated to strengthen the spatial correlation of highlevel semantic features. Experiments show that the proposed method (YOLO-PSA) achieves $\mathbf{9 1. 6 \%}$ detection accuracy on an industrial dataset containing bearings, nuts, gears, and bolts, which is a $3.1 \%$ improvement over the baseline model, and the number of parameters only increases by $\mathbf{2 5. 5 \%}$. Combined with digital twin technology to build a virtual-realistic inspection process, the simulation software predicts the robot behavior and optimizes the inspection path, realizing real-time classification and localization of multi-targeted parts in the factory environment, and providing reliable technical support for intelligent manufacturing.