Research on deformation detection method of three-dimensional weaving guide rods for composite preforms based on deep learning

In the flexible three-dimensional weaving process for composite preforms, the guide rod is a crucial element for fiber orientation and path regulation. Minor deformations can substantially impact the forming quality and mechanical properties of the preform. The guide rod’s narrow and flexible shape renders it susceptible to intricate deformations, including bending and misalignment, during high-speed weaving or uneven loading conditions. Conventional detection methods are inadequate in terms of reliability and precision, hence complicating real-time monitoring. This research offers a deep learning-based approach for detecting guide rod deformation—YOLOv12-ASHN. The method employs a Space-to-Depth Convolution module to reconstruct the convolutional architecture. It integrates a lightweight detection head along with Dynamic Convolution to improve the visualization of tiny targets and multi-scale characteristics. The A2C2f-ACmix attention mechanism facilitates the collaborative modeling of both local and global features, whereas the Normalized Gaussian Wasserstein Distance module substitutes the conventional loss function to enhance resilience against minor positional shifts of objects. Experimental findings indicate that the proposed strategy enhances precision and recall by 15.2% and 20.0%, respectively, reduces the parameter count by 17.8%, and reduces the computing burden by 31.1%. The mAP50 and mAP50-95 metrics improve by 10.4% and 13.1%, respectively, with a detection speed of 588 FPS, satisfying real-time criteria. An experimental platform using the AUBO-i5 robot further substantiates the method’s efficacy and applicability in complex industrial settings.