<p>Nowadays printed circuit board plays a vital role in communication, computer, electronics and other industries. Existing PCB welding defect detection algorithms have the problems of low accuracy and poor real-time performance in identifying small or irregular targets and dense solder joints. This is due to the limited receptive field of standard convolutional kernels, which hinder global feature extraction and focus on local details. Moreover, the effects of kernel count and feature extraction dimensions are often overlooked, leading to the loss of important features. Conventional upsampling methods, such as nearest-neighbor interpolation, can further degrade critical information. To address these challenges, we propose FDDC-YOLO, a novel defect detection network. First, we introduce a new full-dimensional dynamic convolution module FDDC, which integrates full-dimensional dynamic convolution with the newly designed od_ottleneck structure to enhance the feature extraction ability by using the four dimensions of the convolution kernel. Secondly, the CECA attention module in the neck improves the ability of the model to detect small defects by enhancing the local interaction between channels. Third, the Dy-Up module is used to improve image resolution and prevent the loss of detailed information during the detection process. Finally, we replace the CIoU loss with IShapeIoU to reduce the overlap of detection boxes in densely packed solder joints, improving both localization accuracy and convergence speed.The mAP of FDDC-YOLO is improved by 5.4% on the PCBSP_dataset, and a Frame Per Second (FPS) of 189. It improves by 3.8% on the public PCB Defect-Augmented dataset, which proves its good generalization ability.</p>