Search

Welding processes significantly impact product quality as a crucial part of industrial production. Due to the reflection, diversity, complexity, and minuteness of solder defects, traditional detection methods struggle to detect surface defects in solder points effectively. Although object detection based on deep learning has made significant advances, detecting smaller objects remains challenging. To address these issues, we propose an improved defect detection network based on YOLOv9, named attention flexible kernel YOLO (YOLO-AFK). In particular, we propose a fusion attention network (FANet) that can enhance the model’s ability to detect small defects by adaptively adjusting the receptive field of targets during feature extraction. Meanwhile, we use the alterable kernel convolution (AKConv), a variable kernel convolution, that breaks away from traditional convolutions limited to fixed local windows and sampling shapes. It can flexibly adjust the size and shape of the convolution kernels according to the solder targets, leading to more efficient feature extraction, thus achieving a lighter network. To gather more contextual and high-resolution information and enhance the detection accuracy and generalization ability for small objects and low-contrast targets, the cross-stage partial network fusion (C2f) module is designed to fuse feature maps from different levels. We evaluated the model using the publicly available NEU dataset and our proprietary solder point dataset, the fine-grain solder defect dataset (FG-SDD). Compared to previous studies, YOLO-AFK outperforms other state-of-the-art networks in terms of mean Average Precision (mAP) and Precision, with the parameter count increasing by only 12.4M, Precision improving by 10.1%, mAP increasing by 5.6%, and FPS improving by 23%. These results demonstrate the superior performance of the proposed network in detecting defects with complex structures. In particular, for industrial solder joint defect detection, YOLO-AFK not only improves detection accuracy but also significantly enhances the recognition of small targets and complex solder joint defects, showcasing the network’s substantial potential and practical value in real-world production environments. The code is available at: https://github.com/Lwsk-wxy/yolo_afk.git.

With the wide interconnection of power systems and extensive application of phasor measurement units (PMUs), the secure operation of power systems is facing considerable challenges. To satisfy the demand of online dynamic security assessment (DSA) for modern power systems, a data-driven scheme based on sparse projection oblique randomer forests (SPORF) is proposed, which includes offline training, periodic update and online assessment. In the first stage, an improved adaptive synthetic sampling (ADASYN) method is developed to mitigate the class imbalance problem for the data-driven DSA approach. Then, the SPORF-based DSA model is trained using crucial features with low redundancy selected by a feature selection procedure based on the minimal-redundancy-maximal-relevance (MRMR) criterion. In the second stage, the periodic update of the DSA model for unseen system topologies is executed to enhance the robustness of the model. In the third stage, the trained model can provide the DSA result immediately when the real-time operation information of a system is received. The satisfactory performance of the proposed scheme is demonstrated through a series of tests and the comparisons on a 23-bus system and a practical 1648-bus system.