In the realm of machine learning-based target detection, there exist several challenges that require attention, namely limited detection range, complex feature extraction, suboptimal detection precision, and significant subjectivity. In this paper, the strengths and weaknesses of existing deep learning target detection algorithms have been investigated in order to address these issues by integrating the actual welding process of heat exchangers. The objective is to improve the model’s detection accuracy and speed. To achieve this, we employ the YOLOv5 model to detect and identify weld defects of the heat exchanger tube plate, and propose an enhancement method based on the YOLOv5s model. By implementing several enhancements, such as incorporating the attention mechanism, updating the loss function, and optimizing the feature fusion network, the model’s overall performance is enhanced, with a focus on addressing the issues of low detection accuracy, slow convergence, and inadequate real-time performance in detecting small target defects compared to the YOLOv5s model. The improved YOLOv5s_m model improves the detection accuracy by 4.52% and the speed by 4.4 FPS, which solves the problems of low detection accuracy, weak sensitivity of small target defect detection and poor convergence of the bounding box loss function of the YOLOv5s model. These improvements lay the groundwork for enhancing the automation and intelligence of weld quality inspections.