In the machinery manufacturing industry, it is essential to ensure consistency and reliability of product quality. Traditional visual inspection methods often rely on manual operation, which is not only expensive but also easily influenced by subjective factors, so it is not easy to meet the needs of high-speed operation of modern production lines. In view of this, this study innovatively proposes an intelligent defect detection scheme for mechanical manufacturing parts based on transfer learning fused with the YOLO (You Only Look Once) algorithm, aiming to break through the existing bottleneck and improve detection efficiency and accuracy. The core of the project is to combine transfer learning technology in deep learning and the YOLO target detection algorithm to build an intelligent platform suitable for surface defect identification of various types of parts. First, the convolutional neural network on a large number of general-purpose images is pre-trained to achieve an effective transfer of knowledge from known to unknown domains to reduce the number of labelled samples required for new datasets. Secondly, with the help of YOLOv4’s fast positioning capability, it ensures that the second-level response speed can also be achieved when facing complex backgrounds or tiny defects, and there is no omission in continuous operation on the assembly line. Finally, in order to verify the effectiveness of the scheme, the engine block is selected as the test object for experiments, and more than 7,000 high-definition pictures covering common quality problems such as scratches, cracks and burrs are collected as training sets and test sets. After repeated tuning, the accuracy rate of the final model reaches 97.3%, and it only takes 0.03 seconds to process each image, far exceeding the limit of the human eye. Based on the transfer learning fusion YOLO algorithm, this paper experiments on the intelligent defect detection system of mechanical manufacturing parts, and draws the following conclusions: in the experiment of aluminum profile and cover part dataset, the coordinate attention mechanism has the best effect when the dimensionality reduction parameter r=16; Each improved module can improve the detection accuracy, such as +C+CA Attention is 0.5% higher than +C; The AP value of the attention mechanism network was increased by 1.45% in small-scale defect detection, and the large-scale detection was average. The comprehensive performance of the YOLO v5l model was the best. The optimal learning rate was 10 −3 and 150 iterations. Different models have different recognition effects, and Baseline and YOLO v5-Focal EIoU (Efficient Intersection over Union) have a high recognition rate of interface material shedding. The average accuracy of the final inspection system is 97%, the processing speed is 60 frames per second, and the stable detection rate is over 95% under harsh conditions, which significantly improves the detection quality and efficiency. In addition, the model can still maintain a detection success rate of more than 94% in low light or shadow occlusion environments, showing strong robustness.
Yong Wang (Thu,) studied this question.