CN-121982032-A - Welding seam defect real-time detection method for industrial welding scene

CN121982032ACN 121982032 ACN121982032 ACN 121982032ACN-121982032-A

Abstract

The invention relates to the technical field of industrial intelligent detection and discloses a welding seam defect real-time detection method for an industrial welding scene, which comprises the following steps of S1, image acquisition and preprocessing, acquisition of RGB/gray scale images of a welding seam area to be detected, scale normalization, light denoising and contrast enhancement, and generation of network input; S2, feature extraction, S3, feature fusion, S4, target detection, and S5, post-processing and output. According to the invention, the main feature extraction structure, the multi-scale feature fusion structure and the detection head structure are cooperatively optimized, so that the recognition capability of the model on the weld defects under the conditions of small-scale defects, low contrast defects and complex background interference is enhanced while the light weight and the real-time performance of the model are ensured, the detection precision, the robustness and the engineering deployment adaptability are improved, and the method is suitable for online weld quality detection and real-time edge equipment deployment scenes of industrial production lines.

Inventors

Ge Aidong
LI SHUCHANG
Zhang Diehao
Kang ao
ZHAO LUNING
QI HAOXIANG
FENG YU
Hou Qiuyu
LIU YUEYAO
MAO JIANYU
GUO ZHONGYUAN

Assignees

齐鲁工业大学(山东省科学院)

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (6)

1. The real-time detection method for the weld defects facing the industrial welding scene is characterized by comprising the following steps of: S1, image acquisition and preprocessing, namely acquiring RGB/gray scale images of a weld joint area to be detected, and performing scale normalization, light denoising and contrast enhancement to generate network input; s2, extracting characteristics, namely extracting multi-scale characteristics of the network input image by adopting a lightweight characteristic extraction backbone module; S3, feature fusion, namely inputting the multi-scale features output by the lightweight feature extraction trunk module into a bidirectional weighted multi-scale feature fusion module, carrying out normalized weighted fusion at fusion nodes by bidirectional information flow from top to bottom and multi-layer repeated stacking enhancement cross-scale feature interaction; S4, target detection, namely constructing a suture perception multi-scale attention detection head special for detecting weld seams or suture defects, and carrying out defect positioning and category prediction on the fused multi-scale features; S5, post-processing and outputting, adopting non-maximum suppression or soft non-maximum suppression and confidence level screening to generate a final detection frame, outputting position coordinates and category confidence, and providing real-time and visual warning to an upper system or an edge terminal.
2. The method for detecting the weld defects in real time for industrial welding scenes according to claim 1, wherein the specific step of S2 comprises the following steps: the local convolution divides the input feature into two parts according to the channel dimension, only 3X 3 convolution operation is carried out on one part, other channels are directly passed by a bypass, the number of input channels is d, the proportion of the parts is 1/n, the number of channels participating in the convolution is d/n, the corresponding floating point operation times is 1/n2 of the standard convolution, and the memory access quantity is 1/n of the original convolution; When the proportion of the part is 1/4, the calculated amount of the local convolution is only 1/16 of that of the common convolution, and meanwhile, the memory access cost is obviously reduced, so that the actual reasoning speed is effectively improved.
3. The method for detecting the weld defects in real time for the industrial welding scene according to claim 2, wherein in S2, the lightweight feature extraction backbone module is realized by adopting a lightweight fast neural network structure based on partial convolution and is embedded into a target detection frame so as to reduce model calculation complexity and reasoning delay.
4. The real-time detection method for weld defects facing industrial welding scenes according to claim 1, wherein in S3: at each scale feature fusion node, the bidirectional feature pyramid network distributes independent leachable weight parameters for different input branches, and completes feature fusion by adopting a normalized weighted summation mode, and a fusion expression is expressed as follows: (1) Wherein: Input for different scales; Is a learnable weight; is a numerical stable term.
5. The method for detecting the weld defects in real time for industrial welding scenes according to claim 4, wherein in S3: the bidirectional weighted multi-scale feature fusion module is used for replacing a traditional path aggregation structure in an original target detection frame, and a multi-layer stacked bidirectional feature pyramid fusion unit is adopted to enhance expression and interaction capability among different scale features.
6. The method for detecting the weld defects in real time for industrial welding scenes according to claim 1, wherein in S4: The suture sensing multi-scale attention detection head is constructed on the basis of the original decoupling detection head structure, and multi-scale separation enhancement attention sensing modules (MultiSEAM) are respectively embedded in the regression branch and the classification branch to enhance the multi-scale context information modeling capability and the discrimination capability of the weld joint related defect region; the suture sensing multi-scale attention detection head adopts a double-branch decoupling structure; In the regression branch, firstly, carrying out channel transformation and space feature extraction on input features through 3X 3 convolution, then introducing a multi-scale separation enhancement attention perception module to carry out self-adaptive weighting on multi-scale context information, and finally outputting boundary frame distribution parameters through 1X 1 convolution, and completing refined boundary frame regression by combining with distributed focus loss; In the classification branch, the features are subjected to light feature extraction by depth separable convolution and 1×1 convolution, then the discrimination features are enhanced by a multi-scale separation enhancement attention perception module, and finally the prediction probability of each category is output by the 1×1 convolution.

Description

Welding seam defect real-time detection method for industrial welding scene Technical Field The invention relates to the technical field of industrial intelligent detection, in particular to a real-time detection method for weld defects of an industrial welding scene. Background In the industrial fields of aerospace, automobile manufacturing, rail transit, energy equipment and the like, welded structures are widely used, and the quality of the welded structures is directly related to the safety and service reliability of products. Defects such as common air holes, cracks, unfused and slag inclusion in the welding line have the characteristics of small size, irregular shape, fuzzy boundary and the like, and if the defects can not be found in time, structural failure and even safety accidents are extremely easy to occur in the subsequent service process. At present, the engineering site usually adopts traditional nondestructive testing methods such as ultrasonic testing, ray testing, manual visual inspection and the like, but the method generally has the problems of low testing efficiency, strong manual dependency, insufficient automation degree, limited complex defect identification capability and the like, and is difficult to meet the requirements of modern intelligent manufacturing on online and automatic testing. With the development of deep learning and computer vision technology, weld defect recognition methods based on object detection models such as YOLO, fast R-CNN, SSD and the like are gradually applied to the field of industrial detection. The YOLO series model has obvious advantages in real-time detection scenes due to an end-to-end structure and high detection speed. However, the following technical bottlenecks are still faced with directly adopting the general YOLO model: 1. the weld defect size is usually smaller and the form is complex, and the general detection model is easy to leak detection or misdetection when the multi-scale feature expression capability is insufficient; 2. the complex background and noise interference are serious, namely strong reflection, texture interference and weld body structure noise exist in the weld image, and the discrimination capability of the model is easily weakened; 3. In engineering application, especially in the deployment scene of edge equipment, the real-time performance and the precision are difficult to be considered, the model reasoning speed and the computing resource consumption are strictly limited, and a high-precision model is often accompanied by larger parameter quantity and computing complexity. Therefore, in order to solve the above problems, a method for detecting the weld defects in real time for industrial welding is provided. Disclosure of Invention The technical problem to be solved by the invention is to provide the welding seam defect real-time detection method for the industrial welding scene, which aims at improving the main feature extraction structure, the multi-scale feature fusion module and the detection head structure, taking the small target recognition capability and the real-time property of the edge equipment into consideration, improving the accuracy and the robustness of the welding seam defect detection and facilitating the engineering deployment, thereby meeting the actual requirements of the on-line detection and the edge equipment deployment on the industrial production line. The invention adopts the following technical scheme to realize the aim of the invention: a welding seam defect real-time detection method facing industrial welding scenes comprises the following steps: S1, image acquisition and preprocessing, namely acquiring RGB/gray scale images of a weld joint area to be detected, and performing scale normalization, light denoising and contrast enhancement to generate network input; s2, extracting characteristics, namely extracting multi-scale characteristics of the network input image by adopting a lightweight characteristic extraction backbone module; S3, feature fusion, namely inputting the multi-scale features output by the lightweight feature extraction trunk module into a bidirectional weighted multi-scale feature fusion module, carrying out normalized weighted fusion at fusion nodes by bidirectional information flow from top to bottom and multi-layer repeated stacking enhancement cross-scale feature interaction; S4, target detection, namely constructing a suture perception multi-scale attention detection head special for detecting weld seams or suture defects, and carrying out defect positioning and category prediction on the fused multi-scale features; S5, post-processing and outputting, adopting non-maximum suppression or soft non-maximum suppression and confidence level screening to generate a final detection frame, outputting position coordinates and category confidence, and providing real-time and visual warning to an upper system or an edge terminal. As a further limitation of the present technic