CN-122023258-A - Online detection method for angular deformation of thick plate T-shaped joint

CN122023258ACN 122023258 ACN122023258 ACN 122023258ACN-122023258-A

Abstract

The invention discloses an online detection method for angle deformation of a thick plate T-shaped joint, which comprises the steps of designing a multi-dimensional feature aggregation module, designing a feature aggregation ladder pyramid structure network by combining the multi-dimensional feature aggregation module with a small target detection layer, designing a high-efficiency lightweight detection head to reduce calculation parameters of the network and improve detection instantaneity of a model, and providing a cascading processing algorithm based on image segmentation-refinement-feature point detection in the aspect of angle characterization, and accurately extracting two plate edge feature points and two plate coordinates and fitting a spatial position relation. The invention realizes the light weight and the recognition accuracy of the weld characteristic point recognition model, can accurately extract the characteristic points and coordinates of the edges of the two plates and fit the spatial position relationship, and is beneficial to promoting the floor popularization of the laser vision sensing technology in automatic and intelligent welding.

Inventors

YU ZHUOHUA
BAI LU
Hu Zhaoxiao
JIA YE
HE YINSHUI
WU LINGQING
Wang Biaolong
Yao Mugui
GAO MINGLUN
LUO QUAN

Assignees

南昌交通学院

Dates

Publication Date: 20260512
Application Date: 20251229

Claims (6)

1. An on-line detection method for angular deformation of a thick plate T-shaped joint is characterized by comprising the following steps: S1, designing a multidimensional feature aggregation module; s2, designing a characteristic aggregation ladder pyramid structure network by combining the multi-dimensional characteristic aggregation module with the small target detection layer; s3, designing a high-efficiency lightweight detection head to reduce the calculation parameters of a network and improve the detection instantaneity of a model; s4, in the aspect of angle representation, a cascading processing algorithm based on image segmentation-refinement-feature point detection is provided, and two plate edge feature points and two plate coordinates are accurately extracted and a spatial position relation is fitted.
2. The on-line detection method of angular deformation of a thick plate T-joint according to claim 1, wherein the multi-dimensional feature aggregation module in step S1 comprises: S11, selecting features under different scales as input, unifying dimensions through downsampling and upsampling operations, and then fusing all the input features; S12, optimizing the depth separable convolution by using different convolution kernel sizes, and outputting the depth separable convolution as a multi-receptive field feature; S13, introducing a residual structure, and adding the characteristics of the original input and the multi-branch convolution output of the module element by element through jump connection.
3. The method for on-line detection of angular deformations of a thick plate T-joint according to claim 1, characterized in that said network of characteristic converging stepped pyramid structures in step S2 comprises: s21, carrying out continuous convolution treatment on the characteristic diagram Z, and adding a C3k2 module to obtain a treatment result Z1; s22, performing convolution treatment on Z1 once, and then adding the convolutionally treated Z1 into a C3k2 module again to obtain Z2; s23, performing convolution treatment on Z2 once, and then adding the convolutionally treated Z2 into the C3k2 module again to obtain Z3; s24, adding the C3k2 module again after carrying out convolution treatment on the Z3 once again, and adding the SPPF module and the C2PSA module to obtain Z4; s25, after the Z2, the Z3 and the Z4 are processed, adding MSFAM modules to obtain Z5; s26, performing up-sampling on Z5, performing feature fusion with Z2, and adding the Z5 into a C3k2 module to obtain Z6; s27, performing up-sampling on Z6, performing feature fusion on the Z6 and Z1, and adding the Z6 into a C3k2 module to obtain Z7; S28, adding Z5, Z6 and Z7 into MSFAM modules to obtain Z8; and S29, adding Z4, Z5 and Z8 into a detection module respectively, implementing in parallel, and executing a final target detection task.
4. The on-line detection method for angular deformation of thick plate T-joint according to claim 1, wherein the processing steps of the input feature map by the high-efficiency lightweight detection head in step S3 are as follows: s31, continuously passing the input multi-scale feature map through two Group-Conv 3X 3 modules, extracting lightweight features and enhancing channel interaction; s32, inputting the outputs of the group convolution into Conv2d_Reg and Conv2d_Cls respectively.
5. The method for on-line detection of angular deformation of a thick plate T-joint according to claim 1, wherein said step S4 comprises: s41, acquiring a welding picture through a welding system, and then processing the image by utilizing a unified segmentation frame to generate a corresponding binary image; S42, carrying out refinement treatment on the obtained binary image to obtain a stripe structure with each column only containing a single pixel, and further extracting all coordinate points with the pixel value of 255; s43, determining the coordinates of the end points at the two ends of the stripe by sequencing the coordinate points; s44, extracting key point position information of the feature points by using the optimized YOLOv n network; S45, fusing the coordinate information acquired by the two methods to generate a fused image; and S46, fitting visual characterization information of the angular deformation according to the position coordinates of the two plates on the image.
6. The method for on-line detection of angular deformation of a thick plate tee joint of claim 5, wherein fitting the position coordinates on the image to the visual representation of angular deformation comprises: Calculating based on the included angle between the two plates; and calculating based on the included angle between the groove and the bottom plate.

Description

Online detection method for angular deformation of thick plate T-shaped joint Technical Field The invention relates to an automatic welding technology, in particular to an online detection method for angular deformation of a thick plate T-shaped joint. Background At present, thick plate T-shaped joints are typical connection forms of prefabricated steel structures of ocean engineering equipment, and welding quality of the thick plate T-shaped joints directly influences safety and reliability of the structures. At present, the components mainly adopt an arc multilayer multi-pass welding process, but residual stress and angular deformation caused by local heating and rapid cooling in the welding process seriously weaken the structural performance. The research shows that the real-time monitoring of the angular deformation can provide key basis for weld quality evaluation and dynamic regulation and control of welding parameters. Angular deformation is usually expressed in a displacement manner, and is specifically expressed as a displacement change before and after web welding, and the displacement can directly reflect the degree of angular deformation generated in the welding process. However, the severe environments such as high temperature, smoke dust, splashing and the like limit the precision of the traditional measuring means, and the synchronous measurement of multi-point three-dimensional deformation and the capture of a dynamic evolution process are difficult to realize. In recent years, the machine vision technology has shown advantages in welding deformation measurement, but the prior art is limited to static T-state measurement after welding, and lacks an online real-time measurement means for T-joint web angular deformation. The technical bottleneck causes difficulty in realizing closed-loop control of the welding process based on angular deformation feedback, and also restricts the construction of an online dynamic welding quality monitoring system taking angular deformation control as a core. Therefore, the development of the online measurement technology suitable for the angular deformation of the T-shaped joint web has important engineering significance in the aspects of enhancing the intellectualization of the laser vision sensing technology in the welding process, improving the automatic welding quality and the like. Disclosure of Invention The invention mainly aims to provide an online detection method for angular deformation of a thick plate T-shaped joint, and aims to realize synchronous measurement of multi-point three-dimensional deformation and capture of a dynamic evolution process in a welding process, and a welding quality online dynamic monitoring system taking angular deformation control as a core is constructed. The technical scheme adopted by the invention is that the online detection method for the angular deformation of the T-shaped joint of the thick plate comprises the following steps: S1, designing a multidimensional feature aggregation module; s2, designing a characteristic aggregation ladder pyramid structure network by combining the multi-dimensional characteristic aggregation module with the small target detection layer; s3, designing a high-efficiency lightweight detection head to reduce the calculation parameters of a network and improve the detection instantaneity of a model; s4, in the aspect of angle representation, a cascading processing algorithm based on image segmentation-refinement-feature point detection is provided, and two plate edge feature points and two plate coordinates are accurately extracted and a spatial position relation is fitted. Further, the multi-dimensional feature aggregation module in step S1 includes: S11, selecting features under different scales as input, unifying dimensions through downsampling and upsampling operations, and then fusing all the input features; S12, optimizing the depth separable convolution by using different convolution kernel sizes, and outputting the depth separable convolution as a multi-receptive field feature; S13, introducing a residual structure, and adding the characteristics of the original input and the multi-branch convolution output of the module element by element through jump connection. Still further, the feature aggregation ladder pyramid structure network in step S2 includes: s21, carrying out continuous convolution treatment on the characteristic diagram Z, and adding a C3k2 module to obtain a treatment result Z1; s22, performing convolution treatment on Z1 once, and then adding the convolutionally treated Z1 into a C3k2 module again to obtain Z2; s23, performing convolution treatment on Z2 once, and then adding the convolutionally treated Z2 into the C3k2 module again to obtain Z3; s24, adding the C3k2 module again after carrying out convolution treatment on the Z3 once again, and adding the SPPF module and the C2PSA module to obtain Z4; s25, after the Z2, the Z3 and the Z4 are processed, adding MSFAM