Weld seam 3D reconstruction contour extraction system and method based on machine vision

Abstract

The present invention belongs to the field of welding technology and discloses a system and method for extracting a three-dimensional weld contour based on machine vision. The method comprises: collecting welding process parameters of N welding robots; collecting welding area images of the N welding robots; collecting welding thermal imaging images and welding spectral images of the N welding robots; performing feature extraction on the welding thermal imaging images and welding spectral images of the N welding robots to obtain welding spectral feature data and welding temperature feature data corresponding to the N welding robots; and reconstructing a three-dimensional weld contour based on the welding spectral feature data, welding temperature feature data and welding area images of the N welding robots. The present invention realizes accurate recognition and reconstruction of the weld morphology by the welding robot, overcomes the visual error problem caused by thermal distortion and arc interference in the traditional welding process, and improves the intelligence level of the welding automation system.

Description

Technical Field

[0001] The present invention relates to the field of welding technology, and more particularly to a system and method for extracting weld seam contours through three-dimensional reconstruction based on machine vision. Background Art

[0002] Welding technology, a crucial processing method in modern manufacturing, is widely used in various fields. Welding quality directly impacts a product's structural strength, sealing, and service life. Therefore, achieving high-precision and high-stability welding quality control has become a key focus in the industry. In recent years, with the advancement of intelligent manufacturing, the welding process has gradually evolved towards automation and intelligence. With the development of deep learning and image processing algorithms, the application of machine vision technology in weld inspection has become increasingly mature, playing a crucial role in intelligent welding systems.

[0003] However, in the actual welding process, due to the complexity of the welding environment, the change in air refractive index caused by high temperature and the strong interference of the welding arc, the weld recognition accuracy decreases, causing image distortion and increased signal noise in the machine vision system, thereby reducing the accuracy of weld contour extraction, affecting the accuracy of subsequent weld 3D modeling, and ultimately affecting the welding quality, making it difficult to meet the needs of high-precision welding control.

[0004] Based on this, there is an urgent need for a more accurate, efficient, and anti-interference visual inspection and weld reconstruction solution to achieve high-precision weld recognition, three-dimensional modeling, and intelligent welding control, so as to improve the welding quality and automation level of welding robots under complex working conditions. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned purpose, the present invention provides the following technical solution: a method for extracting the contour of a weld seam 3D reconstruction based on machine vision, comprising:

[0006] Collect welding process parameters of N welding robots;

[0007] Collect welding area images of N welding robots;

[0008] Collect welding thermal imaging images and welding spectrum images of N welding robots;

[0009] Feature extraction is performed on the welding thermal imaging images and welding spectrum images of N welding robots to obtain welding spectrum feature data and welding temperature feature data corresponding to the N welding robots;

[0010] The three-dimensional contour of the weld is reconstructed based on the welding spectrum characteristic data, welding temperature characteristic data and welding area images of N welding robots.

[0011] Furthermore, the method for reconstructing the three-dimensional profile of the weld includes:

[0012] Inputting the welding spectrum characteristic data and the welding temperature characteristic data of N welding robots into the modeling diagnosis model respectively to obtain the modeling diagnosis results of the N welding robots; the modeling diagnosis results include the welds that can be directly modeled and the welds that cannot be directly modeled;

[0013] Extracting weld contours from weld area images whose modeling diagnosis results indicate that weld modeling can be performed directly, and adding the extracted weld contours to the weld reconstructed three-dimensional image;

[0014] According to the preset method, the weld area image whose modeling diagnosis result indicates that it is not possible to directly model the weld is marked to obtain the weld area image to be processed:

[0015] The welding spectrum characteristic data and welding temperature characteristic data corresponding to the image of the weld area to be processed are input into the parameter setting model to obtain the clarity adjustment parameters. The image of the weld area to be processed is adjusted using the clarity adjustment parameters to obtain a clear processed image. The weld contour of the clear processed image is extracted and added to the reconstructed three-dimensional weld image.

[0016] Furthermore, the method for acquiring the image of the weld area to be processed includes:

[0017] Based on the welding spectrum characteristic data, the optical flow high disturbance area is extracted and marked on the welding area image to obtain the weld area image to be processed;

[0018] Based on the welding temperature characteristic data, the high thermal distortion area, the medium thermal distortion area where the stress concentration factor exceeds the preset stress concentration factor threshold, and the low thermal distortion area where the boundary eccentricity exceeds the preset boundary eccentricity threshold are marked in the welding area image to obtain the image of the weld area to be processed.

[0019] Furthermore, the method for acquiring welding spectrum characteristic data corresponding to N welding robots includes:

[0020] S100: Let the initial value of n be 1, and the value range of n is 1 to N; let the initial value of b be 1, and b is the counting variable of the optical flow high disturbance area;

[0021] S101: Divide the welding spectrum image of the nth welding robot into H spectrum pixels, and calculate the corresponding optical flow disturbance degree according to the image brightness of the H spectrum pixels;

[0022] S102: Marking spectral pixels whose optical flow disturbance degree is greater than a preset optical flow disturbance degree threshold as optical flow high disturbance pixels;

[0023] S103: Select one optical flow high disturbance pixel from all optical flow high disturbance pixels as an optical flow high disturbance pixel seed, and based on the optical flow high disturbance pixel seed, expand to all adjacent optical flow high disturbance pixels of the same type to form the bth optical flow high disturbance area; and continue to expand to adjacent optical flow high disturbance pixels of the same type until there are no optical flow high disturbance pixels that meet the conditions to be added to the bth optical flow high disturbance area; let b=b+1, repeat S103 until all optical flow high disturbance pixels are assigned to the corresponding optical flow high disturbance areas, and finally obtain B optical flow high disturbance areas;

[0024] S104: Constructing B optical flow high disturbance areas and corresponding optical flow disturbance degrees into welding spectrum feature data of the nth welding robot;

[0025] S105: Let n=n+1. If n is less than or equal to N, continue executing S101 to S104; if n is greater than N, end the current process.

[0026] Furthermore, the method for obtaining the welding temperature characteristic data corresponding to N welding robots includes:

[0027] S200: let the initial value of n be 1;

[0028] S201: Inputting the welding process parameters of the nth welding robot into a threshold setting model to obtain a pixel temperature threshold 1 and a pixel temperature threshold 2 corresponding to the welding thermal imaging image of the nth welding robot; the pixel temperature threshold 1 is less than the pixel temperature threshold 2;

[0029] The welding thermal imaging image of the nth welding robot is divided into M hot pixels; the temperatures corresponding to the M hot pixels are obtained and recorded as hot pixel temperatures; the hot pixels whose hot pixel temperatures are greater than or equal to the pixel temperature threshold 2 are marked as high thermal distortion pixels, the hot pixels whose hot pixel temperatures are less than the pixel temperature threshold 2 and greater than or equal to the pixel temperature threshold 1 are marked as medium thermal distortion pixels, and the hot pixels whose hot pixel temperatures are less than the pixel temperature threshold 1 are marked as low thermal distortion pixels;

[0030] The welding thermal imaging image of the nth welding robot is divided based on high thermal distortion pixels, medium thermal distortion pixels and low thermal distortion pixels, and Q high thermal distortion areas, Y medium thermal distortion areas and R low thermal distortion areas are obtained;

[0031] S202: extracting features of Q high thermal distortion regions, Y medium thermal distortion regions, and R low thermal distortion regions according to a preset method to obtain welding temperature feature data of the nth welding robot;

[0032] S203: Let n=n+1. If n is less than or equal to N, continue executing S201 to S202; if n is greater than N, end the current process.

[0033] Furthermore, a method for obtaining Q high thermal distortion regions includes:

[0034] S300: Set the initial value of q to 1, where q is a high thermal distortion area counting variable;

[0035] S301: Select one high thermal distortion pixel from all high thermal distortion pixels as a high thermal distortion pixel seed, and based on the high thermal distortion pixel seed, expand to all adjacent high thermal distortion pixels of the same type to form a qth high thermal distortion region; and continue to expand to adjacent high thermal distortion pixels of the same type until there are no more high thermal distortion pixels that meet the conditions to be added to the qth high thermal distortion region; then execute S302;

[0036] S302: If there are high thermal distortion pixels that have not yet been assigned to corresponding high thermal distortion regions, set q = q + 1 and execute S301. If all high thermal distortion pixels have been assigned to corresponding high thermal distortion regions, stop the current process, and ultimately obtain Q high thermal distortion regions.

[0037] The same method for obtaining Q high thermal distortion regions is used to obtain Y medium thermal distortion regions and R low thermal distortion regions.

[0038] Furthermore, the method for obtaining the welding temperature characteristic data of the nth welding robot includes:

[0039] Let the initial value of q be 1, and the value range of q is 1 to Q; the boundary disturbance index is calculated based on the number and size of the hot pixels in the qth high thermal distortion area of ​​the nth welding robot; let q = q + 1, when q is less than or equal to Q, continue to calculate the boundary disturbance index based on the number and size of the hot pixels in the qth high thermal distortion area of ​​the nth welding robot, and when q is greater than Q, stop the execution and obtain the boundary disturbance index corresponding to the Q high thermal distortion areas;

[0040] Let the initial value of y be 1, and the value range of y be 1 to Y; calculate the stress concentration factor based on the temperature of the hot pixel in the yth medium thermal distortion area of ​​the nth welding robot; let y = y + 1, when y is less than or equal to Y, continue to calculate the stress concentration factor based on the temperature of the hot pixel in the yth medium thermal distortion area of ​​the nth welding robot. When y is greater than Y, stop the calculation and obtain the stress concentration factors corresponding to the Y medium thermal distortion areas;

[0041] Let r be initialized to 1, and the value range of r is 1 to R; construct the r-th low thermal distortion region of the n-th welding robot into a minimum circumscribed ellipse, and calculate the boundary eccentricity based on the major axis length and minor axis length of the minimum circumscribed ellipse; let r = r + 1, when r is less than or equal to R, continue to construct the r-th low thermal distortion region of the n-th welding robot into a minimum circumscribed ellipse, and calculate the boundary eccentricity based on the major axis length and minor axis length of the minimum circumscribed ellipse. When r is greater than R, stop the execution and obtain the boundary eccentricities corresponding to the R low thermal distortion regions;

[0042] The Q high thermal distortion areas and the corresponding boundary disturbance index, Y medium thermal distortion areas and the corresponding stress concentration factor, and R low thermal distortion areas and the corresponding boundary eccentricity of the nth welding robot are constructed into the corresponding welding temperature characteristic data.

[0043] Furthermore, the training method of the modeling diagnosis model includes:

[0044] Pre-constructing a modeling diagnosis data set, the modeling diagnosis data set including C groups of modeling diagnosis data and modeling diagnosis results corresponding to the C groups of modeling diagnosis data, where C is a positive integer greater than 0, and the modeling diagnosis data including welding spectrum feature data and welding temperature feature data; dividing the modeling diagnosis data set into a modeling diagnosis data training set and a modeling diagnosis data validation set, wherein the modeling diagnosis data training set is used for parameter learning of the modeling diagnosis model, and the modeling diagnosis data validation set is used for real-time evaluation of the generalization ability of the modeling diagnosis model;

[0045] During the training process of the modeling diagnosis model, a deep neural network structure based on multi-layer perceptron is used to convert the modeling diagnosis data into a high-dimensional feature vector as input, and the nonlinear features in the data are extracted through several hidden layers. Finally, the softmax activation function is used in the output layer to generate the probability distribution of the modeling diagnosis results, and the modeling diagnosis result corresponding to the maximum probability is output as the final prediction result; the training process aims to minimize the cross-entropy loss function, and at the same time introduces an early stopping strategy to monitor the performance of the modeling diagnosis data validation set. When the prediction accuracy on the modeling diagnosis data validation set reaches the preset threshold, the modeling diagnosis model is considered to have converged and the training of the modeling diagnosis model is stopped.

[0046] Furthermore, the training method of the parameter setting model includes:

[0047] Preliminarily collecting a parameter setting data set, the parameter setting data set comprising J groups of parameter setting data and clarity adjustment parameters corresponding to the J groups of parameter setting data, where J is a positive integer greater than 0, the parameter setting data comprising welding spectrum characteristic data and welding temperature characteristic data; dividing the parameter setting data set into a training set and a validation set, the training set being used to learn parameter setting model parameters, and the validation set being used to evaluate the generalization ability of the parameter setting model to avoid overfitting;

[0048] During the training process of the parameter setting model, a dynamic learning rate adjustment strategy is combined with an early stopping mechanism to minimize the cross-entropy loss function as the optimization goal. When the performance of the validation set reaches the preset performance requirements, training is automatically stopped to ensure the convergence of the parameter setting model. The parameter setting model is implemented based on the neural network model architecture. The neural network model consists of an input layer, a hidden layer, and an output layer. The input layer receives the parameter setting data and converts it into a high-dimensional feature vector. The hidden layer uses an activation function to extract the complex nonlinear pattern of the parameter setting data. The output layer calculates the probability distribution of the clarified adjustment parameter through the softmax activation function, and uses the clarified adjustment parameter corresponding to the maximum probability as the prediction result and outputs it.

[0049] A machine vision-based weld 3D reconstruction contour extraction system implements the machine vision-based weld 3D reconstruction contour extraction method, comprising:

[0050] Parameter acquisition module, used to collect welding process parameters of N welding robots;

[0051] The first image acquisition module is used to acquire welding area images of N welding robots;

[0052] The second image acquisition module is used to acquire welding thermal imaging images and welding spectrum images of N welding robots;

[0053] A data processing module is used to extract features from the welding thermal imaging images and welding spectrum images of the N welding robots to obtain welding spectrum feature data and welding temperature feature data corresponding to the N welding robots;

[0054] The weld reconstruction module reconstructs the three-dimensional contour of the weld based on the welding spectrum characteristic data, welding temperature characteristic data and welding area image of N welding robots.

[0055] Compared with the existing technology, the technical effects and advantages of the weld 3D reconstruction contour extraction system and method based on machine vision of the present invention are as follows:

[0056] This solution acquires and analyzes welding spectral and temperature characteristic data to identify areas of high optical flow disturbance and thermal distortion. Combined with a modeling diagnostic model, it distinguishes between areas suitable for direct modeling and those awaiting processing. For areas not suitable for direct modeling, this solution uses a parameter setting model to dynamically generate sharpening adjustment parameters, precisely correcting visual image distortion and improving the visibility and accuracy of weld contours. Using weld contour extraction technology, the sharpened weld area is seamlessly integrated into the 3D reconstructed weld model, achieving highly accurate weld modeling.

[0057] Compared to existing technologies, the solution of the present invention can not only effectively reduce the interference of the welding environment on the machine vision of the welding robot, but also enable the welding robot to "see" a more realistic and distortion-free weld contour, thereby improving welding accuracy, welding quality, welding consistency and reliability. This solution breaks through the technical bottleneck of traditional weld identification, improves the intelligence level of the welding automation system, provides solid data support for high-precision welding quality control, and is suitable for multi-robot collaborative welding and high-precision manufacturing fields. By constructing a weld three-dimensional reconstruction contour extraction system and method based on machine vision, the present invention realizes the accurate identification and reconstruction of the weld morphology by the welding robot, overcoming the visual error problem caused by thermal distortion and arc interference in the traditional welding process. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of a weld seam 3D reconstruction and contour extraction system based on machine vision according to Example 1 of the present invention;

[0059] Figure 2 This is a flow chart of a method for extracting weld contours based on machine vision and three-dimensional reconstruction according to embodiment 2 of the present invention;

[0060] Figure 3 A flow chart of a method for reconstructing a three-dimensional profile of a weld;

[0061] Figure 4 Flowchart of a method for acquiring welding spectrum characteristic data corresponding to N welding robots;

[0062] Figure 5 The flowchart of the method for obtaining welding temperature characteristic data corresponding to N welding robots is shown in FIG. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present invention will be described in detail, clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. It should be noted that the specific embodiments described below are only used to better illustrate and describe the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and essence of the present invention, those skilled in the art may modify, adjust or make equivalent replacements based on the contents disclosed in the present invention, and these should all be regarded as the scope of protection of the present invention.

[0064] Example 1

[0065] See also Figure 1 As shown, the machine vision-based weld 3D reconstruction contour extraction system described in this embodiment includes a parameter acquisition module, a first image acquisition module, a second image acquisition module, a data processing module and a weld reconstruction module. Each module is connected through wired and / or wireless connections to achieve data transmission.

[0066] A parameter acquisition module is used to collect welding process parameters of N welding robots; the welding process parameters include welding current, welding voltage, welding speed, welding wire type, welding wire diameter and welding gas type; the welding current, welding voltage and welding speed are obtained through corresponding sensors, and the welding wire type, welding wire diameter and welding gas type are obtained from a welding process database.

[0067] It should be noted that welding current and welding voltage affect welding energy input, thereby affecting temperature distribution, and further affecting the degree of thermal distortion of the visual image; welding speed affects heat input and cooling rate, and thus affects temperature distribution; different welding wire types and wire diameters correspond to different melting temperatures, which affects the temperature range and temperature distribution of the weld; the type of welding gas affects the local temperature distribution and refractive index change.

[0068] The first image acquisition module is used to acquire welding area images of N welding robots; the welding area images are acquired by industrial cameras carried by the welding robots.

[0069] It should be noted that welding workpieces with complex curved structures (such as serpentine tubes) typically requires the collaborative operation of multiple welding robots to improve welding efficiency and ensure welding quality. However, due to the structural complexity of the workpiece and the parallel operation of multiple robots, different welding robots face different welding environmental factors (such as welding angle, weld shape, heat input distribution, arc interference, etc.) within their respective welding areas. These factors affect the precise construction of the weld's three-dimensional contour. Therefore, it is necessary to collect welding environment images and weld feature data from N welding robots in their respective welding areas, and comprehensively process the information collected by different welding robots to pave the way for the subsequent establishment of a globally consistent three-dimensional weld contour.

[0070] The second image acquisition module is used to acquire welding thermal imaging images and welding spectral images of N welding robots; the welding thermal imaging images are acquired by the thermal imaging camera carried by the welding robot, and the welding spectral images are acquired by the spectral imaging device (such as a spectral camera) carried by the welding robot.

[0071] The data processing module is used to extract features from the welding thermal imaging images and welding spectrum images of N welding robots to obtain welding spectrum feature data and welding temperature feature data corresponding to the N welding robots.

[0072] like Figure 4 As shown, the method for obtaining welding spectrum characteristic data corresponding to N welding robots includes:

[0073] S100: Let the initial value of n be 1, and the value range of n is 1 to N; let the initial value of b be 1, and b is the counting variable of the optical flow high disturbance area;

[0074] S101: Divide the welding spectrum image of the nth welding robot into H spectrum pixels, obtain the image brightness of the H spectrum pixels, and record it as the spectrum pixel brightness; calculate the optical flow disturbance degree corresponding to the H spectrum pixels according to the H spectrum pixel brightness;

[0075] S102: Marking spectral pixels whose optical flow disturbance degree is greater than a preset optical flow disturbance degree threshold as optical flow high disturbance pixels;

[0076] S103: Select one optical flow high disturbance pixel from all optical flow high disturbance pixels as an optical flow high disturbance pixel seed, and based on the optical flow high disturbance pixel seed, expand to all adjacent optical flow high disturbance pixels of the same type to form the bth optical flow high disturbance area; and continue to expand to adjacent optical flow high disturbance pixels of the same type until there are no optical flow high disturbance pixels that meet the conditions to be added to the bth optical flow high disturbance area; let b=b+1, repeat S103 until all optical flow high disturbance pixels are assigned to the corresponding optical flow high disturbance areas, and finally obtain B optical flow high disturbance areas;

[0077] S104: Constructing B optical flow high disturbance areas and corresponding optical flow disturbance degrees into welding spectrum feature data of the nth welding robot;

[0078] S105: Let n=n+1. If n is less than or equal to N, continue to execute S101 to S104; if n is greater than N, obtain the welding spectrum characteristic data corresponding to N welding robots and end the current process.

[0079] The method for obtaining the optical flow disturbance degree includes:

[0080] ;

[0081] in, is the optical flow perturbation degree of the h-th spectral pixel, is the brightness of the hth spectral pixel, is the brightness gradient of the hth spectral pixel in the horizontal direction, is the brightness gradient of the hth spectral pixel in the vertical direction.

[0082] It should be noted that the high brightness of the welding arc and its uneven radiation can cause overexposure or underexposure in localized areas of the visual image, reducing the contrast of the weld boundary in the image, thereby affecting the accuracy of weld contour recognition. The dynamic changes in the welding arc light create random optical flow perturbations in the visual image, causing the weld boundary to appear unrealistically deformed, thereby exacerbating visual thermal distortion. Furthermore, fluctuations in the welding arc light intensity lead to uneven brightness gradients in the weld area, making it difficult for traditional edge detection algorithms to accurately locate the weld contour, thus affecting the accuracy of 3D weld reconstruction.

[0083] like Figure 5 As shown, the method for obtaining the welding temperature characteristic data corresponding to N welding robots includes:

[0084] S200: let the initial value of n be 1;

[0085] S201: Inputting the welding process parameters of the nth welding robot into a threshold setting model to obtain a pixel temperature threshold 1 and a pixel temperature threshold 2 corresponding to the welding thermal imaging image of the nth welding robot; the pixel temperature threshold 1 is less than the pixel temperature threshold 2;

[0086] The welding thermal imaging image of the nth welding robot is divided into M hot pixels; the temperatures corresponding to the M hot pixels are obtained and recorded as hot pixel temperatures; the hot pixels whose hot pixel temperatures are greater than or equal to the pixel temperature threshold 2 are marked as high thermal distortion pixels, the hot pixels whose hot pixel temperatures are less than the pixel temperature threshold 2 and greater than or equal to the pixel temperature threshold 1 are marked as medium thermal distortion pixels, and the hot pixels whose hot pixel temperatures are less than the pixel temperature threshold 1 are marked as low thermal distortion pixels;

[0087] The welding thermal imaging image of the nth welding robot is divided based on high thermal distortion pixels, medium thermal distortion pixels and low thermal distortion pixels, and Q high thermal distortion areas, Y medium thermal distortion areas and R low thermal distortion areas are obtained;

[0088] S202: extracting features of Q high thermal distortion regions, Y medium thermal distortion regions, and R low thermal distortion regions according to a preset method to obtain welding temperature feature data of the nth welding robot;

[0089] S203: Let n=n+1. If n is less than or equal to N, continue to execute S201 to S202; if n is greater than N, obtain the welding temperature characteristic data of N welding robots and end the current process.

[0090] The training method of the threshold setting model includes:

[0091] Pre-constructing a threshold setting data set, the threshold setting data set including a group A of threshold setting data and a pixel temperature threshold 1 and a pixel temperature threshold 2 corresponding to the group A of threshold setting data, where A is a positive integer greater than 0, and the threshold setting data including welding process parameters; dividing the threshold setting data set into a threshold setting data training set and a threshold setting data validation set, wherein the threshold setting data training set is used for parameter learning of a threshold setting model, and the threshold setting data validation set is used for real-time evaluation of the generalization ability of the threshold setting model;

[0092] During the training process of the threshold setting model, a deep neural network structure based on multi-layer perceptron is adopted to convert the threshold setting data into a high-dimensional feature vector as input, extract the nonlinear features in the data through several hidden layers, and finally use the softmax activation function in the output layer to generate the probability distribution of pixel temperature threshold one and pixel temperature threshold two, and output the pixel temperature threshold one and pixel temperature threshold two corresponding to the maximum probability as the final prediction result; the training process aims to minimize the cross entropy loss function, and at the same time introduces an early stopping strategy to monitor the performance of the threshold setting data validation set. When the prediction accuracy on the threshold setting data validation set reaches the preset threshold, the threshold setting model is considered to have converged and the training of the threshold setting model is stopped.

[0093] The method for obtaining Q high thermal distortion regions includes:

[0094] S300: Set the initial value of q to 1, where q is a high thermal distortion area counting variable;

[0095] S301: Select one high thermal distortion pixel from all high thermal distortion pixels as a high thermal distortion pixel seed, and based on the high thermal distortion pixel seed, expand to all adjacent high thermal distortion pixels of the same type to form a qth high thermal distortion region; and continue to expand to adjacent high thermal distortion pixels of the same type until there are no more high thermal distortion pixels that meet the conditions to be added to the qth high thermal distortion region; then execute S302;

[0096] S302: If there are high thermal distortion pixels that have not yet been assigned to corresponding high thermal distortion regions, set q = q + 1 and execute S301. If all high thermal distortion pixels have been assigned to corresponding high thermal distortion regions, stop the current process and ultimately obtain Q high thermal distortion regions.

[0097] Methods for obtaining Y medium thermal distortion regions include:

[0098] S400: Let the initial value of y be 1, where y is the medium thermal distortion area counting variable;

[0099] S401: Select one moderately thermally distorted pixel from all moderately thermally distorted pixels as a moderately thermally distorted pixel seed. Based on the moderately thermally distorted pixel seed, expand to all adjacent moderately thermally distorted pixels of the same type to form a yth moderately thermally distorted region. Continue expanding to adjacent moderately thermally distorted pixels of the same type until no more moderately thermally distorted pixels meet the conditions and can be added to the yth moderately thermally distorted region. Execute S402.

[0100] S402: If there are moderate thermal distortion pixels that have not yet been assigned to a corresponding moderate thermal distortion region, set y = y + 1 and execute S401. If all moderate thermal distortion pixels have been assigned to a corresponding moderate thermal distortion region, the current process is terminated, ultimately obtaining Y moderate thermal distortion regions.

[0101] The method for obtaining R low thermal distortion regions includes:

[0102] S500: Let the initial value of r be 1, where r is the low thermal distortion area counting variable;

[0103] S501: Select one low thermal distortion pixel from all low thermal distortion pixels as a low thermal distortion pixel seed, and based on the low thermal distortion pixel seed, expand to all adjacent low thermal distortion pixels of the same type to form an r-th low thermal distortion region; and continue to expand to adjacent low thermal distortion pixels of the same type until there are no low thermal distortion pixels that meet the conditions to be added to the r-th low thermal distortion region; then execute S502;

[0104] S502: If there are low thermal distortion pixels that have not yet been assigned to the corresponding low thermal distortion region, set r = r = 1 and execute S501. If all low thermal distortion pixels have been assigned to the corresponding low thermal distortion region, stop the current process and finally obtain R low thermal distortion regions.

[0105] A method for extracting features of Q high thermal distortion regions, Y medium thermal distortion regions, and R low thermal distortion regions according to a preset method to obtain welding temperature feature data of an n-th welding robot includes:

[0106] Let the initial value of q be 1, and the value range of q is 1 to Q; the boundary disturbance index is calculated based on the number and size of the hot pixels in the qth high thermal distortion area of ​​the nth welding robot; let q = q + 1, when q is less than or equal to Q, continue to calculate the boundary disturbance index based on the number and size of the hot pixels in the qth high thermal distortion area of ​​the nth welding robot, and when q is greater than Q, stop the execution and obtain the boundary disturbance index corresponding to the Q high thermal distortion areas;

[0107] Let the initial value of y be 1, and the value range of y be 1 to Y; calculate the stress concentration factor based on the temperature of the hot pixel in the yth medium thermal distortion area of ​​the nth welding robot; let y = y + 1, when y is less than or equal to Y, continue to calculate the stress concentration factor based on the temperature of the hot pixel in the yth medium thermal distortion area of ​​the nth welding robot. When y is greater than Y, stop the calculation and obtain the stress concentration factors corresponding to the Y medium thermal distortion areas;

[0108] Let r be initialized to 1, and the value range of r is 1 to R; construct the r-th low thermal distortion region of the n-th welding robot into a minimum circumscribed ellipse, and calculate the boundary eccentricity based on the major axis length and minor axis length of the minimum circumscribed ellipse; let r = r + 1, when r is less than or equal to R, continue to construct the r-th low thermal distortion region of the n-th welding robot into a minimum circumscribed ellipse, and calculate the boundary eccentricity based on the major axis length and minor axis length of the minimum circumscribed ellipse. When r is greater than R, stop the execution and obtain the boundary eccentricities corresponding to the R low thermal distortion regions;

[0109] The Q high thermal distortion areas and the corresponding boundary disturbance index, Y medium thermal distortion areas and the corresponding stress concentration factor, and R low thermal distortion areas and the corresponding boundary eccentricity of the nth welding robot are constructed into the corresponding welding temperature characteristic data.

[0110] The method for obtaining the boundary disturbance index includes:

[0111] ;

[0112] in, is the boundary perturbation index corresponding to the qth high thermal distortion region, is the number of hot pixels in the qth high thermal distortion area, is the size of the hot pixel in the qth high thermal distortion area, is the material elastic modulus of the welding material.

[0113] It should be noted that the material elastic modulus is an inherent property of the welding material and reflects its ability to respond to external forces. For example, the material elastic modulus of steel is approximately 200 GPa, the material elastic modulus of aluminum is approximately 69 GPa, the material elastic modulus of copper is approximately 110 GPa, and the material elastic modulus of titanium alloy is approximately 120 GPa.

[0114] The method for obtaining the stress concentration factor includes:

[0115] ;

[0116] ;

[0117] ;

[0118] in, is the stress concentration factor corresponding to the yth medium thermal distortion area, is the maximum temperature of the hot pixel in the yth medium thermal distortion area, is the average temperature of the hot pixels in the yth medium thermal distortion area, is the number of hot pixels in the y-th medium thermal distortion area, is the temperature of the i-th hot pixel in the y-th medium thermal distortion area, and e is a constant;

[0119] is the total thermal stress gradient of the yth medium thermal distortion region, is the thermal stress gradient of the ith thermal pixel in the yth medium thermal distortion area in the horizontal direction, is the thermal stress gradient in the vertical direction of the ith thermal pixel in the yth medium thermal distortion area, is a logarithmic function, 、 and The thermal stress normalization coefficient is set by a person skilled in the art as an initial value, and is optimized and adjusted through multiple experiments to finally determine the optimal parameter value, which is the corresponding thermal stress normalization coefficient.

[0120] The method for obtaining the boundary eccentricity includes:

[0121] ;

[0122] in, is the boundary eccentricity of the rth low thermal distortion region, is the major axis length of the minimum circumscribed ellipse of the rth low thermal distortion region, is the minor axis length of the minimum circumscribed ellipse of the rth low thermal distortion region. The major axis length is the length of the longest symmetry axis in the ellipse, and the minor axis length is the length of the shortest symmetry axis perpendicular to the major axis.

[0123] It should be noted that during the welding process, the temperature of the weld and the surrounding metal surface can reach hundreds to thousands of degrees Celsius, causing the air to expand due to the heat, creating a local temperature gradient. According to Snell's Law, light refracts when propagating through media of varying densities. In the high-temperature welding area, air layers with varying temperatures result in different refractive indices for light. When light passes through these non-uniform gas layers, optical distortion occurs, causing the visual image to appear blurry, distorted, or jittery—what is known as thermal distortion. Specifically, due to the perturbation in the light propagation path, the light signal received by the camera sensor becomes unstable due to variations in air density, resulting in a "wavy" blur effect on the image. Furthermore, the local non-uniformity of the light's refractive index can distort the weld boundary in the visual image, affecting the accurate recognition of weld features and ultimately reducing the accuracy of the 3D weld modeling.

[0124] High thermal distortion areas manifest as dramatic fluctuations in the weld contour, causing the weld boundary to appear irregular in the welding thermal imaging image, thereby inducing thermal distortion in the visual image and directly affecting the accuracy of weld boundary recognition. Although the boundary morphology of the medium thermal distortion area is relatively stable, the significant thermal stress gradient and local stress concentration phenomenon are obvious, which in turn causes local deformation of the weld in the robot visual image, resulting in errors in the weld contour during digital processing and affecting the accuracy of the automatic welding system. The weld contour morphology in the low thermal distortion area changes less, but there is still a slight offset, resulting in a decrease in weld recognition accuracy.

[0125] This solution calculates the boundary disturbance index for high thermal distortion regions to quantify the complexity and irregularity of the weld boundary; calculates the stress concentration factor for moderate thermal distortion regions to assess the distribution of thermal stress and its impact on weld stability; and calculates the boundary eccentricity for low thermal distortion regions to characterize the overall offset of the weld morphology. By calculating different thermal distortion characteristics for different types of thermal distortion regions, it achieves precise quantification of thermal distortion in visual images, providing high-precision data support for weld identification and 3D modeling, further enhancing the stability and intelligence of the welding automation system.

[0126] The weld reconstruction module reconstructs the three-dimensional contour of the weld based on the welding spectrum characteristic data, welding temperature characteristic data and welding area image of N welding robots.

[0127] like Figure 3 As shown in FIG, the method for reconstructing the three-dimensional contour of the weld includes:

[0128] Inputting the welding spectrum characteristic data and the welding temperature characteristic data of N welding robots into the modeling diagnosis model respectively to obtain the modeling diagnosis results of the N welding robots; the modeling diagnosis results include the welds that can be directly modeled and the welds that cannot be directly modeled;

[0129] Extracting weld contours from weld area images whose modeling diagnosis results indicate that weld modeling can be performed directly, and adding the extracted weld contours to the weld reconstructed three-dimensional image;

[0130] According to the preset method, the weld area image whose modeling diagnosis result indicates that it is not possible to directly model the weld is marked to obtain the weld area image to be processed:

[0131] The welding spectrum characteristic data and welding temperature characteristic data corresponding to the image of the weld area to be processed are input into the parameter setting model to obtain the clarity adjustment parameters. The image of the weld area to be processed is adjusted using the clarity adjustment parameters to obtain a clear processed image. The weld contour of the clear processed image is extracted and added to the reconstructed three-dimensional weld image.

[0132] It should be noted that the image sharpening adjustment parameters include a contrast enhancement factor (increasing the contrast between the weld area and the background to make the weld boundary easier to identify), an adaptive gamma correction value (adjusting the image's brightness distribution to reduce local overexposure or underexposure caused by the welding arc), an edge sharpening parameter (enhancing weld boundary clarity and improving contour extraction accuracy), and an optical flow compensation weight (performing motion compensation in areas of high optical flow disturbance to reduce pixel shifts caused by thermal airflow). By dynamically generating optimal sharpening parameters by inputting welding spectral and temperature characteristic data, rather than using a fixed algorithm, the weld reconstruction effect dynamically adapts to different welding environments, reducing the complexity of manually setting parameters.

[0133] The method for obtaining the image of the weld area to be processed includes:

[0134] Based on the welding spectrum characteristic data, the optical flow high disturbance area is extracted and marked on the welding area image to obtain the weld area image to be processed;

[0135] Based on the welding temperature characteristic data, the high thermal distortion area, the medium thermal distortion area where the stress concentration factor exceeds the preset stress concentration factor threshold, and the low thermal distortion area where the boundary eccentricity exceeds the preset boundary eccentricity threshold are marked in the welding area image to obtain the image of the weld area to be processed.

[0136] The training method of the modeling diagnosis model includes:

[0137] Pre-constructing a modeling diagnosis data set, the modeling diagnosis data set including C groups of modeling diagnosis data and modeling diagnosis results corresponding to the C groups of modeling diagnosis data, where C is a positive integer greater than 0, and the modeling diagnosis data including welding spectrum feature data and welding temperature feature data; dividing the modeling diagnosis data set into a modeling diagnosis data training set and a modeling diagnosis data validation set, wherein the modeling diagnosis data training set is used for parameter learning of the modeling diagnosis model, and the modeling diagnosis data validation set is used for real-time evaluation of the generalization ability of the modeling diagnosis model;

[0138] During the training process of the modeling diagnosis model, a deep neural network structure based on multi-layer perceptron is used to convert the modeling diagnosis data into a high-dimensional feature vector as input, and the nonlinear features in the data are extracted through several hidden layers. Finally, the softmax activation function is used in the output layer to generate the probability distribution of the modeling diagnosis results, and the modeling diagnosis result corresponding to the maximum probability is output as the final prediction result; the training process aims to minimize the cross-entropy loss function, and at the same time introduces an early stopping strategy to monitor the performance of the modeling diagnosis data validation set. When the prediction accuracy on the modeling diagnosis data validation set reaches the preset threshold, the modeling diagnosis model is considered to have converged and the training of the modeling diagnosis model is stopped.

[0139] The training method of the parameter setting model includes:

[0140] Preliminarily collecting a parameter setting data set, the parameter setting data set comprising J groups of parameter setting data and clarity adjustment parameters corresponding to the J groups of parameter setting data, where J is a positive integer greater than 0, the parameter setting data comprising welding spectrum characteristic data and welding temperature characteristic data; dividing the parameter setting data set into a training set and a validation set, the training set being used to learn parameter setting model parameters, and the validation set being used to evaluate the generalization ability of the parameter setting model to avoid overfitting;

[0141] During the training process of the parameter setting model, a dynamic learning rate adjustment strategy is combined with an early stopping mechanism to minimize the cross-entropy loss function as the optimization goal. When the performance of the validation set reaches the preset performance requirements, training is automatically stopped to ensure the convergence of the parameter setting model. The parameter setting model is implemented based on the neural network model architecture. The neural network model consists of an input layer, a hidden layer, and an output layer. The input layer receives the parameter setting data and converts it into a high-dimensional feature vector. The hidden layer uses an activation function to extract the complex nonlinear pattern of the parameter setting data. The output layer calculates the probability distribution of the clarified adjustment parameter through the softmax activation function, and uses the clarified adjustment parameter corresponding to the maximum probability as the prediction result and outputs it.

[0142] The softmax activation function is:

[0143] ;

[0144] in, is the output probability corresponding to the num-th eigenvector, is the num-th eigenvector, is the total number of eigenvectors, and e is a constant.

[0145] Example 2

[0146] See also Figure 2 As shown, this embodiment provides a weld seam 3D reconstruction contour extraction method based on machine vision, including:

[0147] Collect welding process parameters of N welding robots;

[0148] Collect welding area images of N welding robots;

[0149] Collect welding thermal imaging images and welding spectrum images of N welding robots;

[0150] Feature extraction is performed on the welding thermal imaging images and welding spectrum images of N welding robots to obtain welding spectrum feature data and welding temperature feature data corresponding to the N welding robots;

[0151] The three-dimensional contour of the weld is reconstructed based on the welding spectrum characteristic data, welding temperature characteristic data and welding area images of N welding robots.

Claims

1. A weld seam 3D reconstruction contour extraction method based on machine vision, characterized in that: include: Collect welding process parameters of N welding robots; Collect welding area images of N welding robots; Collect welding thermal imaging images and welding spectrum images of N welding robots; Feature extraction is performed on the welding thermal imaging images and welding spectrum images of N welding robots to obtain welding spectrum feature data and welding temperature feature data corresponding to the N welding robots; Reconstruct the three-dimensional contour of the weld based on the welding spectrum characteristic data, welding temperature characteristic data and welding area image of N welding robots; Methods for reconstructing the 3D contour of welds include: Inputting the welding spectrum characteristic data and the welding temperature characteristic data of N welding robots into the modeling diagnosis model respectively to obtain the modeling diagnosis results of the N welding robots; the modeling diagnosis results include the welds that can be directly modeled and the welds that cannot be directly modeled; Extracting weld contours from weld area images whose modeling diagnosis results indicate that weld modeling can be performed directly, and adding the extracted weld contours to the weld reconstructed three-dimensional image; According to a preset method, the image of the welding area whose modeling diagnosis result indicates that the weld cannot be directly modeled is marked to obtain an image of the weld area to be processed; The welding spectrum characteristic data and welding temperature characteristic data corresponding to the image of the weld area to be processed are input into the parameter setting model to obtain the clarity adjustment parameters, and the image of the weld area to be processed is adjusted using the clarity adjustment parameters to obtain a clear processed image; the weld contour of the clear processed image is extracted and added to the weld reconstructed three-dimensional image; the clear adjustment parameters include a contrast enhancement coefficient, an adaptive gamma correction value, an edge sharpening parameter and an optical flow compensation weight.

2. The method for extracting weld contours by three-dimensional reconstruction based on machine vision according to claim 1, characterized in that: The method for obtaining the image of the weld area to be processed includes: Based on the welding spectrum characteristic data, the optical flow high disturbance area is extracted and marked on the welding area image to obtain the weld area image to be processed; Based on the welding temperature characteristic data, the high thermal distortion area, the medium thermal distortion area where the stress concentration factor exceeds the preset stress concentration factor threshold, and the low thermal distortion area where the boundary eccentricity exceeds the preset boundary eccentricity threshold are marked in the welding area image to obtain the image of the weld area to be processed.

3. The method for extracting weld contours based on machine vision for 3D reconstruction according to claim 2, characterized in that: The method for obtaining welding spectrum characteristic data corresponding to N welding robots includes: S100: Let the initial value of n be 1, and the value range of n is 1 to N; let the initial value of b be 1, and b is the counting variable of the optical flow high disturbance area; S101: Divide the welding spectrum image of the nth welding robot into H spectrum pixels, and calculate the corresponding optical flow disturbance degree according to the image brightness of the H spectrum pixels; S102: Marking spectral pixels whose optical flow disturbance degree is greater than a preset optical flow disturbance degree threshold as optical flow high disturbance pixels; S103: Select one optical flow high disturbance pixel from all optical flow high disturbance pixels as an optical flow high disturbance pixel seed, and based on the optical flow high disturbance pixel seed, expand to all adjacent optical flow high disturbance pixels of the same type to form the bth optical flow high disturbance area; and continue to expand to adjacent optical flow high disturbance pixels of the same type until there are no optical flow high disturbance pixels that meet the conditions to be added to the bth optical flow high disturbance area; let b=b+1, repeat S103 until all optical flow high disturbance pixels are assigned to the corresponding optical flow high disturbance areas, and finally obtain B optical flow high disturbance areas; S104: Constructing B optical flow high disturbance areas and corresponding optical flow disturbance degrees into welding spectrum feature data of the nth welding robot; S105: Let n=n+1. If n is less than or equal to N, continue executing S101 to S104; if n is greater than N, end the current process.

4. The method for extracting weld contours by three-dimensional reconstruction based on machine vision according to claim 3, characterized in that: The method for obtaining welding temperature characteristic data corresponding to N welding robots includes: S200: let the initial value of n be 1; S201: Inputting the welding process parameters of the nth welding robot into a threshold setting model to obtain a pixel temperature threshold 1 and a pixel temperature threshold 2 corresponding to the welding thermal imaging image of the nth welding robot; the pixel temperature threshold 1 is less than the pixel temperature threshold 2; The welding thermal imaging image of the nth welding robot is divided into M hot pixels; the temperatures corresponding to the M hot pixels are obtained and recorded as hot pixel temperatures; the hot pixels whose hot pixel temperatures are greater than or equal to the pixel temperature threshold 2 are marked as high thermal distortion pixels, the hot pixels whose hot pixel temperatures are less than the pixel temperature threshold 2 and greater than or equal to the pixel temperature threshold 1 are marked as medium thermal distortion pixels, and the hot pixels whose hot pixel temperatures are less than the pixel temperature threshold 1 are marked as low thermal distortion pixels; The welding thermal imaging image of the nth welding robot is divided based on high thermal distortion pixels, medium thermal distortion pixels and low thermal distortion pixels, and Q high thermal distortion areas, Y medium thermal distortion areas and R low thermal distortion areas are obtained; S202: extracting features of Q high thermal distortion regions, Y medium thermal distortion regions, and R low thermal distortion regions according to a preset method to obtain welding temperature feature data of the nth welding robot; S203: Let n=n+1. If n is less than or equal to N, continue executing S201 to S202; if n is greater than N, end the current process.

5. The method for extracting weld contours by three-dimensional reconstruction based on machine vision according to claim 4, characterized in that: The method for obtaining Q high thermal distortion regions includes: S300: Set the initial value of q to 1, where q is a high thermal distortion area counting variable; S301: Select one high thermal distortion pixel from all high thermal distortion pixels as a high thermal distortion pixel seed, and based on the high thermal distortion pixel seed, expand to all adjacent high thermal distortion pixels of the same type to form a qth high thermal distortion region; and continue to expand to adjacent high thermal distortion pixels of the same type until there are no more high thermal distortion pixels that meet the conditions to be added to the qth high thermal distortion region; then execute S302; S302: If there are high thermal distortion pixels that have not yet been assigned to corresponding high thermal distortion regions, set q = q + 1 and execute S301. If all high thermal distortion pixels have been assigned to corresponding high thermal distortion regions, stop the current process, and ultimately obtain Q high thermal distortion regions. The same method for obtaining Q high thermal distortion regions is used to obtain Y medium thermal distortion regions and R low thermal distortion regions.

6. The method for extracting weld contours by three-dimensional reconstruction based on machine vision according to claim 5, characterized in that: The method for obtaining the welding temperature characteristic data of the nth welding robot includes: Let the initial value of q be 1, and the value range of q is 1 to Q; the boundary disturbance index is calculated based on the number and size of the hot pixels in the qth high thermal distortion area of ​​the nth welding robot; let q = q + 1, when q is less than or equal to Q, continue to calculate the boundary disturbance index based on the number and size of the hot pixels in the qth high thermal distortion area of ​​the nth welding robot, and when q is greater than Q, stop the execution and obtain the boundary disturbance index corresponding to the Q high thermal distortion areas; Let the initial value of y be 1, and the value range of y be 1 to Y; calculate the stress concentration factor based on the temperature of the hot pixel in the yth medium thermal distortion area of ​​the nth welding robot; let y = y + 1, when y is less than or equal to Y, continue to calculate the stress concentration factor based on the temperature of the hot pixel in the yth medium thermal distortion area of ​​the nth welding robot. When y is greater than Y, stop the calculation and obtain the stress concentration factors corresponding to the Y medium thermal distortion areas; Let r be initialized to 1, and the value range of r is 1 to R; construct the r-th low thermal distortion region of the n-th welding robot into a minimum circumscribed ellipse, and calculate the boundary eccentricity based on the major axis length and minor axis length of the minimum circumscribed ellipse; let r = r + 1, when r is less than or equal to R, continue to construct the r-th low thermal distortion region of the n-th welding robot into a minimum circumscribed ellipse, and calculate the boundary eccentricity based on the major axis length and minor axis length of the minimum circumscribed ellipse. When r is greater than R, stop the execution and obtain the boundary eccentricities corresponding to the R low thermal distortion regions; The Q high thermal distortion areas and the corresponding boundary disturbance index, Y medium thermal distortion areas and the corresponding stress concentration factor, and R low thermal distortion areas and the corresponding boundary eccentricity of the nth welding robot are constructed into the corresponding welding temperature characteristic data.

7. The method for extracting weld contours by three-dimensional reconstruction based on machine vision according to claim 1, characterized in that: The training method of the modeling diagnosis model includes: Pre-constructing a modeling diagnosis data set, the modeling diagnosis data set including C groups of modeling diagnosis data and modeling diagnosis results corresponding to the C groups of modeling diagnosis data, where C is a positive integer greater than 0, and the modeling diagnosis data including welding spectrum feature data and welding temperature feature data; dividing the modeling diagnosis data set into a modeling diagnosis data training set and a modeling diagnosis data validation set, wherein the modeling diagnosis data training set is used for parameter learning of the modeling diagnosis model, and the modeling diagnosis data validation set is used for real-time evaluation of the generalization ability of the modeling diagnosis model; During the training process of the modeling diagnosis model, a deep neural network structure based on multi-layer perceptron is used to convert the modeling diagnosis data into a high-dimensional feature vector as input, and the nonlinear features in the data are extracted through several hidden layers. Finally, the softmax activation function is used in the output layer to generate the probability distribution of the modeling diagnosis results, and the modeling diagnosis result corresponding to the maximum probability is output as the final prediction result; the training process aims to minimize the cross-entropy loss function, and at the same time introduces an early stopping strategy to monitor the performance of the modeling diagnosis data validation set. When the prediction accuracy on the modeling diagnosis data validation set reaches the preset threshold, the modeling diagnosis model is considered to have converged and the training of the modeling diagnosis model is stopped.

8. The method for extracting weld contours by three-dimensional reconstruction based on machine vision according to claim 1, characterized in that: The training method of the parameter setting model includes: Preliminarily collecting a parameter setting data set, the parameter setting data set comprising J groups of parameter setting data and clarity adjustment parameters corresponding to the J groups of parameter setting data, where J is a positive integer greater than 0, the parameter setting data comprising welding spectrum characteristic data and welding temperature characteristic data; dividing the parameter setting data set into a training set and a validation set, the training set being used to learn parameter setting model parameters, and the validation set being used to evaluate the generalization ability of the parameter setting model; During the training process of the parameter setting model, a dynamic learning rate adjustment strategy is combined with an early stopping mechanism to minimize the cross-entropy loss function as the optimization goal. When the performance of the validation set reaches the preset performance requirements, training is automatically stopped to ensure the convergence of the parameter setting model. The parameter setting model is implemented based on the neural network model architecture. The neural network model consists of an input layer, a hidden layer, and an output layer. The input layer receives the parameter setting data and converts it into a high-dimensional feature vector. The hidden layer uses an activation function to extract the complex nonlinear pattern of the parameter setting data. The output layer calculates the probability distribution of the clarified adjustment parameter through the softmax activation function, and uses the clarified adjustment parameter corresponding to the maximum probability as the prediction result and outputs it.

9. A machine vision-based weld 3D reconstruction contour extraction system, implementing the machine vision-based weld 3D reconstruction contour extraction method according to any one of claims 1 to 8, characterized in that: include: Parameter acquisition module, used to collect welding process parameters of N welding robots; The first image acquisition module is used to acquire welding area images of N welding robots; The second image acquisition module is used to acquire welding thermal imaging images and welding spectrum images of N welding robots; A data processing module is used to extract features from the welding thermal imaging images and welding spectrum images of the N welding robots to obtain welding spectrum feature data and welding temperature feature data corresponding to the N welding robots; The weld reconstruction module reconstructs the three-dimensional contour of the weld based on the welding spectrum characteristic data, welding temperature characteristic data and welding area image of N welding robots.

Images