A deep learning method for predicting weld appearance from weld pool images

Abstract

The application discloses a kind of welding molten pool image prediction weld appearance deep learning method, it is related to welding technical field, including the following steps: first, by the two CCD cameras of upper and lower arrangement from different angles to the molten pool picture of the same position at the same time when welding is collected, molten pool picture is carried out cutting, filtering and other image processing, after using target paper calibrates camera, by three-dimensional reconstruction, obtain the three-dimensional cloud atlas of molten pool, the three-dimensional curved surface relationship z=f (x, y) of molten pool surface is fitted, using pixel instead of the horizontal and vertical coordinates of molten pool, the size of gray scale value indicates the fluctuation of molten pool surface, it is converted from three-dimensional image into two-dimensional image containing three-dimensional information, using optimized loss function to improve convolutional neural network calculation, extract and learn the molten pool information in two-dimensional molten pool image, the appearance of weld is predicted.The application solves the problem of insufficient accuracy caused by using only welding parameters to predict the appearance of weld.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of welding, in particular to a deep learning method for predicting weld appearance from welding pool image. BACKGROUND

[0002] Gas metal arc welding is one of the most widely used welding methods, because it has the advantages of low welding cost, wide application range, mechanization and automation, all-position welding, high production efficiency, etc., and has a large number of applications in shipbuilding, automobile manufacturing, pipeline welding, additive manufacturing and pressure vessels. However, with the wide application of automatic and intelligent welding in the industrial field, people's requirements for weld forming quality are getting higher and higher, and weld forming quality detection is particularly important. The traditional welding quality detection method is to detect the weld quality after welding through non-destructive detection methods such as ultrasonic, X-ray, or destructive detection methods such as slicing, and to measure the weld width, excess height and fusion ratio and other parameters by manual calculation, which cannot meet the needs of modern welding production quality management automation and digital development. Therefore, how to detect the weld quality in real time during welding becomes particularly important.

[0003] At home and abroad, the main method for predicting weld appearance is to use parameters such as welding current, voltage and welding speed that are easy to measure. The use of molten pool information closely related to weld appearance is mostly to obtain the molten width from the collected molten pool pictures, and the three-dimensional information contained in the molten pool is not used. The molten pool contains a wealth of information, such as the depression under the arc, the accumulation of tail metal and the flow of liquid metal. Skilled welders adjust welding parameters by observing the changes in the molten pool shape under the arc to control the good forming of the welding product. Therefore, detecting the molten pool shape under the arc is of great significance to the prediction of weld appearance. However, using only welding parameters to predict weld appearance will result in insufficient accuracy, so there is an urgent need for a deep learning method for predicting weld appearance from welding pool image to change this situation. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art and provide a deep learning method for predicting weld appearance from welding pool image.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0006] A deep learning method for predicting weld appearance from welding pool image, comprising the following steps:

[0007] Step one: first, the same position at the same time during welding is collected by two CCD cameras arranged up and down from different angles, the molten pool pictures are processed by image processing such as cutting and filtering, the corresponding molten pool feature point pairs in the two molten pool images are obtained, the three-dimensional coordinates corresponding to the molten pool feature points are calculated after the target paper is calibrated, the three-dimensional cloud diagram of the molten pool is obtained, the three-dimensional surface relationship z = f(x, y) of the molten pool surface and the three-dimensional reconstruction image of the molten pool are obtained by fitting;

[0008] Step two: the three-dimensional reconstruction image of the molten pool is processed into a two-dimensional image containing three-dimensional information;

[0009] Step three: a convolutional neural network is constructed to obtain the information in the image and predict the geometric morphology of the weld; the model uses an adjusted mean square error loss function (MSE), since the penetration and reinforcement of the weld are quite different from the width of the weld, the width of the weld is about 2-7 times the penetration and about 3-5 times the reinforcement, in order to improve the prediction accuracy of the network for the penetration and reinforcement of the weld, the original loss function is optimized, different coefficients α, β and γ are used according to the difference of the terms based on the original mean square error, and the proportion of the penetration and reinforcement of the weld in the loss function is increased; wherein α is 2-7 times β, and γ is 3-5 times β, so as to obtain the optimized loss function;

[0010] Step four: the training molten pool image is imported into the convolutional neural network to obtain the predicted values of the penetration, width and reinforcement of the weld, which are compared with the measured values, if the error between the predicted results and the actual measured results is less than 5%, the training is stopped and the model is saved; if the error requirement is not met, the error is calculated using the optimized MSE loss function, the error is back propagated and the convolutional neural network model parameters are optimized using the Adam optimizer, and the next training is performed;

[0011] Step five: 280 groups of data are collected, of which 240 groups are used for training the network and 40 groups are used for verification.

[0012] Preferably, the step of processing the three-dimensional reconstruction image of the molten pool into a two-dimensional image is to determine the size of the two-dimensional molten pool image, obtain the pixel value of the two-dimensional molten pool image, obtain the penetration, width and reinforcement of the weld, and assemble the data set.

[0013] Preferably, when determining the size of the 2D molten pool image, it is important to ensure that the size information of the molten pool remains unchanged before and after the 2D conversion. That is, the true distance represented by two pixels along the welding direction after conversion is the same as the true distance represented by two pixels perpendicular to the welding direction. The size of the 2D molten pool image after conversion facilitates convolution operations during subsequent model training. It is necessary to ensure that the number of pixel values ​​for the length and width of the 2D molten pool image is the same, that is, the side length of the 2D molten pool image is the same. The length and width of all reconstructed 3D molten pools are calculated and statistically analyzed. The molten pool width is between (5, 15) mm, and the molten pool length is between (5, 14) mm. To ensure that the information of the molten pool is completely contained in the 2D molten pool image, while reserving a certain range outside the molten pool to represent the surface of the welding base material, the side length of the 2D molten pool image is... The actual length represented by the number of pixels should be greater than the maximum value of the molten pool length and molten pool width. Therefore, the actual length represented by the side length of the two-dimensional molten pool image is between (16, 20) mm. The three-dimensional molten pool is divided using the three-dimensional molten pool surface partitioning formula. First, the center coordinates (x0, y0) of the molten pool are obtained. Then, a suitable actual length value b represented by the side length of the two-dimensional molten pool image is selected to obtain the x-coordinate range (x0-b / 2, x0+b / 2) and y-coordinate range (y0-b / 2, y0+b / 2) represented by the transformed two-dimensional molten pool image. These are then interpolated and divided into 200 equal parts to obtain a 200*200 x, y coordinate point matrix (x′, y′). Substituting (x′, y′) into the surface formula obtained in step one, the z′ value of the corresponding point (x′, y′) is obtained.

[0014] Preferably, the formula for dividing the three-dimensional molten pool surface is as follows: Where x is the abscissa of the molten pool in the 3D reconstructed image; y is the ordinate of the molten pool in the 3D reconstructed image; x0 is the center of the abscissa of the molten pool in the 3D reconstructed image; y0 is the center of the ordinate of the molten pool in the 3D reconstructed image; (x′, y′) are the interpolated x and y axis coordinates; and b is the actual length represented by the side length of the 2D molten pool image.

[0015] Preferably, when acquiring the pixel values ​​of the two-dimensional molten pool image, it should be noted that in the three-dimensional coordinates of the molten pool, the surface of the welding base material is a plane with z = 0, z > 0 represents the portion of the molten pool surface above the base material surface, and z < 0 represents the portion of the molten pool below the base material surface. The surface span of the three-dimensional reconstructed molten pool is calculated and statistically analyzed. The distance from the highest point of the molten pool surface to the base material surface is statistically between (0.1, 1) mm, and the distance from the lowest point of the molten pool bottom to the base material surface is between (0.5, 3) mm. To completely include the molten pool, the distances from the highest and lowest points of the molten pool to the base material surface should be... For values ​​greater than 3mm, to preserve the fluctuation of the molten pool surface, the value should not be too large. Z should be between (3,5). The converted two-dimensional molten pool image is a grayscale image with a grayscale value range of [0,255]. Therefore, the median value of 128, representing the grayscale range of the base material surface, is taken. The formula for calculating the grayscale value of the points on the other molten pool surface is gray=(z′ / Z+1)×128, where gray is the converted grayscale value; z′ is the z-axis coordinate of the molten pool surface obtained after x and y interpolation in the previous step; and Z is the actual distance to the base material surface represented by the maximum grayscale value of 255.

[0016] Preferably, when obtaining the weld penetration, weld width, and weld reinforcement height, attention should be paid to cutting the weld location of the acquired molten pool image to obtain a cross section, and then measuring the weld penetration, weld width, and weld reinforcement height through metallographic preparation.

[0017] Preferably, the original loss function is The optimized loss function is: Where p and p′ represent the predicted and measured values ​​of weld penetration depth, w and w′ represent the predicted and measured values ​​of weld width, and h and h′ represent the predicted and measured values ​​of weld reinforcement height, respectively; α, β and γ are the coefficients of the variance between the predicted and measured values ​​of weld penetration depth, weld width and reinforcement height, respectively, where α is 2 to 7 times β and γ is 3 to 5 times β.

[0018] The beneficial effects of this invention are as follows:

[0019] This deep learning method for predicting weld morphology from weld pool images divides the entire prediction process into four steps: weld pool image acquisition, camera calibration, 3D weld pool information acquisition, training sample construction, algorithm training, and testing. It employs a binocular vision system to reconstruct the 3D weld pool. After obtaining the 3D morphology of the weld pool, it uses a convolutional neural network to extract its 3D information and predict the weld morphology, thus solving the problem of insufficient accuracy caused by using only welding parameters to predict weld morphology. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the learning process structure of a deep learning method for predicting weld morphology from weld pool images proposed in this invention.

[0021] Figure 2 This is a flowchart of the molten pool image processing method for predicting weld morphology from molten pool images, as proposed in this invention.

[0022] Figure 3 This is a flowchart of the three-dimensional reconstruction of the weld pool and the two-dimensionalization process of the deep learning method for predicting weld morphology from weld pool images proposed in this invention.

[0023] Figure 4 This is a flowchart of the convolutional neural network training process for a deep learning method for predicting weld morphology from weld pool images, as proposed in this invention.

[0024] Figure 5 This is a schematic diagram of the error curve structure of a deep learning method for predicting weld morphology from weld pool images proposed in this invention.

[0025] Figure 6 This is a comparison chart of the predicted and measured weld penetration values ​​in the training set of the deep learning method for predicting weld morphology from weld pool images proposed in this invention.

[0026] Figure 7 This is a comparison chart of the predicted and measured weld width values ​​in the training set of the deep learning method for predicting weld morphology from weld pool images proposed in this invention.

[0027] Figure 8 This is a comparison image of the predicted and measured weld reinforcement height values ​​from the training set of the deep learning method for predicting weld morphology from weld pool images proposed in this invention.

[0028] Figure 9 This is a comparison chart of the predicted and measured weld penetration values ​​for the test set of the deep learning method for predicting weld morphology from weld pool images proposed in this invention.

[0029] Figure 10 This is a comparison chart of the predicted and measured weld width values ​​for the test set of the deep learning method for predicting weld morphology from weld pool images proposed in this invention.

[0030] Figure 11 This is a comparison chart of the predicted and measured values ​​of weld reinforcement height in the test set of a deep learning method for predicting weld morphology from weld pool images proposed in this invention. Detailed Implementation

[0031] The technical solution of this patent will be further described in detail below with reference to specific embodiments.

[0032] The embodiments of this patent are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this patent, and should not be construed as limiting this patent.

[0033] In the description of this patent, it should be understood that the terms “center,” “upper,” “lower,” “front,” “back,” “left,” “right,” “vertical,” “horizontal,” “top,” “bottom,” “inner,” and “outer,” etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this patent and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this patent.

[0034] In the description of this patent, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "setting" should be interpreted broadly. For example, they can refer to a fixed connection or setting, a detachable connection or setting, or an integral connection or setting. Those skilled in the art can understand the specific meaning of the above terms in this patent according to the specific circumstances.

[0035] Reference Figures 1-4 A deep learning method for predicting weld morphology from weld pool images includes the following steps:

[0036] Step 1: First, two CCD cameras, positioned vertically and vertically, acquire images of the molten pool at the same time and location during welding from different angles. The images are then processed by cropping and filtering to obtain corresponding molten pool feature point pairs in the two images. After calibrating the cameras using target paper, the three-dimensional coordinates corresponding to the molten pool feature points are calculated to obtain the three-dimensional cloud map of the molten pool. The three-dimensional surface relationship z = f(x, y) of the molten pool surface and the three-dimensional reconstructed image of the molten pool are obtained by fitting the three-dimensional cloud map of the molten pool.

[0037] Step 2: Convert the 3D reconstructed image of the molten pool into a 2D image containing 3D information.

[0038] Step 3: Construct a convolutional neural network to acquire information from the image and predict the geometric shape of the weld using the convolution formula. The model uses an adjusted mean squared error (MSE) loss function. Since the weld penetration and height differ significantly from the weld width (approximately 2-7 times the penetration and 3-5 times the height), the original loss function is optimized to improve the network's prediction accuracy. Different coefficients α, β, and γ are used based on the original mean squared error to increase the proportion of weld penetration and height in the loss function. Here, α is 2-7 times β, and γ is 3-5 times β, thus obtaining the optimized loss function.

[0039] Step 4: After importing the training molten pool image into the convolutional neural network, the predicted values ​​of weld penetration, weld width, and weld height are obtained. These values ​​are then compared with the measured values. If the error between the predicted and measured results is less than 5%, training is stopped and the model is saved. If the error requirement is not met, the error is calculated using the optimized MSE loss function, the error is backpropagated, and the Adam optimizer is used to optimize the parameters of the convolutional neural network model before the next training iteration.

[0040] Step 5: 280 sets of data were collected, of which 240 sets were used for training the network and 40 sets were used for validating the trained network.

[0041] In this embodiment, the Adam optimizer optimizes the model parameters by comprehensively considering the first moment estimate (mean of the gradient) and the second moment estimate (uncentered variance of the gradient) of the gradient and automatically selecting the update step size.

[0042] In this embodiment, the steps for converting the 3D reconstructed image of the molten pool into a 2D image are as follows: determining the size of the 2D image of the molten pool, obtaining the pixel values ​​of the 2D molten pool image, obtaining the weld depth, weld width, and weld height, and constructing a dataset.

[0043] In this embodiment, when determining the size of the 2D image of the molten pool, it is important to ensure that the size information of the molten pool remains unchanged before and after the 2D conversion. That is, the true distance represented by two pixels along the welding direction after conversion is the same as the true distance represented by two pixels perpendicular to the welding direction. The size of the 2D molten pool image after conversion is convenient for convolution operations during subsequent model training. It is necessary to ensure that the number of pixel values ​​in the length and width of the 2D molten pool image is the same, that is, the side length of the 2D molten pool image is the same. The length and width of all reconstructed 3D molten pools are calculated and statistically analyzed. The width of the molten pool is between (5, 15) mm, and the length of the molten pool is between (5, 14) mm. To ensure that the information of the molten pool is completely contained in the 2D molten pool image, and at the same time, a certain range is reserved outside the molten pool to represent the surface of the welding base material, the side length of the 2D molten pool image is... The actual length represented by the number of pixels should be greater than the maximum value of the molten pool length and molten pool width. Therefore, the actual length represented by the side length of the two-dimensional molten pool image is between (16, 20) mm. The three-dimensional molten pool is divided using the three-dimensional molten pool surface partitioning formula. First, the center coordinates (x0, y0) of the molten pool are obtained. Then, a suitable actual length value b represented by the side length of the two-dimensional molten pool image is selected to obtain the x-coordinate range (x0-b / 2, x0+b / 2) and y-coordinate range (y0-b / 2, y0+b / 2) represented by the transformed two-dimensional molten pool image. These are then interpolated and divided into 200 equal parts to obtain a 200*200 x, y coordinate point matrix (x′, y′). Substituting (x′, y′) into the surface formula obtained in step one, the z′ value of the corresponding point (x′, y′) is obtained.

[0044] In this embodiment, the formula for dividing the three-dimensional molten pool surface is as follows: Where z is the abscissa of the molten pool in the 3D reconstructed image; y is the ordinate of the molten pool in the 3D reconstructed image; x0 is the center of the abscissa of the molten pool in the 3D reconstructed image; y0 is the center of the ordinate of the molten pool in the 3D reconstructed image; (x′, y′) are the interpolated x and y axis coordinates; and b is the actual length represented by the side length of the 2D molten pool image.

[0045] In this embodiment, when obtaining the pixel values ​​of the two-dimensional molten pool image, it should be noted that the surface of the base material is a plane with z = 0. z > 0 indicates that the molten pool surface is higher than the base material surface, and z < 0 indicates that the molten pool is lower than the weld surface. The surface span of the molten pool in the three-dimensional reconstruction is calculated and statistically analyzed. The distance from the highest point of the molten pool surface to the base material surface is statistically between (0.1, 1) mm, and the distance from the lowest point of the molten pool bottom to the base material surface is between (0.5, 3) mm. To completely include the molten pool, the distance from the highest and lowest points of the molten pool to the base material surface should be greater than 3 mm. To preserve the fluctuations on the surface of the molten pool, the value should not be too large. Z should be between (3,5). The converted two-dimensional molten pool image is a grayscale image with a grayscale value range of [0,255]. Therefore, the median value of 128, representing the grayscale range of the base material surface, is taken. The formula for calculating the grayscale value of the points on the other molten pool surface is gray=(z′ / Z+1)×128, where gray is the converted grayscale value; z′ is the z-axis coordinate of the molten pool surface obtained after x and y interpolation in the previous step; and Z is the actual distance to the base material surface represented by the maximum grayscale value of 255.

[0046] In this embodiment, to obtain the weld penetration, weld width, and weld reinforcement height, attention should be paid to cutting the weld location of the acquired molten pool image to obtain a cross section, and then measuring the weld penetration, weld width, and weld reinforcement height through metallographic preparation.

[0047] In this embodiment, when constructing the dataset, attention should be paid to collecting images of the molten pool and obtaining the cross-sectional dimensions of the weld seam during the welding process under different welding parameters, and repeating the above steps to construct training samples.

[0048] In this embodiment, the convolution formula is: Where o is the size of the output image; i is the size of the input image; f and s are the size of the convolution kernel and the stride, respectively; and p is the size of the padding boundary.

[0049] In this embodiment, the original loss function is The optimized loss function is, Where p and p′ represent the predicted and measured values ​​of weld penetration depth, w and w′ represent the predicted and measured values ​​of weld width, h and h′ represent the predicted and measured values ​​of weld reinforcement height, respectively; α, β and γ are the coefficients of the variances of the predicted and measured values ​​of weld penetration depth, weld width and reinforcement height.

[0050] like Figure 5 The figure shows the curve of the loss function and the training batch during neural network training. As shown in the figure, after 150 steps, the training mean square error basically stabilizes.

[0051] like Figure 6 , 78 and 8 represent the predicted and measured values ​​of weld penetration, weld width, and weld reinforcement in the training set, respectively. The training set has good accuracy, with the predicted values ​​being more than 95% of the measured values.

[0052] like Figure 9 , 10 11 and 11 are the predicted and measured values ​​of weld penetration, weld width and reinforcement height in the test set, respectively. The accuracy of the test set decreased slightly, but still remained above 90%.

[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning method for predicting weld morphology from weld pool images, characterized in that, Includes the following steps: Step 1: First, two CCD cameras, positioned vertically and vertically, acquire images of the molten pool at the same time and location during welding from different angles. The images are then processed by cropping and filtering to obtain corresponding molten pool feature point pairs in the two images. After calibrating the cameras using target paper, the three-dimensional coordinates corresponding to the molten pool feature points are calculated to obtain the three-dimensional cloud map of the molten pool. The three-dimensional surface relationship z = f(x, y) of the molten pool surface and the three-dimensional reconstructed image of the molten pool are obtained by fitting the three-dimensional cloud map of the molten pool. Step 2: Convert the 3D reconstructed image of the molten pool into a 2D image containing 3D information. Step 3: Construct a convolutional neural network to acquire information from the image and predict the geometric shape of the weld. The model uses an adjusted mean squared error (MSE) loss function. Since the weld penetration and height differ significantly from the weld width (approximately 2-7 times the penetration and 3-5 times the height), the original loss function is optimized to improve the network's prediction accuracy. Different coefficients α, β, and γ are used based on the original mean squared error to increase the proportion of weld penetration and height in the loss function. Here, α is 2-7 times β, and γ is 3-5 times β, thus obtaining the optimized loss function. Step 4: After importing the training molten pool image into the convolutional neural network, the predicted values ​​of weld penetration, weld width, and weld height are obtained. These values ​​are then compared with the measured values. If the error between the predicted and measured results is less than 5%, training is stopped and the model is saved. If the error requirement is not met, the error is calculated using the optimized MSE loss function, the error is backpropagated, and the Adam optimizer is used to optimize the parameters of the convolutional neural network model before the next training iteration. Step 5: 280 sets of data were collected, of which 240 sets were used for network training and 40 sets were used for network validation.

2. The deep learning method for predicting weld morphology from weld pool images according to claim 1, characterized in that, The steps for converting the 3D reconstructed image of the molten pool into a 2D image are as follows: determining the size of the 2D image of the molten pool, obtaining the pixel values ​​of the 2D molten pool image, obtaining the weld depth, weld width, and weld height, and constructing a dataset.

3. The deep learning method for predicting weld morphology from weld pool images according to claim 2, characterized in that, When determining the size of the 2D molten pool image, it is important to ensure that the size information of the molten pool remains unchanged before and after the 2D conversion. That is, the true distance represented by two pixels along the welding direction after conversion is the same as the true distance represented by two pixels perpendicular to the welding direction. The size of the 2D molten pool image after conversion facilitates convolution operations during subsequent model training. It is necessary to ensure that the number of pixels in the length and width of the 2D molten pool image is the same, i.e., the side length of the 2D molten pool image is the same. The length and width of all reconstructed 3D molten pools are calculated and statistically analyzed. The molten pool width is between (5, 15) mm, and the molten pool length is between (5, 14) mm. To ensure that the information of the molten pool is completely contained in the 2D molten pool image, while reserving a certain range outside the molten pool to represent the surface of the base material, the side length and number of pixels of the 2D molten pool image are determined. The actual length represented by the quantity should be greater than the maximum value of the molten pool length and molten pool width. Therefore, the actual length represented by the side length of the two-dimensional molten pool image is between (16, 20) mm. The three-dimensional molten pool is divided using the three-dimensional molten pool surface division formula. First, the center coordinates (x0, y0) of the molten pool are obtained. Then, a suitable actual length value b represented by the side length of the two-dimensional molten pool image is selected to obtain the x-coordinate range (x0-b / 2, x0+b / 2) and y-coordinate range (y0-b / 2, y0+b / 2) represented by the transformed two-dimensional molten pool image. These are then interpolated and divided into 200 equal parts to obtain a 200*200 x, y coordinate matrix (x′, y′). Substituting (x′, y′) into the surface formula obtained in step one, the z′ value of the corresponding point (x′, y′) is obtained.

4. The deep learning method for predicting weld morphology from weld pool images according to claim 3, characterized in that, The formula for dividing the three-dimensional molten pool surface is as follows: Where x is the abscissa of the molten pool in the 3D reconstructed image; y is the ordinate of the molten pool in the 3D reconstructed image; x0 is the center of the abscissa of the molten pool in the 3D reconstructed image; y0 is the center of the ordinate of the molten pool in the 3D reconstructed image; (x′, y′) are the interpolated x and y axis coordinates; and b is the actual length represented by the side length of the 2D molten pool image.

5. The deep learning method for predicting weld morphology from weld pool images according to claim 4, characterized in that, When acquiring the pixel values ​​of the two-dimensional molten pool image, it is important to note that the surface of the base material is a plane with z = 0. z > 0 represents the portion of the molten pool surface above the base material surface, and z < 0 represents the portion of the molten pool below the base material surface. The surface span of the 3D reconstructed molten pool is calculated and statistically analyzed. The distance from the highest point of the molten pool surface to the base material surface is statistically between (0.1, 1) mm, and the distance from the lowest point of the molten pool bottom to the base material surface is between (0.5, 3) mm. To completely include the molten pool, the distance from the highest and lowest points of the molten pool to the base material surface should be greater than 3 mm. This is to preserve the molten pool's shape and shape. The fluctuation of the pool surface should not be too large. Z should be between (3,5). The converted two-dimensional molten pool image is a grayscale image with a grayscale value range of [0,255]. Therefore, the median value of 128, which represents the grayscale range of the base material surface, is taken. The formula for calculating the grayscale value of the points on the other molten pool surface is gray=(z′ / Z+1)×128, where gray is the converted grayscale value; z′ is the z-axis coordinate of the molten pool surface obtained after x and y interpolation in the previous step; and Z is the actual distance to the base material surface represented by the maximum grayscale value of 255.

6. The deep learning method for predicting weld morphology from weld pool images according to claim 5, characterized in that, To obtain the weld penetration, width, and height, it is important to cut the weld area in the acquired molten pool image to obtain a cross-section, and then perform metallographic preparation to measure the weld penetration, width, and height.

7. The deep learning method for predicting weld morphology from weld pool images according to claim 1, characterized in that, The original loss function is: The optimized loss function is: Where p and p′ represent the predicted and measured values ​​of weld penetration depth, w and w′ represent the predicted and measured values ​​of weld width, and h and h′ represent the predicted and measured values ​​of weld reinforcement height, respectively; α, β and γ are the coefficients of the variance between the predicted and measured values ​​of weld penetration depth, weld width and reinforcement height, respectively, where α is 2 to 7 times β and γ is 3 to 5 times β.

Images