A method and system for predicting the cross-sectional size of laser cladding on an inclined base surface

Abstract

The present application relates to the technical field of laser cladding, and particularly relates to a laser cladding cross-section size prediction method and system for an inclined base surface. In view of the problems of boundary fracture, holes and noise in the prediction results of a deep learning model, an endpoint extension repair algorithm based on graph structure analysis is proposed, and through morphological operation, connected domain analysis and path reverse extension and other steps, the boundary defects are effectively repaired. In order to improve the prediction accuracy of the cladding layer cross section, the Pix2Pix model is applied to the cladding layer cross section shape prediction for the first time, and the perception loss is introduced to enhance the perception ability of the model to the boundary details and global structure. In addition, the structural similarity index is introduced, and the high-similarity molten pool images are extracted from the molten pool videos under different parameter conditions as an expanded training set, so that the generalization ability and robustness of the model are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser cladding, and particularly relates to a laser cladding cross-section size prediction method and system for an inclined base surface. BACKGROUND

[0002] As an important additive manufacturing method in the field of advanced manufacturing, laser cladding technology has been widely used in the fields of aerospace, automobile manufacturing, mold repair and mechanical manufacturing in recent years, because it can realize high-precision and high-quality metal surface repair and functional coating preparation. The technology melts and deposits metal powder or wire on the surface of the substrate through a high-energy laser beam to form a cladding layer that is metallurgically combined with the substrate, thereby significantly improving the wear resistance, corrosion resistance and service life of the workpiece.

[0003] In the laser cladding process, the cladding layer cross-section size (such as width, height, area, etc.) is an important parameter for measuring the cladding quality, which is directly related to the mechanical properties of the cladding layer and the subsequent processing quality. The traditional cladding size prediction method mainly relies on empirical formula or numerical simulation technology based on physical mechanism, mainly using numerical simulation methods such as finite element analysis (FEA) and computational fluid dynamics (CFD) to model the heat conduction, molten pool flow, solidification shrinkage and other physical processes in the laser cladding process. These methods can accurately predict the cladding layer cross-section size in theory, but have the disadvantages of long calculation time, high calculation complexity, strong parameter dependence, high cost of parameter acquisition, poor adaptability of the model to changes in working conditions, etc.

[0004] With the rapid development of artificial intelligence, especially deep learning, traditional convolutional neural network (CNN), Transformer and generative adversarial network (GAN) have been introduced into the field of cladding cross-section prediction, achieving high prediction accuracy. For example, some studies use a hybrid CNN-Transformer architecture, using ResNet for spatial feature extraction and Transformer for time series modeling, effectively capturing the spatio-temporal features of the heat image sequence of the molten pool surface through the self-attention mechanism, thereby accurately predicting the two-dimensional depth profile of the molten pool. Some studies combine an improved PredNet for long-term prediction of molten pool images (up to 140 ms in advance) and use ResNet-34 with SE channel attention module (SERes network) for regression analysis, using molten pool image sequences and welding speed inputs to accurately predict the reinforcement height of the weld. Some studies use a GRU-GAN model that combines Gated Recurrent Unit (GRU) and Conditional Generative Adversarial Nets (CGAN), using GRU to model time series dependencies and generating high-fidelity future molten pool images through CGAN, combined with welding speed sequences, to dynamically predict and adjust the molten pool morphology, assisting human operators in making Model Predictive Control (MPC) decisions based on predictions.

[0005] Although the above deep learning models can rely on data-driven advantages to achieve certain accuracy in the prediction of cladding cross-section size, they are prone to defects such as boundary fracture, incomplete closure and noise regions when directly outputting predicted cross-section images. This is because most deep learning models are prone to feature information loss in the transition and mutation regions of the cross-section boundary when extracting local features, leading to boundary generation fracture. Some deep learning models can model global associations, but the distinction between background and boundary features in molten pool images is low, and the model is prone to misjudgment of irrelevant background information as boundary features or omission of some weak boundary signals, resulting in incomplete closure of the boundary and a large difference between the predicted results and the true cross-section in terms of size accuracy and morphology consistency, making it difficult to meet the requirements of industrial applications for precision and stability in predicting the size of laser cladding cross-sections on inclined base surfaces. SUMMARY

[0006] To this end, the technical problem to be solved by the present application is to overcome the problem that the existing deep learning model-based cladding cross-section size prediction method is prone to boundary fracture, incomplete closure and noise area when directly outputting a prediction image, resulting in poor cladding cross-section size prediction accuracy.

[0007] To solve the above technical problems, the present application provides a laser cladding cross-section size prediction method based on inclined base surface, comprising:

[0008] The molten pool image to be predicted is input into the cladding cross-section size prediction model to obtain an initial cladding layer boundary image;

[0009] The areas of all connected regions in the initial cladding layer boundary image are sorted, the largest connected region is taken as the main boundary region, and the remaining regions are taken as non-main regions, only the pixels of the main boundary region are retained, and a processed cladding layer boundary image is obtained;

[0010] For each end point pixel in the processed cladding layer boundary image, the end point pixel is extended in the opposite direction of the adjacent pixel to which the end point pixel belongs to the main boundary region, until the extension path of the end point pixel coincides with the pixel of the main boundary region or reaches the boundary of the processed cladding layer boundary image;

[0011] If the extension path of the end point pixel reaches the boundary of the processed cladding layer boundary image, the boundary of the cladding layer boundary image is deleted pixel by pixel in the opposite direction of the extension direction of the end point pixel until the intersection point of the extension path of the end point pixel and the main boundary region is reached, and a target cladding layer boundary image is obtained;

[0012] Based on the target cladding layer boundary image, the size parameters of the molten pool image to be predicted are obtained.

[0013] Preferably, it further comprises:

[0014] The initial cladding layer boundary image is subjected to morphological opening operation and closing operation to obtain a target initial cladding layer boundary image.

[0015] Preferably, it further comprises:

[0016] The distance between each non-main region and the geometric center of the initial main boundary region is calculated, and the non-main regions with a distance less than a set threshold value are merged with the main boundary region as a target main boundary region.

[0017] Preferably, the intersection point determination method comprises:

[0018] It is judged whether there are three adjacent pixels belonging to the main boundary region in the 8-neighborhood of the current pixel, if there are, it is judged whether the extension paths formed by the current pixel and each of its adjacent pixels are all not coincident, if they are all not coincident, the current pixel is an intersection point.

[0019] Preferably, the prediction model for the cladding cross-section size is the Pix2Pix model.

[0020] Preferably, the training process of the Pix2Pix model includes:

[0021] Obtain the training set of molten pool images, input the molten pool images into the generator of the Pix2Pix model, and output the initial cladding layer boundary map of the molten pool images;

[0022] The initial cladding layer boundary map of the molten pool image is processed through a pre-trained feature extraction network to extract predicted features from multiple levels of the feature extraction network.

[0023] The labeled image of the molten pool image is passed through a pre-trained feature extraction network to extract real features from multiple levels of the feature extraction network;

[0024] The first confidence level is obtained by passing the molten pool image and its initial cladding layer boundary map through the discriminator of the Pix2Pix model;

[0025] The second confidence level is obtained by passing the molten pool image and its label image through the discriminator of the Pix2Pix model;

[0026] The Pix2Pix model is trained based on the conditional adversarial loss between the first and second confidence levels, the pixel-level L1 loss between the initial cladding layer boundary map of the molten pool image and the label image, and the perceptual loss between the predicted features and the real features of multiple layers of the feature extraction network to obtain the target Pix2Pix model.

[0027] Preferably, the formula for the perceptual loss between the predicted features and the true features at multiple levels of the feature extraction network is:

[0028] ,

[0029] in, In order to perceive loss, This represents the number of layers in the feature extraction network. It is a molten pool image. It is a label image. It is random noise. This represents the initial cladding layer boundary diagram. Represents a generator. For the feature extraction network Predictive features of the layer For the feature extraction network The true characteristics of the layer For the feature extraction network The height of the layer's features For the feature extraction network The width of the layer's features, For the feature extraction network The number of channels in the layer's features. For layer index, It is an L2 norm.

[0030] Preferably, the process of obtaining the molten pool image training set includes:

[0031] Acquire molten pool videos under different parameter conditions, select any frame of molten pool image from the molten pool videos under different parameter conditions, and the cross-sectional image of the cladding layer under that parameter condition as label images, and use them as samples in the molten pool image training set.

[0032] Data augmentation was performed on the molten pool images selected under different parameter conditions to obtain data-augmented molten pool images under different parameter conditions;

[0033] In the molten pool video under different parameter conditions, the molten pool images other than the selected molten pool images are used as candidate images under that parameter condition;

[0034] Calculate the structural similarity index between each candidate image and the data-enhanced cladding pool image under different parameter conditions, sort them in descending order, and use the candidate images of the previous preset number of frames and the cross-sectional images of the cladding layer under the same parameter conditions as label images, which are then used as augmented samples in the cladding pool image training set.

[0035] Preferably, the size parameters of the molten pool image to be predicted include: the cross-sectional area, width, and height of the cladding layer.

[0036] The present invention also provides a system for predicting the cross-sectional dimensions of laser cladding on an inclined datum, comprising:

[0037] A memory for storing computer programs; a processor for executing the computer programs to implement the steps of the above-described method for predicting the cross-sectional dimensions of laser cladding on an inclined datum.

[0038] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0039] The method and system for predicting the cross-sectional size of laser cladding on an inclined base surface can directly filter out small-area non-main regions, such as isolated noise pixels or fine pseudo-boundaries, generated due to noise interference during model prediction, by determining the largest connected region in the initial cladding layer boundary map as the main boundary region and retaining the pixels thereof, thereby reducing the interference of irrelevant regions on subsequent size calculation from the source, and can effectively repair the boundary fracture problem caused by insufficient feature learning during model prediction by extending the endpoint pixel in the processed cladding layer boundary map in the opposite direction of the adjacent pixel to which the endpoint pixel belongs to the main boundary region until the extension path of the endpoint pixel coincides with the pixels of the main boundary region or reaches the boundary of the processed cladding layer boundary map, so that the broken boundary is reconnected to form a continuous structure. Finally, the endpoint whose extension path reaches the image boundary is deleted from the boundary in reverse and pixel by pixel until the intersection point of the extension path of the endpoint pixel and the main boundary region is reached, which can clean up the invalid extension pixels at the edge of the boundary map, further ensure the closed integrity and morphological accuracy of the main boundary, and ultimately calculate the size parameters based on the target cladding layer boundary map that is regular, continuous and free of noise interference, thereby effectively improving the accuracy and reliability of the prediction of the cross-sectional size of laser cladding on an inclined base surface.

[0040] Like traditional convolutional neural networks and Transformer deep learning models, when facing typical cross-domain image conversion needs such as molten pool images and cladding layer cross-sectional images, they cannot accurately capture the mapping relationship between them. In order to solve the above problems, the present application introduces the Pix2Pix model, which is used as a cladding cross-sectional size prediction model. This model is inherently capable of handling the conversion of input and output images from different domains, and can gradually learn the potential mapping rules between cross-domain images through the adversarial training of the generator and the discriminator. To further improve the conversion accuracy, the present application additionally introduces a perception loss function to capture the deep semantic features of the image using a pre-trained feature extraction network, so that the Pix2Pix model not only focuses on the surface differences at the pixel level, but also understands the deep feature correlation between the two types of images under the process logic, effectively enhancing the model's understanding ability and generation accuracy of the cross-domain feature mapping between molten pool images and cladding layer cross-sectional images, thereby improving the prediction accuracy of the cross-sectional size of laser cladding on an inclined base surface.

[0041] In addition, in order to solve the problems of high data acquisition cost, long acquisition cycle and complicated cross-section annotation process of the existing inclined base surface cladding data, and the limited scale of the high-quality data set available for training, the present application selects an arbitrary molten pool image and the corresponding real cladding layer cross-section image of the molten pool video under different parameter conditions as the initial sample, generates the data enhanced molten pool image under the parameter condition by combining data enhancement, sets the remaining molten pool images in the molten pool video under the same parameter condition except the initial frame as the candidate images, calculates the structural similarity index of each candidate image and the data enhanced molten pool image under the parameter condition, selects the images and the corresponding cross-section images of the first preset number of frames as the expanded samples, expands the feature coverage range of the single initial image through data enhancement, and ensures the consistency of the candidate images and the initial samples in the molten pool core features by means of the structural similarity index, so that the cross-section image of the expanded sample does not need to be collected or annotated, the data acquisition and annotation cost is greatly reduced, the sample number and diversity of the molten pool image training set are enriched, and the prediction accuracy of the cladding cross-section size prediction model is effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to make the content of the present application more easily and clearly understood, the present application will be further described in detail below according to specific embodiments of the present application and in conjunction with the drawings, in which:

[0043] Figure 1 is a flowchart of a laser cladding cross-section size prediction method of the present application.

[0044] Figure 2 is a flowchart of the endpoint extension repair based on the graph structure, Figure 2 (a) in is the acquisition process of the processed cladding layer boundary graph, Figure 2 (b) in is the acquisition process of the target cladding layer boundary graph.

[0045] Figure 3 is a structural diagram of the cladding cross-section size prediction model of the present application.

[0046] Figure 4 is a schematic diagram of the acquisition process of the molten pool image training set.

[0047] Figure 5 is a comparison experiment diagram of different cladding cross-section size prediction models and the method of the present application, Figure 5 (a) in is a comparison diagram of different cladding cross-section size prediction models and the method of the present application for extracting the target cladding layer boundary graph of the first molten pool image, Figure 5 (b) in is a comparison diagram of different cladding cross-section size prediction models and the method of the present application for extracting the target cladding layer boundary graph of the second molten pool image, Figure 5The (c) behavior of different cladding cross-section size prediction models in the figure and the comparison schematic of the target cladding layer boundary map of the fourth image of the molten pool extracted by the method of the present application. Figure 5 The (d) behavior of different cladding cross-section size prediction models in the figure and the comparison schematic of the target cladding layer boundary map of the fourth image of the molten pool extracted by the method of the present application.

[0048] Figure 6 The (c) behavior of different cladding cross-section size prediction models in the figure and the comparison schematic of the target cladding layer boundary map of the fourth image of the molten pool extracted by the method of the present application. Figure 6 The (a) behavior of different cladding cross-section size prediction models in the figure and the comparison schematic of the predicted cladding layer cross-sectional area by the method of the present application. Figure 6 The (b) behavior of different cladding cross-section size prediction models in the figure and the comparison schematic of the predicted width by the method of the present application. Figure 6 The (c) behavior of different cladding cross-section size prediction models in the figure and the comparison schematic of the predicted height by the method of the present application. DETAILED DESCRIPTION

[0049] The present application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it, but the embodiments are not limiting to the present application.

[0050] When the existing cladding cross-section size prediction method based on deep learning model directly outputs the predicted image, the problems of boundary fracture, incomplete closure and noise region are prone to occur. In the face of this technical problem, the skilled person in the art usually starts from the optimization of the model itself, such as increasing the number of training samples, adjusting the model network structure (such as deepening the network layer, introducing attention mechanism), optimizing the loss function (such as increasing the loss weight of the boundary region), etc., trying to make the model learn the boundary features better in the training stage to reduce the prediction defects; or using simple image post-processing methods, such as deleting obvious noise pixels by fixed threshold filtering, using single morphological operation (such as closing operation) to fill small gaps.

[0051] But these conventional ideas have obvious defects. Increasing samples or adjusting model structure not only increases data acquisition and calculation cost, but also is difficult to fundamentally solve the problem of insufficient fitting of boundary details by the model in the small sample scene, and is still prone to fracture in complex contour area; optimizing the loss function can enhance the attention of the model to the boundary, but cannot avoid the interference of noise features in the training process, and there will still be residual fine false boundaries in the prediction result; and simple post-processing methods lack pertinence, fixed threshold filtering may delete effective boundary pixels, single morphological operation is difficult to simultaneously consider noise removal and boundary repair, and cannot solve the problems of boundary closure and invalid edge extension.

[0052] Therefore, the existing method is difficult to completely solve the above-mentioned defects, so that the prediction result is different from the true cross section in size accuracy and topography consistency, and it is difficult to meet the requirements of industrial application on precision and stability. The present application is aimed at this technical blank, and proposes an endpoint extension repair method based on graph structure through progressive post-processing of main boundary region screening, endpoint reverse extension, intersection point judgment and edge cleaning, which accurately overcomes the difficulties of existing methods in noise discrimination, boundary repair integrity and invalid edge processing, effectively solves the core problem of cladding cross section boundary prediction, and the specific scheme is as follows:

[0053] Referring to Figure 1 The embodiment one provides a laser cladding cross section size prediction method of inclined base surface, which comprises:

[0054] As Figure 2 shown, Figure 2 is a flow chart of endpoint extension repair based on graph structure, Figure 2 (a) is the acquisition process of the processed cladding layer boundary graph, Figure 2 (b) is the acquisition process of the target cladding layer boundary graph.

[0055] Step S1: obtaining an initial cladding layer boundary graph by inputting the molten pool image to be predicted into the cladding cross section size prediction model;

[0056] In the embodiment, preferably, the initial cladding layer boundary graph is subjected to morphological opening operation and closing operation to obtain a target initial cladding layer boundary graph. The opening operation (first erosion and then expansion) can accurately remove the small noise pixels and fine protruding structures attached to the boundary region in the initial boundary graph. Such noise is mainly caused by the misjudgment of local interference features during the prediction of the deep learning model (such as background impurities in the training data and pseudo-boundary protrusions caused by model fitting deviation), and the opening operation can eliminate these irrelevant interference without damaging the overall morphology of the main boundary, thereby avoiding the misjudgment of noise into the effective boundary during the subsequent screening of the main boundary region.

[0057] The closing operation (first expansion and then erosion) can effectively fill the small holes inside the main boundary in the initial boundary graph, and connect the gaps between adjacent boundary pixels caused by insufficient model feature learning. Such small holes and gaps are important causes of increased difficulty in subsequent boundary extension repair and deviation in size calculation. The closing operation can make the main boundary form a more continuous and complete contour shape by regularizing the local structure of the boundary. The target initial cladding layer boundary graph formed by the combination of the two has a smoother and more continuous boundary shape than the initial cladding layer boundary graph, which not only reduces the interference of noise on the subsequent steps, but also lays a more regular and clear boundary foundation for the accurate screening of the main boundary region and the endpoint extension repair, thereby providing a guarantee for improving the accuracy of the final cladding cross section size prediction from the preprocessing stage.

[0058] Step S2: sort the areas of all connected regions in the initial cladding layer boundary map, take the largest connected region as the main boundary region, and the rest as non-main regions, only keep the pixels of the main boundary region, and obtain the processed cladding layer boundary map;

[0059] The determination method of the connected region includes:

[0060] If two pixels in the initial cladding layer boundary map are adjacent in the horizontal, vertical or diagonal direction of their 8-neighborhood, they are considered to be connected with each other; a connected region is the largest set composed of all the connected pixels, that is, any two pixels in the region can be connected through a series of adjacent pixels, and any boundary pixel outside the region cannot be connected with the pixels in the region through the 8-neighborhood relationship.

[0061] The area of the connected region refers to the total number of pixels contained in the connected region, as shown in (a) of Figure 2 Figure 2 There are three connected regions in the initial cladding layer boundary map in (a) of (a), and the areas of the connected regions are sorted, only the pixels of the main boundary region are kept, and the processed cladding layer boundary map is obtained.

[0062] In this embodiment, preferably, the distance between each non-main region and the geometric center of the initial main boundary region is calculated, the non-main regions with a distance less than a set threshold value are merged with the main boundary region as the target main boundary region.

[0063] By calculating the distance between each non-main region and the main boundary region, the non-main regions with a distance less than a set threshold value are merged into the main boundary region as the target main boundary region, which can effectively avoid the problem of effective boundary loss caused by simply selecting by area. In the cladding cross-section boundary prediction, part of the non-main regions (such as small branches near the main boundary, local protrusions or broken short boundaries) are small in area, but are actually an organic part of the cladding layer profile. If they are deleted as noise only because of their small area, the boundary structure will be incomplete, which will affect the accuracy of subsequent size calculation. By distance threshold determination and merging of close non-main regions, these effective local structures closely related to the main boundary can be accurately retained, which not only eliminates isolated noise regions far from the main boundary, but also ensures the integrity and continuity of the main boundary region, so that the processed boundary map is more consistent with the actual profile characteristics of the cladding layer.

[0064] In this embodiment, the distance for determining the relevance of the non-main region and the main boundary region can be calculated in various ways, such as Euclidean distance, Manhattan distance, Chebyshev distance.

[0065] In this embodiment, the minimum Euclidean distance between each non-main region and the main region is calculated, and if the distance is less than a preset threshold value, the small region is kept and merged into the main boundary region. ​

[0066] The minimum Euclidean distance (i.e. the straight-line distance between the nearest two points between the non-main region and the main boundary region) is selected as the basis for judgment, which has the following advantages: on the one hand, the minimum Euclidean distance can most intuitively reflect the spatial proximity of the two regions, and can better capture the local connection relationship between the non-main region and the main boundary (such as the close proximity of an end point of the non-main region to the main boundary), avoiding the deletion of effective structures that are locally connected due to the distance between the centers of the regions as a whole; on the other hand, the Euclidean distance is more sensitive to the slight spatial position changes of the boundary, and can accurately distinguish between "effective small regions that are locally close to the main boundary" and "completely isolated noise regions", and is particularly suitable for scenarios where the main boundary of the cladding layer may have local protrusions, depressions or slight branches, thereby maximizing the retention of effective structures while minimizing the removal of noise, further improving the accuracy of the target main boundary region.

[0067] To avoid the deletion of small regions on the boundary path, the minimum Euclidean distance between each non-main region and the main region is further calculated, and if the distance is less than a predetermined threshold, the small region is retained and merged into the main boundary region.

[0068] Step S3: For each end point pixel in the processed cladding layer boundary map, starting from the end point pixel, extending in the opposite direction of the end point pixel pointing to its adjacent pixel belonging to the main boundary region, until the extension path of the end point pixel coincides with the pixels of the main boundary region or reaches the boundary of the processed cladding layer boundary map;

[0069] In this embodiment, specifically, the determination rule of the end point pixel includes two cases:

[0070] (1) There is only one adjacent pixel in the 8-neighborhood of the pixel;

[0071] (2) There are two adjacent pixels in the 8-neighborhood of the pixel, and the two adjacent pixels have a connection relationship in a single direction, i.e. they do not form a closed or multi-directional extension structure.

[0072] If any of the above conditions is met, the pixel is an end point pixel.

[0073] Step S4: If the extension path of the end point pixel reaches the boundary of the processed cladding layer boundary map, starting from the boundary of the cladding layer boundary map, in the opposite direction of the extension direction of the end point pixel, performing pixel-by-pixel deletion until reaching the intersection point of the extension path of the end point pixel and the main boundary region, to obtain a target cladding layer boundary map; wherein the intersection point satisfies that there are three adjacent pixels in the 8-neighborhood of the pixel, and each adjacent pixel has an extension direction independent of the main path, such pixels constitute intersection positions in the path structure, and the extension and deletion operations are terminated based on the intersection positions.

[0074] For example, Figure 2As shown in (b) in FIG. 4, the end point pixel 4 is extended along the direction opposite to the direction in which it points to its adjacent pixel belonging to the main boundary region until the extension path of the end point pixel coincides with the pixel of the main boundary region, while the extension paths of the end point pixels 1, 2, 3 and 5 reach the boundary of the processed cladding layer boundary map and need to be further processed. Starting from the boundary of the cladding layer boundary map, the pixels are deleted in the direction opposite to the extension direction of the end point pixels 1, 2, 3 and 5 respectively until the intersection point of the extension path of the end point pixels 1, 2, 3 and 5 with the main boundary region is reached.

[0075] In this embodiment, the determination method of the intersection point specifically includes:

[0076] It is determined whether there are three adjacent pixels belonging to the main boundary region in the 8-neighborhood of the current pixel. If there are, it is determined whether the extension paths formed by the current pixel and each of its adjacent pixels do not coincide. If they do not coincide, the current pixel is an intersection point.

[0077] Through the above steps, the application can effectively repair the boundary breakage and loss caused by model prediction error, ensure the continuity and geometric accuracy of the cladding layer boundary, and thus improve the reliability of subsequent size measurement and feature extraction.

[0078] Step S5: based on the target cladding layer boundary map, obtaining the size parameter of the molten pool image to be predicted.

[0079] In this embodiment, specifically, the size parameter of the molten pool image to be predicted includes the cross-sectional area, width and height of the cladding layer.

[0080] Although traditional convolutional neural networks and Transformers have shown excellent performance in computer vision tasks such as image classification and object detection, there are natural limitations in their feature extraction and modeling logic. The design of such models is usually based on the assumption that there is a high visual correlation between the input image and the output result. For example, in image classification, the input is a clear image of the target, and the output is the corresponding category label. In the detection task, both the input and the output are around the target position information of the same image, which essentially does not deviate from the category of the same domain feature association. Therefore, when facing the typical cross-domain image conversion demand such as the molten pool image and the cladding layer cross-sectional image, the traditional model is difficult to break through the essential difference between the two types of images in the visual domain (the molten pool image presents a dynamic high-temperature region form, and the cladding layer cross-sectional image exhibits a static cross-sectional contour structure), cannot accurately capture the process mapping relationship between them, and naturally cannot achieve effective conversion from the molten pool feature to the cladding layer cross-sectional feature.

[0081] To solve the above problem, the application adopts a Pix2Pix model based on a conditional generative adversarial network (CGAN) as a core model framework, and combines a perception loss function, the model is naturally capable of processing the conversion of images in different domains, and can gradually learn the potential mapping rules between cross-domain images through the adversarial training of the generator and the discriminator; at the same time, in order to further improve the conversion accuracy, the application additionally introduces a perception loss function, uses a pre-trained feature extraction network (such as a VGG network) to capture the deep semantic features of the image, so that the model not only focuses on the surface differences at the pixel level, but also understands the deep feature correlation of the two types of images under the process logic, and finally significantly enhances the understanding ability and generation accuracy of the model for the cross-domain feature mapping of the molten pool image and the cladding layer cross-section image.

[0082] As shown in Figure 3 , Figure 3 is a structural diagram of the cladding cross-section size prediction model of the application.

[0083] In this embodiment, preferably, the training process of the Pix2Pix model comprises:

[0084] obtaining a molten pool image training set, inputting the molten pool image into the generator of the Pix2Pix model, and outputting an initial cladding layer boundary image of the molten pool image;

[0085] The generator (Generator) of the Pix2Pix model adopts a U-Net (UnetGenerator) structure, the input channel number is 3, the output channel number is 3, the output is an RGB format cladding layer boundary image, the basic channel number is 64, and a total of 8 layers of down-sampling structures (num_down=8) are set. In the encoding and decoding process, a batch normalization (Batch Normalization, BN) layer is used to speed up the convergence speed, and a Dropout operation is introduced in part of the layers to enhance the generalization ability of the model. The generator weight initialization method is a normal distribution (Normal Initialization) with a mean of 0 and a variance of 0.022.

[0086] The U-Net structure extracts multi-scale features through an encoder, and uses a skip connection to fuse high-resolution spatial information in the decoding process, so as to retain the structural features of the input image.

[0087] extracting prediction features of multiple levels of the feature extraction network through the pre-trained feature extraction network of the initial cladding layer boundary image of the molten pool image;

[0088] The label image of the molten pool image is extracted through a pre-trained feature extraction network to obtain real features of multiple levels of the feature extraction network;

[0089] In this embodiment, the feature extraction network adopts a VGG network, and the predicted features and the real features of the 4th layer, the 9th layer and the 16th layer of the feature extraction network are extracted.

[0090] The molten pool image and the initial cladding layer boundary image thereof are input into a discriminator of a Pix2Pix model to obtain a first confidence;

[0091] The molten pool image and the label image thereof are input into the discriminator of the Pix2Pix model to obtain a second confidence;

[0092] The discriminator (Discriminator) of the Pix2Pix model adopts a PatchGAN structure (PatchDiscriminator), the number of input channels is 6 (obtained by splicing a 3-channel molten pool image (input) and a 3-channel target boundary image (label / generated result) in the channel dimension, the number of basic channels is 64, the number of convolution layers is 3, and feature standardization is performed using batch normalization. PatchGAN takes a local region (Patch) of an image as a discrimination unit, each Patch independently determines its authenticity, and the fidelity of texture and details can be improved.

[0093] The Pix2Pix model is trained based on conditional adversarial loss between the first confidence and the second confidence, pixel-level L1 loss between the initial cladding layer boundary image of the molten pool image and the label image, and perceptual loss between the predicted features and the real features of multiple levels of the feature extraction network, to obtain a target Pix2Pix model.

[0094] The formula of the conditional adversarial loss (GAN Loss) between the first confidence and the second confidence is:

[0095] ,

[0096] Among them, is the conditional adversarial loss, is the generator, is the discriminator, is the source domain image (molten pool image), is the target domain image (label image), is random noise, is the mathematical expectation, is the second confidence, is the first confidence, is the initial cladding layer boundary image, indicates that the generator minimizes the conditional adversarial loss, This indicates that the discriminator should maximize the conditional resistance to loss.

[0097] The generator relies not only on random noise vectors when generating target data, but also on pre-existing conditional information in the dataset. This allows it to establish a mapping between input and output. The discriminator aims to maximize the probability difference between real and generated samples, while the generator aims to minimize this difference.

[0098] This invention introduces a pixel-wise L1 loss between the initial cladding layer boundary map of the molten pool image and the label image to constrain the pixel-level difference between the generated image and the real image, making the generated image closer to the target domain image. The loss weight is set to 100.0 to ensure structural consistency. The formula is as follows:

[0099] ,

[0100] in, For pixel-level L1 loss, It is an L1 norm.

[0101] To compensate for the structural differences between the source and target domains and improve the similarity of the generated results in the high-level semantic feature space, this invention introduces a perceptual loss based on the original pix2pix loss function. The perceptual loss is calculated based on the multi-layer feature representations (layers 4, 9, and 16) of the pre-trained feature extraction network (VGG network), using the mean squared error (MSE) form. The formula for the perceptual loss between the predicted features and the true features at multiple layers of the feature extraction network is as follows:

[0102] ,

[0103] in, In order to perceive loss, This represents the number of layers in the feature extraction network. It is a molten pool image. It is a label image. It is random noise. This represents the initial cladding layer boundary diagram. Represents a generator. For the feature extraction network Predictive features of the layer For the feature extraction network The true characteristics of the layer For the feature extraction network The height of the layer's features For the feature extraction network The width of the layer's features, For the feature extraction network the number of channels of the feature of the layer, is the layer index, is the L2 norm.

[0104] In actual industrial production, the acquisition cost of the inclined base surface cladding data is high, the collection cycle is long, and the cross-section labeling process is tedious, so that the scale of the high-quality data set available for training is limited. The existing prediction methods based on deep learning generally need large-scale data to obtain good generalization ability, and in the case of lack of data, these methods are prone to overfitting, which leads to a significant decline in prediction effect under new working conditions or new material conditions.

[0105] In order to solve the above problems, the present application uses a variety of data enhancement methods (such as image inversion, brightness / contrast adjustment, cropping, channel transformation, etc.), boundary conversion and enhancement method based on structural similarity index (SSIM) to increase the data size by about ten times, thereby significantly alleviating the small sample problem, and the specific scheme is as follows:

[0106] As shown in Figure 4 , the acquisition process of the molten pool image training set is shown in Figure 4 .

[0107] In this embodiment, preferably, the acquisition process of the molten pool image training set comprises:

[0108] Obtaining molten pool videos under different parameter conditions, selecting any one frame of molten pool image in the molten pool videos under different parameter conditions, and the cross-section image of the cladding layer under the parameter condition as a label image as a sample in the molten pool image training set;

[0109] In this embodiment, specifically, the different parameter conditions refer to the core process parameters that affect the cladding process (such as laser cladding, electric arc cladding, etc. heat processing technology), such as laser power, powder feeding rate, scanning speed.

[0110] In this embodiment, in the molten pool videos obtained under each different parameter condition, first, a single frame of molten pool image is selected as an input image, and a cross-section image of the cladding layer under the condition is found as a label image to form a group of training samples. Due to the high cost and time-consuming of the acquisition process of the molten pool image and the cross-section image, the experiment finally only obtains 108 groups of sample data, and after being divided into training set and test set, the number of samples available for training is more limited.

[0111] Data augmentation is performed on the molten pool images selected under different parameter conditions to obtain data augmented molten pool images under different parameter conditions;

[0112] In order to solve the small sample problem and improve the model generalization ability, the present application adopts the following measures in the data augmentation link:

[0113] Random rotation: randomly rotate the image to a certain angle (range [-degrees, +degrees]) to simulate different shooting directions; where degrees is the maximum offset of the rotation angle;

[0114] Random cropping: randomly select a region of a specified size in the image for cropping to change the local information of the input image;

[0115] Random style conversion: convert the image to a grayscale image with a certain probability (e.g., 30%) and maintain the RGB channel structure to enhance the model's robustness to brightness and texture changes;

[0116] Color disturbance: randomly adjust the brightness, contrast, saturation, and hue of the image to simulate different lighting environments.

[0117] Since the prediction task of the present application is only for the outer contour of the cladding layer cross section, in the enhancement process, the boundary line of about 5 pixels of the outermost periphery of the cladding layer cross section in the label image corresponding to the data enhanced molten pool image under different parameter conditions is extracted to form an independent cladding layer boundary image; then the cladding layer boundary image is spliced with the corresponding data enhanced molten pool image to generate a data enhanced molten pool image with stronger structural information.

[0118] To make full use of video information, the present application extracts multiple frames of images from each molten pool video as additional samples, but in order to avoid the model learning deviation caused by too large differences between different frames in the same video, the present application uses SSIM for screening. In the molten pool video under different parameter conditions, the molten pool images other than the selected molten pool images are used as candidate images under that parameter condition;

[0119] The structural similarity index of each candidate image under different parameter conditions and the data enhanced molten pool image under that parameter condition is calculated, and the candidate images are sorted in descending order, and the first pre-set number of candidate images and the cladding layer cross section image under that parameter condition are used as label images, which are used as expansion samples in the molten pool image training set.

[0120] In this embodiment, taking the originally selected molten pool image as the reference, the pre-set frame number is set to 9, and the 9 most similar frames with the highest structural similarity index are selected for data expansion. Through the above processing, the original 108 group data set is expanded to about 10 times the size, effectively increasing the number of effective samples for model training and enhancing the diversity and representativeness of the data.

[0121] In this embodiment, in order to obtain the laser cladding molten pool and cladding layer cross-section data of different inclined surfaces, the metal substrate is first fixed on a workbench with adjustable inclination angle. By adjusting the inclination angle of the workbench, the substrate can assume different inclined states. At the same time, the posture angle of the mechanical arm is adjusted to ensure that the coaxial powder feeding nozzle is perpendicular to the substrate surface under any substrate inclination, thereby ensuring the stability of powder feeding and the forming quality.

[0122] During the experiment, four substrate inclination posture angles were selected for laser cladding test, namely 0°, 30°, 60° and 90°. Under each posture angle, the following process parameter combinations were set:

[0123] Laser power: 1200 W, 1000 W, 800 W;

[0124] Scanning speed: 3 mm / s, 3.5 mm / s, 7 mm / s;

[0125] Powder feeding rate: 4 g / min, 6 g / min, 8 g / min.

[0126] During the laser cladding process, a high-speed camera is used to take lateral shots of the molten pool. The camera and the substrate maintain a 35° angle, and the shooting distance is 20 cm. The equipment used is Acuteye V4.0 welding high-speed camera, which can capture clear images at high speed to ensure the fidelity of the details of the molten pool dynamic process.

[0127] After the cladding is completed, the single-track cladding layer is line-cut to obtain the cross-section sample. The sample is inlaid, polished and processed to ensure that the cross-section is flat and smooth, facilitating microscopic observation. Then, the MX6R optical microscope is used to take the cross-section image of the cladding layer, which serves as the label data required for subsequent training models.

[0128] This embodiment is based on the MMgeneration environment and is performed on an NVIDIA RTX3090 GPU. The training process of the cladding cross-section size prediction model uses Adam optimizer for both the generator and the discriminator, with a learning rate of , the momentum parameter in the Adam optimizer, the exponential decay rate of the second moment estimate of the Adam optimizer The total number of training iterations is 300,000, a training log is recorded every 100 iterations, a model checkpoint is saved and a performance evaluation is performed every 10,000 iterations, the evaluation indexes include average surface distance (ASD) and Hausdorff distance (Hausdorff), the number of evaluation samples is 214, and the image resolution is 512*512. In order to improve the training efficiency and stability, distributed data parallelism (NCCL backend) is used, cudnn_benchmark is enabled to optimize convolution calculation, OpenCV multi-threading is disabled and multi-processes are started in fork mode; at the same time, the generated result is visualized and saved every 5000 iterations to monitor the generation quality of the model.

[0129] In order to verify the effectiveness of the laser cladding cross-section size prediction method of the inclined base surface, the method of the present application is compared with a plurality of existing mainstream models, including Unet and DeepLabV3+ in the convolutional neural network, ViT based on the Transformer structure, and the Pix2Pix model in CGAN.

[0130] The data set is divided into a training set and a test set in a ratio of 8:2, and the experimental results are shown in Table 1, which is a comparison of experimental results of different models.

[0131] Table 1

[0132]

[0133] From the experimental results, it can be seen that the method of the present application is superior to other comparison models in terms of two key indicators, average surface distance (Average Surface Distance, ASD) and Hausdorff distance (Hausdorff Distance).

[0134] The average surface distance (ASD) is used to measure the average shortest distance between the predicted boundary and the real boundary, reflecting the accuracy of the overall boundary fitting. The smaller the ASD, the closer the predicted result is to the real contour. The formula is:

[0135] ,

[0136] Among them, is the average surface distance between and , is the boundary pixel set of the target cladding layer boundary graph, is the boundary pixel set of the real cladding layer cross-section graph, is the predicted contour in the target cladding layer boundary graph, This represents the true outline of the actual cladding layer cross-section. for The elements in for The elements in For a single pixel arrive The shortest distance between all pixels in the array; in this embodiment, the distance is the Euclidean distance. For a single pixel arrive The shortest distance between all pixels in the array. The total number of pixels in the boundary pixel set of the target cladding layer boundary map. This represents the total number of pixels in the boundary pixel set of the actual cladding layer cross-sectional image.

[0137] The Hausdorff distance measures the maximum distance between two sets of boundary points, focusing on the farthest deviation on the boundary and reflecting extreme error conditions. A smaller Hausdorff distance indicates a stronger fit to outliers or marginal regions. Its formula is:

[0138] ,

[0139] in, for and The distance between Hausdorf and the other side, It is a norm.

[0140] Traversing point sets Each pixel in (i.e., for any) ), in point set Find the pixel The closest pixel in space (i.e., determine) And satisfy the distance (minimum), for Each pixel in Find the nearest points to form a pair; calculate the distance between all such pairs, and select the largest distance, denoted as . Similarly, we can obtain ,Pick , The larger value in the middle is used as .

[0141] like Figure 5 As shown, Figure 5 These are experimental comparison figures showing the prediction models for different cladding cross-section sizes and the method of this invention. Figure 5(a) is a schematic diagram comparing the prediction models for different cladding cross-sectional dimensions with the target cladding layer boundary map extracted from the first molten pool image by the method of this invention. Figure 5 (b) is a schematic diagram comparing the prediction models for different cladding cross-sectional dimensions with the target cladding layer boundary map extracted from the second molten pool image by the method of this invention. Figure 5 The diagram in (c) shows a comparison between the prediction models for different cladding cross-section sizes and the target cladding layer boundary map extracted from the third molten pool image by the method of this invention. Figure 5 The diagram shows a comparison between the prediction models for different cladding cross-sectional dimensions in (d) and the target cladding layer boundary map extracted from the fourth molten pool image by the method of this invention.

[0142] Figure 5 The visualization comparison results of the method of this invention with several existing mainstream models are presented. It is clearly observed from the comparison figures that the method of this invention outperforms other methods in terms of the integrity of the cladding layer cross-sectional profile, boundary continuity, and detail reproduction, demonstrating higher prediction accuracy and superior visual effects.

[0143] like Figure 6 As shown, Figure 6 This diagram illustrates a comparison between prediction models for different cladding cross-section sizes and prediction results for cladding layer size using the method of this invention. Figure 6 (a) is a schematic diagram comparing the prediction models for different cladding cross-sectional dimensions with the method of this invention for predicting the cross-sectional area of ​​the cladding layer. Figure 6 (b) is a schematic diagram comparing the predicted width of different cladding cross-section size prediction models with that predicted by the method of this invention. Figure 6 The diagram shows a comparison between the prediction models for different cladding cross-section dimensions and the height predicted by the method of this invention, with behavior (c) in the figure.

[0144] Figure 6 This paper presents a detailed comparison of the best-performing ViT model, Pix2Pix model, and the method of this invention in three dimensions: cross-sectional area, width, and height of the cladding layer. The evaluation metrics include the coefficient of determination (COP). The mean absolute error (MAE) and mean squared error (MSE) of the method are clearly observed in the scatter plot. The prediction results of the method of the present invention are closely distributed near the ideal fitting line (45-degree line) for these three key parameters, indicating that its fitting accuracy to the actual size is significantly better than the comparison model, and the prediction results are more accurate and stable.

[0145] Among them, the coefficient of determination ( : represents the goodness of fit between the predicted and actual values, with a numerical range of [−∞,1], where 1 indicates a perfect fit, 0 indicates that the model has no predictive ability, negative values ​​indicate that the model performs worse than the simple mean, and higher values ​​indicate that the model performs better than the simple mean. The value indicates that the model has a strong ability to explain changes in the data.

[0146] Mean Absolute Error (MAE): is the average of the absolute values of the errors between predicted and actual values, reflecting the average size of the prediction error. The smaller the MAE, the more accurate the prediction result and the more concentrated the error.

[0147] Mean Squared Error (MSE): is the average of the squares of the prediction errors, emphasizing the punishment effect of large errors. The smaller the MSE, the smaller the overall error of the model and the better the prediction stability.

[0148] As shown in Table 2, Table 2 is the prediction performance of the cladding layer size of different methods and the method of the present application on the test set.

[0149] Table 2

[0150]

[0151] As shown in Table 3, Table 3 is the ablation experiment result of the method of the present application.

[0152] Table 3

[0153]

[0154] The present application verifies the influence of perceptual loss, enhanced method based on structural similarity index (SSIM) and endpoint extension repair algorithm based on graph structure analysis on the prediction performance of cladding cross-sectional size.

[0155] The second embodiment provides a laser cladding cross-sectional size prediction system for inclined base surface, comprising:

[0156] The memory is used to store the computer program, and the processor is used to execute the computer program to realize the steps of the above-mentioned laser cladding cross-sectional size prediction method for inclined base surface.

[0157] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems or computer program products. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0158] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0159] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 one or more flow or blocks

[0160] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 Figure 1 one or more flow or blocks

[0161] Obviously, the above-described embodiments are only examples for clarity of description and are not limiting on the embodiments. Based on the above description, one of ordinary skill in the art can further make other different forms of changes or modifications. Here, all the embodiments are not required to be enumerated, and the obvious changes or modifications derived therefrom are still within the protection scope of the present application.

Claims

1. A method for predicting the cross-sectional dimensions of laser cladding on an inclined datum, characterized in that, include: The initial cladding layer boundary map is obtained by using the cladding cross-section size prediction model to predict the image of the molten pool to be predicted. The areas of all connected regions in the initial cladding layer boundary map are sorted, and the connected region with the largest area is taken as the main boundary region, while the remaining regions are taken as non-main regions. Only the pixels of the main boundary region are retained to obtain the processed cladding layer boundary map. For each endpoint pixel in the processed cladding layer boundary map, starting from the endpoint pixel, extend in the opposite direction to the adjacent pixel belonging to the main boundary region until the extension path of the endpoint pixel coincides with the pixel of the main boundary region or reaches the boundary of the processed cladding layer boundary map. If an endpoint pixel's extension path reaches the boundary of the processed cladding layer boundary map, then starting from the boundary of the cladding layer boundary map, pixel-by-pixel deletion is performed in the opposite direction of the endpoint pixel's extension direction until the intersection of the endpoint pixel's extension path and the main boundary region is reached, thus obtaining the target cladding layer boundary map; wherein, the method for determining the intersection point includes: Determine if there are 3 adjacent pixels belonging to the main boundary region in the 8-neighborhood of the current pixel. If so, determine if the extension paths formed by the current pixel and each of its adjacent pixels do not overlap. If they do not overlap, the current pixel is an intersection point. Based on the target cladding layer boundary map, obtain the size parameters of the molten pool image to be predicted.

2. The method for predicting the cross-sectional dimensions of laser cladding on an inclined datum surface according to claim 1, characterized in that, Also includes: The initial cladding layer boundary map of the target is obtained by performing morphological opening and closing operations on the initial cladding layer boundary map.

3. The method for predicting the cross-sectional dimensions of laser cladding on an inclined datum surface according to claim 1, characterized in that, Also includes: Calculate the distance between each non-main region and the geometric center of the initial main boundary region. Merge non-main regions with a distance less than a set threshold with the main boundary region to form the target main boundary region.

4. The method for predicting the cross-sectional dimensions of laser cladding on an inclined datum surface according to claim 1, characterized in that, The prediction model for cladding cross-section size is the Pix2Pix model.

5. The method for predicting the cross-sectional dimensions of laser cladding on an inclined datum surface according to claim 4, characterized in that, The training process of the Pix2Pix model includes: Obtain the training set of molten pool images, input the molten pool images into the generator of the Pix2Pix model, and output the initial cladding layer boundary map of the molten pool images; The initial cladding layer boundary map of the molten pool image is processed through a pre-trained feature extraction network to extract predicted features from multiple levels of the feature extraction network. The labeled image of the molten pool image is passed through a pre-trained feature extraction network to extract real features from multiple levels of the feature extraction network; The first confidence level is obtained by passing the molten pool image and its initial cladding layer boundary map through the discriminator of the Pix2Pix model; The second confidence level is obtained by passing the molten pool image and its label image through the discriminator of the Pix2Pix model; The Pix2Pix model is trained based on the conditional adversarial loss between the first and second confidence levels, the pixel-level L1 loss between the initial cladding layer boundary map of the molten pool image and the label image, and the perceptual loss between the predicted features and the real features of multiple layers of the feature extraction network to obtain the target Pix2Pix model.

6. The method for predicting the cross-sectional dimensions of laser cladding on an inclined datum surface according to claim 5, characterized in that, The formula for the perceptual loss between the predicted features and the true features at multiple levels of the feature extraction network is as follows: ， in, In order to perceive loss, This represents the number of layers in the feature extraction network. It is a molten pool image. It is a label image. It is random noise. This represents the initial cladding layer boundary diagram. Represents a generator. For the feature extraction network Predictive features of the layer For the feature extraction network The true characteristics of the layer For the feature extraction network The height of the layer's features For the feature extraction network The width of the layer's features, For the feature extraction network The number of channels in the layer's features. For layer index, It is an L2 norm.

7. The method for predicting the cross-sectional dimensions of laser cladding on an inclined datum surface according to claim 5, characterized in that, The process of obtaining the training set of molten pool images includes: Acquire molten pool videos under different parameter conditions, select any frame of molten pool image from the molten pool videos under different parameter conditions, and the cross-sectional image of the cladding layer under that parameter condition as label images, and use them as samples in the molten pool image training set. Data augmentation was performed on the molten pool images selected under different parameter conditions to obtain data-augmented molten pool images under different parameter conditions; In the molten pool video under different parameter conditions, the molten pool images other than the selected molten pool images are used as candidate images under that parameter condition; Calculate the structural similarity index between each candidate image and the data-enhanced cladding pool image under different parameter conditions, sort them in descending order, and use the candidate images of the previous preset number of frames and the cross-sectional images of the cladding layer under the same parameter conditions as label images, which are then used as augmented samples in the cladding pool image training set.

8. The method for predicting the cross-sectional dimensions of laser cladding on an inclined datum surface according to claim 1, characterized in that, The dimensional parameters of the molten pool image to be predicted include: the cross-sectional area, width, and height of the cladding layer.

9. A system for predicting the cross-sectional dimensions of laser cladding on an inclined datum, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the method for predicting the cross-sectional dimensions of a tilted datum laser cladding as described in any one of claims 1 to 8.

Images