Intelligent and standardized welding position identification method and system for hot-rolled aluminum coil

Abstract

The invention discloses an intelligent and standardized welding position identification method and system for a hot-rolled aluminum coil, belongs to the technical field of welding position identification, and aims to realize accurate planning of a welding path of a large aluminum coil. The method comprises the following steps: firstly, accurately identifying welding areas, namely identifying three welding spot areas on the end surface of an outer ring, one welding spot area on the end surface of an inner ring and one welding spot area on a belt head; according to the method, warping, staggered layers and other abnormal areas which are not suitable for welding in the aluminum coil can be automatically eliminated, and therefore high-quality welding area recognition is achieved. Normalized welding path planning is carried out by utilizing an accurate identification result, and accurate guidance information is provided for subsequent welding work.

Description

Technical Field

[0001] The invention belongs to the technical field of welding position recognition, and in particular relates to an intelligent and standardized welding position recognition method and system for hot-rolled aluminum coils. Background Art

[0002] In the current production process of aluminum coils, manual welding is required after cold rolling to enhance the stability of the aluminum coils and prevent the aluminum coils from cracking during subsequent transportation. However, there are two main problems with manual welding: first, standardization. Manual welding is usually performed by welders based on experience, which results in the number and position of welding points not being fixed, and the height problem causes the welding points to be concentrated in the lower half, which poses a potential risk of cracking; second, safety issues. The high-temperature and toxic working environment poses a major threat to the health of workers. Therefore, it is necessary to use an intelligent welding system that uses a robotic arm and deep learning methods for automated intelligent welding instead of manual welding.

[0003] However, due to the inherent metallic properties of aluminum, the unique structure of aluminum coils, and the weak light conditions that are prevalent in actual production environments, this poses a series of challenges to intelligent welding systems. Specifically: 1) Structural challenges. The coiled structure of aluminum causes misalignment between layers, resulting in welds that are not on the same plane. This is significantly different from the flat seam welding of steel plates, making the seam recognition module designed for steel plates ineffective. 2) Lighting challenges. There will be a low-light environment in aluminum production, which makes it difficult to identify the welds in the aluminum coil during this period. In this case, this increases the complexity of weld identification. 3) Material sensitivity. As a soft metal, artificial intelligence is very sensitive to force. The welding point will affect the internal stress distribution of the entire aluminum coil, and improper handling may lead to uneven stress and internal collapse. Therefore, the selection of the welding path is crucial. Summary of the invention

[0004] In response to the above challenges, the present invention proposes a new intelligent identification method for the welding area of ​​aluminum coils. This method comprehensively considers the characteristics of aluminum coils and various abnormal situations that may occur during the production process. By sampling and analyzing the data of the aluminum coils at multiple angles, it can accurately identify a total of ten specific weldable areas on both sides of the head and end face areas. Secondly, this technology adopts a standardized identification process, combined with advanced image processing technology and deep learning algorithms, to achieve intelligent identification of the welding area of ​​aluminum coils. This not only significantly reduces the involvement of operators in working in high-risk environments, but also improves the production efficiency and welding quality of aluminum coils. Through this invention and innovation, while ensuring the health and safety of operators, it also brings a double improvement in efficiency and safety to the hot rolling production of aluminum coils.

[0005] The technical solution provided by the present invention is: an intelligent and standardized welding position identification method for hot-rolled aluminum coils, comprising:

[0006] Collect 2D images and 3D point cloud data of aluminum coils, enhance the brightness of 2D images with low light, and remove noise from 3D point clouds; the shooting area includes an end face head area and three outer ring end face peripheral areas; the three outer ring end face peripheral areas are located at the vertices of the regular triangle inscribed in the end face circle;

[0007] Construct a target instance segmentation network for key feature detection of aluminum coils. The network includes two branches, 2D and 3D. The 2D branch is based on the YOLOv8 model and is used to extract a weld area on the inner wall of the 2D image. The 3D branch is based on the PointNet++ architecture and is used to extract the end face area of ​​the ring in the 3D point cloud, including three weld areas on the outer end face and one weld area on the inner end face. The three weld areas on the outer end face and one weld area on the inner end face are further classified into non-weldable areas and weldable areas.

[0008] The welding area is further divided within the weldable area, and accurate path planning is performed on the welding point area of ​​the inner wall and the welding area based on the actual size of the welding gun.

[0009] Furthermore, a two-stage exposure method is used to collect 2D images and 3D point cloud data, with the first stage being long exposure and the second stage being short exposure.

[0010] Furthermore, adaptive block γ-correction is used to enhance the brightness of the 2D image. The specific implementation method is as follows:

[0011] First, the image is divided into blocks, and the γ value is adaptively adjusted according to the average brightness of each block:

[0012]

[0013] In the formula, LU(p ij ) represents the brightness of the pixel, M and N are the length and width of the image after the block is divided; calculate the average brightness, using 128 as the threshold: areas with brightness higher than 128 are considered darker, and a γ value less than 1 is applied to enhance the brightness. The smaller the γ value, the stronger the enhancement; areas below 128 are considered brighter, and a γ value greater than 1 is used to compress the brightness. The larger the γ value, the stronger the compression; then use a lookup table on the image block to speed up the γ correction:

[0014]

[0015] k represents the current pixel, Represents the enhanced pixel brightness value, and performs inverse γ enhancement on the pixel.

[0016] Furthermore, a statistical method based on the z-axis is used to denoise the noise points. The intervals are divided according to the length of the z-axis to obtain the interval length h:

[0017]

[0018] Among them, h is the length of the interval, l is the length of the z-axis interval occupied by the point cloud, and z max is the maximum value of the z-axis data in the point cloud data, z min is the minimum value of the z-axis data in the point cloud data, and then record the number of points in each interval, using the interval with the most points as a reference:

[0019]

[0020] P j is the set of all points in the jth interval, and all points p in the interval i Satisfy the z-axis value z i In a size of The limits are set successively by extending to adjacent zero intervals:

[0021]

[0022] j lower is the lower limit of the z-axis size of the final selected valid point, j upper The upper limit of the z-axis size of the final valid points is selected, and the points outside these limits are regarded as noise and removed.

[0023] Furthermore, the input of the two-dimensional branch is the 2D image after brightness enhancement, and it outputs the feature map for predicting the box, the feature map for predicting the classification, and the mask coefficient for segmentation. At the same time, the output of the P5 layer in the Head is upsampled to obtain the mask prototype; the model output is filtered into two prediction boxes by NMS, and the mask prototype and mask coefficient are used to generate the target prediction mask. The prediction box is cropped and thresholded to obtain the segmentation result.

[0024] Furthermore, the input of the three-dimensional branch is a set of three-dimensional coordinates. First, a clustering algorithm is used in the sampling layer to select representative points from the input point cloud to form a smaller subset. Then, for each sampling point, a group of points are selected in its neighborhood in the grouping layer, and a spherical or K-nearest neighbor search is used to define the neighborhood. Finally, the feature extraction layer applies a small PointNet to the points in each neighborhood to extract local features. The local features are processed by MLP to generate a feature vector of fixed length. Through layer-by-layer aggregation, the local features are gradually summarized to a higher level to obtain a global feature vector. The fully connected layer further processes and classifies the extracted global feature vector, and finally outputs the classification label to divide the 3D point cloud into non-weldable areas and weldable areas.

[0025] Furthermore, for a welding point area of ​​the inner wall, the center line is the actual welding path, and the actual welding path is an arc.

[0026] Furthermore, the welding area is further divided within the weldable area according to the thickness n of the plate. When n is less than a certain threshold, the welding area moves radially inward and downward by k1 mm from the boundary line of the weldable area. When n is greater than a certain threshold, the welding area moves radially inward and downward by k2 mm from the boundary line of the weldable area. The outer edge of the welding area is parallel to the outer edge of the weldable area. k1 and k2 are constants.

[0027] Furthermore, the thickness of the plate and strip is obtained by transferring parameters during the rolling process of the aluminum coil, and the welding area is a rectangular arc.

[0028] The present invention also provides an intelligent and standardized welding position identification system for hot-rolled aluminum coils, comprising:

[0029] A processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an intelligent and standardized welding position identification method for hot-rolled aluminum coils as described in the above technical solution.

[0030] Compared with the prior art, the beneficial effects and advantages of the present invention are as follows:

[0031] ① Improve production efficiency: Traditionally, the identification of welding areas relies on experienced workers operating at close range, especially in high temperature environments, where workers need to take frequent breaks and adjust. The system proposed in this invention can work continuously for 8 to 10 hours at 50°C, greatly improving production efficiency.

[0032] ② Improve production standardization: During manual welding, the selection of welding positions and welding points usually depends on the workers' experience and is highly subjective. This system can perform scientific and accurate welding point selection under the premise of accurate feature recognition, ensuring the standardized selection of welding points, meeting engineering requirements, and providing a standardized solution for the automation of welding point area recognition in the entire industry.

[0033] ③ Reduce the impact of the production environment on human health: The temperature in the aluminum coil production workshop is usually above 40°C. In addition, the temperature is high when the aluminum coil is welded. When workers are exposed to high temperature for a long time during close operation, it may have an adverse effect on their health. This system significantly reduces the exposure time of workers to high temperatures and reduces the harm of high temperature operations to workers' health. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The following is the algorithm workflow diagram.

[0035] Figure 2 Schematic diagram of data collection and processing.

[0036] Figure 3 Schematic diagram of normalization of welding area. DETAILED DESCRIPTION

[0037] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0038] The embodiment of the present invention proposes an intelligent planning method for the welding area path of a large aluminum coil. The method aims to solve the intelligent problem in the aluminum coil welding process, thereby realizing the efficiency and standardization of the welding operation. The method includes three parts: data collection, feature extraction and path planning.

[0039] The data acquisition part comprehensively considers the temperature, shooting accuracy and receptive field of the actual production workshop, and optimizes the best shooting order. And the image preprocessing is performed according to the actual captured images: brightness enhancement is performed for the low light problem of 2D images, and noise removal is performed for the noise problem of 3D point clouds. This ensures the accuracy of model recognition. The feature extraction part uses the image segmentation method to segment the head part in the 2D image and the staggered part in the 3D point cloud, which plays a vital role in the accurate planning of the welding path. The path planning part reasonably plans the number and position of welding points based on the feature extraction results, including the welding area with an equilateral triangle on the outer ring end face, the welding area with the inner ring end face extending to the head position, and the welding area at the outer end of the head. And according to the actual size of the welding gun, accurate path planning is performed to provide a reliable and accurate path for the welding of the welding gun.

[0040] The specific working process of the present invention is as follows:

[0041] The present invention aims to achieve accurate planning of the welding path of large aluminum coils. First, the welding area is accurately identified, including the three welding spot areas on the outer ring end face, one welding spot area on the inner ring end face, and one welding spot area on the lead. This method can automatically exclude abnormal areas such as warping and staggered layers in the aluminum coil that are not suitable for welding, thereby achieving high-quality welding area identification. After that, using the accurate identification results, standardized welding path planning is carried out to provide accurate guidance information for subsequent welding work. The specific workflow is as follows:

[0042] Lead welding path planning: The robot arm carries the camera and moves to a distance of 100-150cm from the center of the aluminum coil (the distance depends on the size of the aluminum coil) to collect data. The collected data is input into the pre-trained neural network model to segment the coil head welding area. The segmentation results are used to generate subsequent welding guidance information. Workers can monitor and supervise in the background to prevent occasional abnormal situations from affecting the segmentation accuracy and causing recognition errors; path planning is performed based on the recognition results, the discrete point coordinates of the path are given, and the path is converted from the image coordinate system to the robot arm coordinate system to guide the welding work of the welding gun.

[0043] End face welding path planning: According to the height and position of the aluminum coil, the robot arm carrying the camera first moves to the highest point in the center of the aluminum coil for data collection. The collected data is input into the network model to exclude abnormal positions such as staggered layers and warping. Finally, five three-dimensional coordinate points based on the robot arm are output at the outermost circle to calibrate the welding arc area at the highest point. The arc area is 50mm long and the width is fixed at 5mm. In addition, the robot arm needs to move to the two specified positions below to collect information, and return the coordinate information of the rectangular welding area along the diameter after the same processing. In this process, since the point cloud coordinates of the MechProS camera are centered on the camera, additional coordinate conversion is required through the coordinates of the center of the aluminum coil to obtain the true rectangular welding area along the diameter.

[0044] The present invention is a method for realizing intelligent identification of welding areas of large aluminum coils, including data collection, feature extraction and path planning. The overall flow chart is as follows: Figure 1 The detailed design is as follows:

[0045] 1. Data Collection

[0046] The shooting sequence and receiving field cover four areas, including an end face leading area and three end face peripheral areas. The shooting of the end face leading area will simultaneously capture the leading area and the inner circle of the end face, and obtain two data at one time. The three end face peripheral areas are the vertices of the equilateral triangle inscribed in the end face circle. These areas are selected for feature-rich data collection. The two-stage exposure method is adopted: long exposure is 50 milliseconds and short exposure is 7 milliseconds. The two-stage exposure method avoids the problems of dark light and incomplete point cloud data in a single exposure information collection. The two-stage exposure acquisition ensures that the details of the light and dark areas in high-contrast environments are retained, reduces noise, and improves imaging quality and recognition accuracy. In order to optimize the point cloud accuracy, the shooting distance is set to 1 meter. In order to completely cover the central inner circle, the distance is increased to 1.5 meters, slightly sacrificing accuracy. Data preprocessing is carried out according to the shooting characteristics of the two parts of the data to ensure the accuracy of feature recognition. The flow chart is as follows Figure 2 shown.

[0047] 2D image brightness enhancement: During the data collection process, workshop conditions and environmental changes brought some problems: in 2D images, low brightness caused by natural light changes and production activities reduces image clarity and needs to be enhanced. We use adaptive block-wise γ-correction to enhance brightness and improve segmentation performance. Traditional γ-correction applies a uniform transformation across the entire image, which may over-enhance or compress details in certain areas, thereby reducing image quality. To solve this problem, we divide the image into blocks and adaptively adjust the γ value based on the average brightness of each block:

[0048]

[0049] In the formula, LU(p ij ) represents the brightness of the pixel, M and N are the length and width of the image after the block is divided. Calculate the average brightness and use 128 as the threshold: areas with brightness higher than 128 are considered darker, and a γ value less than 1 is applied to enhance the brightness. The smaller the γ value, the stronger the enhancement; areas below 128 are considered brighter, and a γ value greater than 1 is used to compress the brightness. The larger the γ value, the stronger the compression. Then use a lookup table on the image block to speed up the γ correction:

[0050]

[0051] k represents the current pixel, Represents the enhanced pixel brightness value, and performs inverse γ enhancement on the pixel.

[0052] 3D point cloud noise removal: The collected point cloud data has a high density, and the main body density of the object reaches 200 points / cm 3 , the noise distribution is characterized by a small number of points clustered. In order to reduce the noise of the point cloud, we use a statistical method based on the z-axis to denoise, divide the interval according to the length of the z-axis, and get the interval length h:

[0053]

[0054] Where l is the length of the z-axis interval occupied by the point cloud, z max is the maximum value of the z-axis data in the point cloud data, z min is the minimum value of the z-axis data in the point cloud data, and then record the number of points in each interval, using the interval with the most points as a reference:

[0055]

[0056] P j is the set of all points in the jth interval, P is the set of all points, and all points p in the interval i Satisfy the z-axis value z i In a size of The limits are set successively by extending to adjacent zero intervals:

[0057]

[0058] j lower is the lower limit of the z-axis size of the final selected valid point, j upper The upper limit of the z-axis size of the final valid points is selected, and the points outside these limits are regarded as noise and removed.

[0059] 2. Feature Extraction

[0060] In the initial positioning process, it is crucial to accurately extract the features of the aluminum coil. We propose a target instance segmentation network for key feature detection of aluminum coils, which includes two branches: 2D and 3D. The 2D branch is used to extract the head area of ​​the inner wall, and the 3D branch is used to extract the end face area of ​​the ring, including the inner and outer ring parts of the end face, and further classified into non-weldable areas and weldable areas.

[0061] Head region recognition: For the two-dimensional image input to the model, the brightness of the image with lower brightness is enhanced, and the brightness of low-brightness pixels is doubled using the block γ-correction method. The image is then resized to 3×640×640. The network model is based on the YOLOv8 model and outputs feature maps for prediction boxes, feature maps for prediction classification, and mask coefficients for segmentation. At the same time, the output of the P5 layer in the Head is upsampled to obtain the mask prototype. The network prediction output is filtered into two prediction boxes by NMS, and the target prediction mask is generated using the mask prototype and mask coefficients. The prediction box is cropped and thresholded to obtain the segmentation result.

[0062] End face area recognition: For the input 3D point cloud data, noise is removed from images with high noise levels. Since the number of points in the original image can usually reach 2 million, 100,000 points are randomly extracted from the original image as the input of CNN. The network model is based on the PointNet++ architecture, which has a hierarchical feature learning mechanism and is tailored for irregular point cloud data. The input is a set of 3D coordinates. First, a clustering algorithm is used in the sampling layer to select representative points from the input point cloud to form a smaller subset; then, for each sampling point, a set of points is selected in its neighborhood in the grouping layer, and the neighborhood is defined using a spherical or K-nearest neighbor search. Finally, the feature extraction layer applies a small PointNet to the points in each neighborhood to extract local features, which are processed by MLP (multi-layer perceptron) to generate a fixed-length feature vector. Through layer-by-layer aggregation, local features are gradually summarized to higher levels, similar to the feature extraction process from local to global in CNN, thereby obtaining a global feature vector. The fully connected layer further processes and classifies the extracted global feature vector, and finally outputs the classification label, which is divided into non-weldable areas and weldable areas. In practical applications, the point cloud data of three welding points around the aluminum coil have similar shapes and can be regarded as a single category through different rotations. After the input point cloud data is randomly downsampled, the segmentation prediction of each point is generated through the PointNet++ network.

[0063] 3. Path Planning

[0064] According to the result of feature extraction, the target area is further divided into paths, such as Figure 3According to actual production experience and demand, five welding areas are obtained on one side: one for the head part, i.e. Figure 3 The welding area 1 in the end face is obtained by the 2D branch; the inner circle part of the end face is one, that is, Figure 3 The welding area 2 in the end face has three outer ring parts, namely Figure 3 The welding areas 3-5 in the figure are all obtained by the 3D branch.

[0065] Leading area path planning: Based on the recognition results, such as Figure 3 The blue area at the head of the middle belt is used as the target area. The approximate center line is the actual weld path. Limited by the welding width and the penetration angle of the welding gun, the path is set to 100mm from the boundary and divided into arc areas with a length of 50mm and a width of 5mm.

[0066] End face area path planning: Based on the recognition results, such as Figure 3 Take the lower right result of the middle outer ring end face as an example, the blue area is the target area. The selection of the area here is limited by the flatness and compactness required for welding, which requires that there should be no misalignment and warping in the welding area.

[0067] In order to prevent the collision of the welding gun, we offset the welding area inward. The green part indicates the non-weldable area (including abnormal conditions such as depression, protrusion and warping), and the blue part indicates the weldable area. Both areas have obvious boundary lines. In actual application scenarios, the welding area division is close to the boundary line between the weldable and non-weldable areas, and the welding area division is carried out in the weldable area. Considering that the welding width of the welding gun is 5mm, the maximum thickness of the aluminum coil is 6mm, and the minimum requirement is to cover a layer of weld seam, this part of the area is divided according to the thickness of the plate and strip n (obtained by the transfer parameters of the aluminum coil rolling process). When n<4, the welding area moves 1mm radially inward from the boundary line, and when n>4, the welding area moves down from the boundary area (i.e. moves radially inward) 3mm. The welding area is a 5mm*50mm approximate rectangular arc. The welding area moves down according to the thickness of the plate and strip. The outer edge of the welding area is parallel to the outer edge of the weldable area everywhere, so it is an arc, and the inner edge is a straight line. Since the length of 50mm is extremely short compared to the entire aluminum coil, it can be approximated as an arc.

[0068] In summary, the present invention provides an advanced intelligent planning solution for the automated welding path of aluminum coils, which ensures the accuracy, reliability and safety of welding operations through meticulous technical innovation and practical verification, while improving production efficiency.

[0069] On the other hand, an embodiment of the present invention further provides an intelligent and standardized welding position recognition system for hot-rolled aluminum coils, comprising:

[0070] A processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an intelligent and standardized welding position identification method for hot-rolled aluminum coils as described in the above technical solution.

[0071] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. An intelligent and standardized welding position identification method for hot-rolled aluminum coils, characterized in that: include: Collect 2D images and 3D point cloud data of aluminum coils, enhance the brightness of 2D images with low light, and remove noise from 3D point clouds; the shooting area includes an end face head area and three outer ring end face peripheral areas; the three outer ring end face peripheral areas are located at the vertices of the regular triangle inscribed in the end face circle; Construct a target instance segmentation network for key feature detection of aluminum coils. The network includes two branches, 2D and 3D. The 2D branch is based on the YOLOv8 model and is used to extract a weld area on the inner wall of the 2D image. The 3D branch is based on the PointNet++ architecture and is used to extract the end face area of ​​the ring in the 3D point cloud, including three weld areas on the outer end face and one weld area on the inner end face. The three weld areas on the outer end face and one weld area on the inner end face are further classified into non-weldable areas and weldable areas. The welding area is further divided within the weldable area, and accurate path planning is performed on the welding point area of ​​the inner wall and the welding area based on the actual size of the welding gun.

2. The intelligent and standardized welding position identification method for hot-rolled aluminum coil according to claim 1, characterized in that: A two-stage exposure method is used to collect 2D images and 3D point cloud data, with the first stage being long exposure and the second stage being short exposure.

3. The intelligent and standardized welding position identification method for hot-rolled aluminum coil according to claim 1, characterized in that: Adaptive block γ-correction is used to enhance the brightness of 2D images. The specific implementation method is as follows: First, the image is divided into blocks, and the γ value is adaptively adjusted according to the average brightness of each block: In the formula, LU(p ij ) represents the brightness of the pixel, M and N are the length and width of the image after being divided into blocks; Calculate the average brightness and use 128 as the threshold: areas with brightness above 128 are considered darker, and a γ value less than 1 is applied to enhance the brightness. The smaller the γ value, the stronger the enhancement. Areas below 128 are considered brighter, and γ values ​​greater than 1 are used to compress the brightness. The larger the γ value, the stronger the compression. Then use a lookup table on the image block to speed up the γ correction: k represents the current pixel, Represents the enhanced pixel brightness value, and performs inverse γ enhancement on the pixel.

4. The intelligent and standardized welding position identification method for hot-rolled aluminum coil according to claim 1, characterized in that: The statistical method based on the z-axis is used to denoise the noise points. The interval is divided according to the length of the z-axis to obtain the interval length h: Among them, h is the length of the interval, l is the length of the z-axis interval occupied by the point cloud, and z max is the maximum value of the z-axis data in the point cloud data, z min is the minimum value of the z-axis data in the point cloud data, and then record the number of points in each interval, using the interval with the most points as a reference: P j is the set of all points in the jth interval, and all points p in the interval i Satisfy the z-axis value z i In a size of The limits are set successively by extending to adjacent zero intervals: j lower is the lower limit of the z-axis size of the final selected valid point, j upper The upper limit of the z-axis size of the final valid points is selected, and the points outside these limits are regarded as noise and removed.

5. The intelligent and standardized welding position identification method for hot-rolled aluminum coil according to claim 1, characterized in that: The input of the two-dimensional branch is a 2D image after brightness enhancement, and it outputs a feature map for predicting the box, a feature map for predicting the classification, and a mask coefficient for segmentation. At the same time, the output of the P5 layer in the Head is upsampled to obtain a mask prototype; the model output is filtered into two prediction boxes by NMS, and the mask prototype and mask coefficient are used to generate the target prediction mask. The prediction box is cropped and thresholded to obtain the segmentation result.

6. The intelligent and standardized welding position identification method for hot-rolled aluminum coil according to claim 1, characterized in that: The input of the 3D branch is a set of 3D coordinates. First, a clustering algorithm is used in the sampling layer to select representative points from the input point cloud to form a smaller subset. Then, for each sampling point, a group of points is selected in its neighborhood in the grouping layer, and the neighborhood is defined using spherical or K-nearest neighbor search. Finally, the feature extraction layer applies a small PointNet to the points in each neighborhood to extract local features. The local features are processed by MLP to generate a feature vector of fixed length. Through layer-by-layer aggregation, the local features are gradually summarized to a higher level to obtain a global feature vector. The fully connected layer further processes and classifies the extracted global feature vector, and finally outputs the classification label to divide the 3D point cloud into non-weldable areas and weldable areas.

7. The intelligent and standardized welding position identification method for hot-rolled aluminum coil according to claim 1, characterized in that: For a welding point area of ​​the inner wall, the center line is the actual weld path, and the actual weld path is an arc.

8. The intelligent and standardized welding position identification method for hot-rolled aluminum coil according to claim 1, characterized in that: The welding area is further divided within the weldable area according to the thickness n of the plate. When n is less than a certain threshold, the welding area moves radially inward from the boundary line of the weldable area by k1 mm. When n is greater than a certain threshold, the welding area moves radially inward from the boundary line of the weldable area by k2 mm. The outer edge of the welding area is parallel to the outer edge of the weldable area. k1 and k2 are constants.

9. The intelligent and standardized welding position identification method for hot-rolled aluminum coil according to claim 8, characterized in that: The thickness of the plate and strip is obtained by transferring parameters during the rolling process of the aluminum coil, and the welding area is a rectangular arc.

10. An intelligent and standardized welding position recognition system for hot-rolled aluminum coils, characterized in that: include: A processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute an intelligent and standardized welding position identification method for hot-rolled aluminum coils as described in any one of claims 1 to 9.

Images