Automatic welding system and method

Abstract

The present invention discloses an automatic welding system, including a sensor calibration module, a data acquisition and processing module, a welding position identification module, and a path planning module; it mainly performs semantic segmentation on a three-dimensional point cloud model including a workpiece to be welded through an improved RandLA-Net network, retaining the point cloud model of the workpiece body to be welded; performs plane segmentation on the three-dimensional point cloud model of the workpiece body to be welded through a RANSAC algorithm, and performs multi-plane alignment on the segmented plane to obtain a point cloud constituting the position to be welded. The motion path of the robot arm is planned according to the point cloud constituting the position to be welded, and finally automatic welding is achieved. The present invention also provides an automatic welding method. The present invention not only realizes the intelligence and automation of welding, but also effectively improves the efficiency and accuracy of welding position identification; at the same time, the present invention adds an attention mechanism module to the local feature aggregation module, effectively improving the recognition accuracy and reducing the situation of missegmentation.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence, and in particular relates to an automatic welding system and method. Background Art

[0002] In view of the current situation in my country, for the models to be welded, traditional manual welding is very inconvenient and consumes human resources. Most of my country's current robot welding systems are systems based on the OpenCV computer vision library. In order to achieve pre-welding guidance, it is necessary to first use the visual sensing system to identify the model and the position to be welded, determine the key point position of welding, establish the coordinates of the key points, send them to the robot, move the robot's mechanical arm to the welding starting point, and automatically complete the pre-welding guidance. The accuracy and recognition precision of the position to be welded directly affect the accuracy of the welding system.

[0003] Modeling the welded parts is particularly important for automatic welding systems. Currently, visual sensors are primarily used to identify the parts to be welded, which inevitably captures scenes other than the parts to be welded. Therefore, it is crucial to remove data unrelated to the parts to be welded during the modeling process. This directly impacts the accuracy and time required to identify the subsequent weld positions. Based on current solutions, existing automatic welding systems lack high welding accuracy and efficiency. Welding accuracy and efficiency directly impact the practicality of automatic welding systems. Improving welding accuracy and efficiency is a critical issue that must be addressed in order for automatic welding systems to completely replace traditional manual welding. Summary of the Invention

[0004] The present invention aims to solve the problems existing in the prior art and proposes an automatic welding system which improves welding precision and working efficiency.

[0005] To achieve the above-mentioned object, the present invention provides an automatic welding system, comprising a sensor calibration module, a data acquisition and processing module, a welding position identification module, and a path planning module;

[0006] Among them, the sensor calibration module is used to calibrate the sensor that collects data of the workpiece to be welded;

[0007] The data acquisition and processing module collects three-dimensional point cloud data of the workpiece to be welded to obtain a first three-dimensional point cloud, and pre-processes the first three-dimensional point cloud to obtain a second three-dimensional point cloud; reconstructs the second three-dimensional point cloud to obtain a three-dimensional point cloud model including the complete workpiece to be welded;

[0008] The welding position recognition module uses the RandLA-Net network combined with the RandLA-Net semantic segmentation algorithm to perform category recognition on the 3D point cloud model of the complete workpiece to be welded, obtained by the data acquisition and processing module, to obtain a 3D point cloud model of the workpiece to be welded. The RANSAC algorithm is then used to perform plane segmentation on the 3D point cloud model of the workpiece to be welded, and the segmented planes are subjected to multi-plane registration to obtain the point cloud that constitutes the welding position.

[0009] The path planning module generates a welding path based on the point cloud constituting the welding position obtained by the welding position recognition module.

[0010] Furthermore, the system also includes a remote monitoring module. The outputs from the sensor calibration module, data acquisition and processing module, welding position identification module, and path planning module are all sent to the remote monitoring module. This module monitors the welding position identification and path planning processes of the robotic arm in real time. This allows users to monitor welding progress in real time, effectively preventing welding errors.

[0011] The present invention also provides an automatic welding method based on the above automatic welding system, comprising the following steps:

[0012] Step 1: Calibrate the sensor;

[0013] Step 2: Use the sensor to collect 3D point cloud data of the workpiece to be welded to obtain the first 3D point cloud;

[0014] Step 3: Preprocess the first three-dimensional point cloud to obtain a second three-dimensional point cloud;

[0015] Step 4: Reconstruct the second 3D point cloud to obtain a 3D point cloud model including the complete workpiece to be welded;

[0016] Step 5: Based on the RandLA-Net network and the RandLA-Net semantic segmentation algorithm, the 3D point cloud model of the complete workpiece to be welded obtained by the data acquisition and processing module is classified to obtain a 3D point cloud model of the workpiece to be welded;

[0017] Step 6: Use the RANSAC algorithm to perform plane segmentation on the 3D point cloud model of the workpiece to be welded, and perform multi-plane registration on the segmented planes to obtain the point cloud that constitutes the welding position;

[0018] Step 7: Perform path planning based on the point cloud of the welding position obtained in step 6.

[0019] Furthermore, in step 6, after plane segmentation of the three-dimensional point cloud model of the workpiece to be welded is performed using the RANSAC algorithm, equations of each plane of the workpiece to be welded are obtained, and the intersection lines of the obtained plane equations are calculated. The points on each plane are projected onto the corresponding intersection lines, and the points obtained by projecting onto the intersection lines are the points constituting the welding position, thereby obtaining a point cloud constituting the welding position. In this way, the point cloud constituting the welding position can be quickly found on the intersection lines of the planes.

[0020] Furthermore, an attention mechanism module is added to the local feature aggregation module of the RandLA-Net network. This attention mechanism module includes a channel attention submodule and a spatial attention submodule. The input features are sequentially passed through the channel attention submodule and the spatial attention submodule to infer an attention map. The input feature map, channel attention map, and spatial attention map are then multiplied in sequence to perform adaptive feature fusion. This method gradually increases the receptive field of each point, allowing it to learn and perceive complex geometric spatial structures, thereby preserving local features and geometric details. While using random sampling to increase speed, it also ensures comprehensive information. This improves the accuracy of the reconstructed model of the part to be welded.

[0021] Furthermore, the channel attention submodule performs mean pooling and maximum pooling on the input point cloud features F respectively; then fully connects the feature channels, and the number of output channels is C / r, where C is the number of channels, that is, the feature dimension of the input; r is the decay rate; then fully connects again to restore the output channels to C; finally, the features after mean pooling and the features after maximum pooling are added, and then the output is obtained by activating the sigmoid function to obtain the channel attention map M c , the final channel attention map M c Multiply it with the input point cloud feature F and perform the weighted attention mechanism to get F'.

[0022] Furthermore, the spatial attention submodule takes the weighted feature F' after the channel attention submodule as input, concatenates the two features after maximum pooling and mean pooling, and then passes the activation function sigmoid to output the spatial attention map M. s , the final spatial attention map M s Multiply it with the input feature F' and perform attention mechanism weighting.

[0023] Beneficial effects: Compared with the existing technology, the present invention not only realizes the intelligence and automation of welding, but also effectively improves the efficiency and accuracy of welding position recognition; at the same time, the present invention improves the RandLA-Net network and adds an attention mechanism module to the local feature aggregation module, which effectively improves the recognition accuracy and reduces the occurrence of missegmentation; it can also improve the overall accuracy while accelerating the learning speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic structural diagram of the automatic welding system provided by the present invention;

[0025] Figure 2 A schematic flow chart of the automatic welding method provided by the present invention;

[0026] Figure 3 A schematic diagram of the RandLA-Net network structure used in the present invention;

[0027] Figure 4 A schematic diagram of the LFA module structure in the RandLA-Net network used in the present invention;

[0028] Figure 5 Schematic diagram of the attention mechanism module. DETAILED DESCRIPTION

[0029] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the examples of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0030] Example 1:

[0031] like Figure 1 As shown, this embodiment discloses an automatic welding system, which mainly includes a sensor calibration module, a data acquisition and processing module, a welding position recognition module, a path planning module and a remote monitoring module.

[0032] The sensor calibration module primarily calibrates the sensors used in the system, which is essential for precise welding. In this embodiment, a structured light 3D camera is primarily used to collect data from the workpiece being welded, so the sensor calibration module primarily calibrates the structured light 3D camera. Alternatively, a lidar sensor could be used as the data acquisition sensor in this system.

[0033] The data acquisition and processing module mainly preprocesses the first three-dimensional point cloud collected by the structured light 3D camera to obtain a second three-dimensional point cloud, and reconstructs the preprocessed second three-dimensional point cloud to obtain a three-dimensional point cloud model that includes a complete workpiece to be welded.

[0034] The welding position recognition module first uses the RandLA-Net semantic segmentation algorithm to classify the three-dimensional point cloud model of the complete workpiece to be welded obtained by the data acquisition and processing module, and deletes other point clouds outside the point cloud of the workpiece to be welded, thereby obtaining a three-dimensional point cloud model of the workpiece to be welded; then the three-dimensional point cloud model of the workpiece to be welded is plane segmented by the RANSAC algorithm. After plane segmentation, the equations of each plane of the workpiece to be welded are obtained, and the intersection line of the obtained plane equations is calculated. The intersection line is the equation at the welding position of the three-dimensional point cloud data. Because the plane equation to be calculated is infinite, the intersection line obtained for the plane is also infinitely long. At this time, the point cloud data on each plane after the plane segmentation of the model to be welded is projected onto the intersection line corresponding to each plane by giving the coefficient of the intersection line equation and using the projection filtering algorithm in the point cloud pcl library. The set of points projected onto the intersection line is the welding position, thereby obtaining the point cloud that constitutes the welding position; this can solve the problem of infinite length of the intersection line.

[0035] The path planning module performs welding path planning based on the improved Q-Learning algorithm for the point set constituting the welding position obtained by the welding position recognition module.

[0036] In the remote monitoring module, a Qt-based visualization interface was developed to acquire, process, and transmit multi-source sensor data. This allows for real-time monitoring of welding position identification and path planning, preventing welding errors. The visualization interface consists of a real-time data transmission status display window, a real-time welding interface display window, and stop and resume buttons. The remote monitoring module allows real-time visibility of data transmission status and displays the welding interface, facilitating operator assessment of welding accuracy and enabling the system to be stopped and resumed.

[0037] like Figure 2 As shown, the automatic welding method based on the above automatic welding system specifically includes the following steps:

[0038] Step 1: Calibrate the structured light 3D camera; the specific calibration method is:

[0039] Step 1-1: Place the calibration plate at any position within the measured range, use the camera to shoot the calibration plate, detect the corner points of the calibration plate, and record the coordinates of each corner point in the imaging plane (u c , v c );

[0040] Step 1-2: Keep the calibration plate still and use a projector to project the sinusoidal fringe pattern onto the calibration plate. Obtain the phase diagram at this time and record the phase of each corner point.

[0041] Step 1-3: Adjust the placement of the calibration plate multiple times and repeat steps 1-1 and 1-2 to obtain multiple sets of corner point coordinates and corresponding phases;

[0042] Step 1-4: Calibrate the camera using Zhang Zhengyou's calibration method to obtain the camera's intrinsic parameter A c , and the external parameters [R, t] corresponding to each calibration plate image, the coordinates (X c ,Y c ,Z c ),

[0043] in (X w ,Y w ,Z w ) are the coordinates of the corner point in the world coordinate system;

[0044] Step 1-5: Use the phase of the calibration plate corner point position obtained in step 1-3 and the coordinates of the calibration plate corner point in the camera coordinate system obtained in step 1-4 to construct a polynomial By using the least square method, eight parameters a1~a8, three-dimensional coordinates (X c ,Y c ,Z c ) and any point (u c , v c ) and the phase corresponding to that point The relationship is shown in formula (1), where f cx , f cy , c cx , c cy All are camera internal parameters.

[0045]

[0046] Step 2: Use a structured light 3D camera to collect 3D point cloud data of the workpiece to be welded to obtain the first 3D point cloud;

[0047] Step 3: Preprocess the first three-dimensional point cloud to obtain a second three-dimensional point cloud; the preprocessing includes point cloud denoising, point cloud outlier removal, and point cloud smoothing.

[0048] Step 4: Reconstruct the second three-dimensional point cloud to obtain a three-dimensional point cloud model including the complete workpiece to be welded.

[0049] Step 4, based on the Iterative Closest Point (ICP) point cloud registration algorithm, stitches and fuses point cloud data from different angles and positions to create a three-dimensional point cloud model of the workpiece to be welded. ICP is currently the most widely used and effective point cloud registration method. It transforms point clouds of the same target in all different coordinate systems into a common reference coordinate system through common feature stitching, forming a complete point cloud model of the target object. After fusing the overlapping areas between the point clouds into a smooth surface, a complete point cloud model is obtained, enabling three-dimensional reconstruction of the welding target and determining the relative positional relationship between the camera and each point in the point cloud. The ICP algorithm is essentially an optimal matching algorithm based on the least squares method. It repeatedly searches for corresponding points and calculates the optimal rigid body transformation until a convergence criterion indicating a correct match is met, at which point the iteration terminates. This results in a more accurate three-dimensional point cloud model of the workpiece to be welded, enabling the precise location of the weld to be determined, thereby accurately completing the automatic welding process. This effectively reduces the error rate of automatic welding.

[0050] Step 5: The 3D point cloud model of the complete workpiece to be welded, obtained in Step 4, is input into the trained RandLA-Net network for recognition. The RandLA-Net semantic segmentation algorithm is used to classify the 3D point cloud model of the complete workpiece to be welded, thereby obtaining a complete 3D point cloud of the workpiece to be welded and a background point cloud. The background point cloud refers to the point cloud of objects other than the workpiece to be welded, such as the welding table on which the workpiece to be welded is placed. All other point clouds besides the point cloud of the workpiece to be welded are deleted to obtain a 3D point cloud model of the workpiece to be welded.

[0051] like Figure 3 As shown in the figure, the RandLA-Net network used in this embodiment, where N represents the number of point clouds, d in Indicates the input dimension, n classRepresenting the number of output categories, the network stacks multiple local feature aggregation (LFA) modules and random sampling layers (RS), following the widely used encoder-decoder architecture with skip connections. The input point cloud is first fed into a multilayer perceptron (MLP) to extract features for each point. Four encoding and decoding layers are then used to learn the features of each point. Each encoding layer consists of an LFA module and a RS. The LFA module increases the feature dimensionality of the points to retain more information, while random downsampling reduces the number of points. For each decoder layer, nearest neighbor interpolation is used to upsample the points to improve efficiency, while skip connections are used to stack the decoded features with the corresponding features from the encoding stage. Finally, three fully connected layers (FC) and a dropout layer (DP) are used to predict the semantic label of each point. In this embodiment, the network is primarily used to classify all point clouds in the input 3D point cloud model of the complete workpiece to be welded into two categories: one is the point cloud that constitutes the workpiece to be welded, and the other is the background point cloud excluding the workpiece to be welded. This results in a point cloud that constitutes the workpiece to be welded.

[0052] The RandLA-Net network is trained by inputting 3D point cloud models containing different types of welded workpieces into the network for training. During the training process, the input 3D point cloud models of the welded workpieces are semantically segmented using the RandLA-Net semantic segmentation algorithm. Local regions with identical features within the 3D point cloud models of the welded workpieces are clustered together, thereby dividing the input 3D point cloud model containing the welded workpieces into the welded workpiece body region and the background region, where the background region is the region other than the welded workpiece body. To ensure that the trained RandLA-Net network can accurately output 3D point cloud models of welded workpieces, the training datasets used include 3D point cloud models of single-line workpieces at different angles, 3D point cloud models of right-angle workpieces, 3D point cloud models of three-intersection workpieces, 3D point cloud models of long straight workpieces, 3D point cloud models of V-shaped workpieces, and 3D point cloud models of overlapped workpieces.

[0053] In order to increase the receptive field of each point, learn and perceive complex geometric spatial structures, thereby retaining local features and geometric details, and ensuring information comprehensiveness while using random sampling to speed up, this embodiment adds an attention mechanism module to the LFA module.

[0054] like Figure 4 As shown in FIG, this embodiment mainly adds an attention mechanism module to the attention pooling module and the dilated residual block in the LFA module. Indicates concat splicing, Indicates sum, represents the Softmax operation, ⊙ represents the multiplication operation, Represents the K nearest neighbor points. N represents the number of input point clouds, d represents the feature dimension of the relevant points, and d in represents the feature dimension of the input point, K represents the number of k nearest neighbors, d out represents the output point feature dimension, and d' represents the attention pooling output dimension.

[0055] like Figure 5 As shown, the attention mechanism module includes a channel attention submodule and a spatial attention submodule.

[0056] Among them, the channel attention submodule mainly focuses on which channels contribute to the final output of the network. The input of this submodule is the feature F of the point cloud data obtained by the data acquisition and processing module. First, the point cloud features are mean pooled and maximum pooled respectively, and then the feature channels are fully connected. In order to reduce parameters, the number of output channels here is C / r, where C is the number of channels, that is, the feature dimension of the input; r is the decay rate, and the r value in this embodiment is 8. The next step is to use full connection to restore the output channel to C. Finally, the features after mean pooling and the features after maximum pooling are added, and then the channel attention map M is obtained after the output through the activation function sigmoid. c , the final channel attention map M c Multiply it with the initial input feature F and perform the weighted attention mechanism to get F'.

[0057] The spatial attention submodule focuses on which positions contribute to the final output of the network. This submodule takes the weighted feature F' after the channel attention submodule as input, concatenates the two features after maximum pooling and mean pooling, and then passes the activation function sigmoid to output the spatial attention map M. s , the final spatial attention map M s Multiply it with the input feature F' and perform attention mechanism weighting. For a feature F, the attention mechanism module will sequentially infer the channel attention map M c and spatial attention map M s , the final output feature is

[0058] Step 6: Use the RANSAC algorithm to perform plane segmentation on the 3D point cloud model of the workpiece to be welded, and perform multi-plane registration on the segmented planes to obtain the point cloud that constitutes the welding position;

[0059] Among them, after the plane is segmented, the equations of each plane of the workpiece to be welded will be obtained, and the intersection line of the obtained plane equations will be calculated. The intersection line is the equation of the welding position of the three-dimensional point cloud data. Because the plane equation to be obtained is infinite, the intersection line of the plane is also infinitely long. At this time, the point cloud data on each plane after the plane of the model to be welded is segmented, and the coefficient of the intersection equation is given. The projection filtering algorithm in the point cloud PCL library is used to project the point cloud data on the intersection line corresponding to each plane. The set of points projected onto the intersection line is the position that needs to be welded, thereby obtaining the point cloud that constitutes the welding position; this can solve the problem of infinite length of the intersection line.

[0060] Step 7: Based on the improved Q-Learning, the paths between adjacent points in the point cloud that make up the welding position are planned separately, and the paths between all points are connected to obtain the welding path. The improved algorithm adds a layer of learning process on the basis of the original algorithm, performs deep learning on the environment, converges faster, reduces the number of learning times, and improves efficiency. The specific method is as follows: The Q update rule in the improved Q-Learning algorithm is:

[0061] Q′(s,a)=r(s,a)+γ(αmaxQ(s′,a′)+(1-α)maxQ(s″,a″))

[0062] Where Q′(s,a) represents the Q value corresponding to the robot arm performing action a in state s; r(s,a) represents the reward obtained by the robot arm from state s performing action a; γ is called the reduction factor; s′ represents the state of the robot arm one step after state s; α represents the dynamic search factor, Q(s′,a′) represents the Q value corresponding to the robot arm performing action a′ from state s′, s″ represents the state of the robot arm two steps after state s, and Q(s″,a″) represents the Q value corresponding to the robot arm performing action a″ from state s″. The dynamic search factor α introduced here is used to ensure the convergence of the Q value. The updated learning rule mainly uses the dynamic search factor α to weigh the rewards obtained in step 1 and step 2. The larger α is, the greater the weight given to the reward of step 1. When α = 1, the update rule is determined only by step 1, which is consistent with the update rule of the classic Q-Learning algorithm.

[0063] The specific steps include:

[0064] Step 7-1: Initialize the Q value table; construct a Q value table with n rows and n columns (n ​​is the number of states), and initialize all values ​​in the table to zero;

[0065] Step 7-2: Select the next action a based on the current Q-value table. Initially, all Q-values ​​are zero, and the next action is randomly selected. As the number of iterations increases, the Q-value table is continuously updated, selecting the action with the highest reward, essentially finding the optimal path. Action a is determined by the direction of the robot's motion.

[0066] Step 7-3: Calculate the action reward. After taking action a, update the Q value of the previous state based on the current state and reward using the following equation:

[0067] Q′(s,a)=r(s,a)+γ(αmaxQ(s′,a′)+(1-α)maxQ(s″,a″));

[0068] Among them, maxQ(s′,a′) represents the maximum Q value of the robot arm in state s′ among all actions; maxQ(s″,a″) represents the maximum Q value of the robot arm in state s′ among all actions;

[0069] Step 7-4: Repeat step 7-3 until the iteration ends and the final Q value table is obtained;

[0070] Step 7-5: Select the best path based on the Q value table.

[0071] The dynamic search factor α is dynamically adjusted based on environmental feedback. If an exploration from the starting point to the end point fails, the learning factor is increased to increase the randomness of the next exploration to avoid falling into the previous local optimization dilemma. Conversely, the learning factor is reduced to increase the purposefulness. A larger dynamic search factor increases the probability of random search actions, which to some extent avoids falling into local optimization.

[0072] During path planning, the action taken from one state to the next is controlled by a policy, which maps states to actions. When obstacles or traps exist in the environment, the policy must ensure that the next action does not encounter obstacles or traps. The reward represents the probability of making the correct decision in state s. A higher probability indicates a higher reward, meaning the closer the robot is to the ideal path. Continuous exploration changes the state to achieve the optimal path. Each step forward by the robot arm represents a new state, but it does not necessarily mean it is nearing the target. Therefore, the state update triggered by each action must be evaluated and accumulated into the previous reward. After exploring state s and all possible actions a at a given moment, the optimal action is selected based on the environmental feedback to enter the next state s, and this process repeats until the destination is reached. Given the point cloud coordinates of the welding location, the robot arm uses the Q-Learning algorithm to explore all possible paths within a specific state. Each step forward seeks a local optimal solution within the current state until it reaches the destination.

Claims

1. An automatic welding system, characterized in that: It includes sensor calibration module, data acquisition and processing module, welding position identification module, and path planning module; Among them, the sensor calibration module is used to calibrate the sensor that collects data of the workpiece to be welded; The data acquisition and processing module collects three-dimensional point cloud data of the workpiece to be welded to obtain a first three-dimensional point cloud, and pre-processes the first three-dimensional point cloud to obtain a second three-dimensional point cloud; reconstructs the second three-dimensional point cloud to obtain a three-dimensional point cloud model including the complete workpiece to be welded; The welding position recognition module uses the RandLA-Net network combined with the RandLA-Net semantic segmentation algorithm to perform category recognition on the 3D point cloud model of the complete workpiece to be welded, obtained by the data acquisition and processing module, to obtain a 3D point cloud model of the workpiece to be welded. The RANSAC algorithm is then used to perform plane segmentation on the 3D point cloud model of the workpiece to be welded, and the segmented planes are subjected to multi-plane registration to obtain the point cloud that constitutes the welding position. The path planning module generates a welding path based on the point cloud constituting the welding position obtained by the welding position recognition module; An attention mechanism module is added to the local feature aggregation module of the RandLA-Net network; the attention mechanism module includes a channel attention submodule and a spatial attention submodule; the input features are sequentially passed through the channel attention submodule and the spatial attention submodule to infer the attention map, and then the input feature map, the channel attention map and the spatial attention map are sequentially multiplied to perform adaptive feature fusion; The channel attention submodule performs mean pooling and maximum pooling on the input point cloud features F respectively; then fully connects the feature channels, and the output channel number is C / r, where C is the number of channels, that is, the feature dimension of the input; r is the decay rate; then fully connects again to restore the output channel to C; finally, the features after mean pooling and the features after maximum pooling are added, and then the output is obtained by activating the sigmoid function to obtain the channel attention map M c , the final channel attention map M c Multiply it with the input point cloud feature F and perform the weighted attention mechanism to get F'. The spatial attention submodule takes the weighted feature F' after the channel attention submodule as input, concatenates the two features after maximum pooling and mean pooling, and then passes the activation function sigmoid to output the spatial attention map M. s , the final spatial attention map M s Multiply it with the input feature F' and perform attention mechanism weighting; Among them, the training method of the RandLA-Net network is: the collected 3D point cloud model containing different types of welding workpieces is input into the RandLA-Net network for training, and during the training process, the input 3D point cloud model of the welding workpiece is semantically segmented by the RandLA-Net semantic segmentation algorithm, and the local areas with the same features in the 3D point cloud model of the welding workpiece are clustered, so as to divide the input 3D point cloud model including the welding workpiece into the welding workpiece body area and the background area, wherein the background area is other areas except the welding workpiece body; the data sets used in the training process include single straight line workpiece 3D point cloud models of different angles, right-angle workpiece 3D point cloud models, three-intersection line workpiece 3D point cloud models, long straight workpiece 3D point cloud models, V-shaped workpiece 3D point cloud models, and overlapped workpiece 3D point cloud models.

2. The automatic welding system according to claim 1, characterized in that: It also includes a remote monitoring module. The output results of the sensor calibration module, data acquisition and processing module, welding position identification module and path planning module are respectively sent to the remote monitoring module. The remote monitoring module monitors the process of welding position identification and path planning of the robotic arm in real time.

3. An automatic welding method based on the automatic welding system according to claim 1, characterized in that: The following steps are involved: Step 1: Calibrate the sensor; Step 2: Use the sensor to collect 3D point cloud data of the workpiece to be welded to obtain the first 3D point cloud; Step 3: Preprocess the first three-dimensional point cloud to obtain a second three-dimensional point cloud; Step 4: Reconstruct the second 3D point cloud to obtain a 3D point cloud model including the complete workpiece to be welded; Step 5: Based on the RandLA-Net network and the RandLA-Net semantic segmentation algorithm, the 3D point cloud model of the complete workpiece to be welded obtained by the data acquisition and processing module is classified to obtain a 3D point cloud model of the workpiece to be welded; Step 6: Use the RANSAC algorithm to perform plane segmentation on the 3D point cloud model of the workpiece to be welded, and perform multi-plane registration on the segmented planes to obtain the point cloud that constitutes the welding position; Step 7: Perform path planning based on the point cloud of the welding position obtained in step 6; Add an attention mechanism module to the local feature aggregation module of the RandLA-Net network; the attention mechanism module includes a channel attention submodule and a spatial attention submodule; The input features are sequentially passed through the channel attention submodule and the spatial attention submodule to infer the attention map, and then the input feature map, channel attention map and spatial attention map are multiplied in sequence for adaptive feature fusion; The channel attention submodule performs mean pooling and maximum pooling on the input point cloud features F respectively; then fully connects the feature channels, and the output channel number is C / r, where C is the number of channels, that is, the feature dimension of the input; r is the decay rate; then fully connects again to restore the output channel to C; finally, the features after mean pooling and the features after maximum pooling are added, and then the output is obtained by activating the sigmoid function to obtain the channel attention map M c , the final channel attention map M c Multiply it with the input point cloud feature F and perform the weighted attention mechanism to get F'. The spatial attention submodule takes the weighted feature F' after the channel attention submodule as input, concatenates the two features after maximum pooling and mean pooling, and then passes the activation function sigmoid to output the spatial attention map M. s , the final spatial attention map M s Multiply it with the input feature F' and perform attention mechanism weighting; Among them, the training method of the RandLA-Net network is: the collected 3D point cloud model containing different types of welding workpieces is input into the RandLA-Net network for training, and during the training process, the input 3D point cloud model of the welding workpiece is semantically segmented by the RandLA-Net semantic segmentation algorithm, and the local areas with the same features in the 3D point cloud model of the welding workpiece are clustered, so as to divide the input 3D point cloud model including the welding workpiece into the welding workpiece body area and the background area, wherein the background area is other areas except the welding workpiece body; the data sets used in the training process include single straight line workpiece 3D point cloud models of different angles, right-angle workpiece 3D point cloud models, three-intersection line workpiece 3D point cloud models, long straight workpiece 3D point cloud models, V-shaped workpiece 3D point cloud models, and overlapped workpiece 3D point cloud models.

4. The automatic welding method according to claim 3, characterized in that: In step 6, after the three-dimensional point cloud model of the workpiece to be welded is plane segmented by the RANSAC algorithm, the equations of each plane of the workpiece to be welded are obtained, the intersection lines of the obtained plane equations are calculated, and the points on each plane are projected onto the corresponding intersection lines. The points projected onto the intersection lines are the points that constitute the welding positions, thereby obtaining the point cloud that constitutes the welding positions.

Images