A method and device for realizing effective riveting of plates of different specifications

Abstract

The present invention discloses a method and device for realizing effective riveting of plates of different specifications, belonging to the technical field of rapid mechanical cold connection. The method of the present invention accurately spatially locates the plate through a visual inspection device based on a deep learning lightweight recognition network, and a motor drives the clamping plate to clamp the relative position, so as to realize the riveting of plates of different thicknesses and widths. The self-piercing riveting device is provided with an intelligent clamp, and the intelligent clamp includes a fixing groove, a lead screw, a rectangular left baffle, a camera, a front convex baffle, a bottom plate, a motor, a positioning through hole, a rectangular right baffle, and a rear convex baffle; the new intelligent clamp device of the present invention realizes accurate positioning of the riveting position of the plate through a camera with a visual inspection function, so that the clamping device clamps quickly and stably, and the plate is easy to disassemble.

Description

Technical Field

[0001] The invention relates to a method and a device for realizing effective riveting of plates with different specifications, belonging to the technical field of rapid mechanical cold connection. Background Art

[0002] Self-piercing riveting technology is a thin plate material connection technology that has developed rapidly in recent years. It can effectively connect two or more layers of plates. Its principle is to rely on the mechanical interlock formed by the plastic deformation and rebound of the upper and lower plates and rivets during the riveting process to achieve effective connection of the plates. Usually, riveting requires the experimenter to hold the plate to operate, which will cause large riveting errors, uneven quality and potential safety hazards during operation. With the widespread application of self-piercing riveting in the automotive and aviation fields, the demand for connecting plates of different materials, different thicknesses, different lengths, and different widths has also increased. However, the riveting operation and plate quality in the existing technology are sometimes difficult to meet the work requirements, so a better intelligent fixture for self-piercing riveting equipment is needed. Summary of the invention

[0003] The present invention aims to solve the problems of large riveting errors and uneven sheet forming quality, and potential safety hazards in operation that are common in traditional self-piercing riveting. The present invention provides a method for realizing effective riveting of sheets of different specifications. The sheet is accurately spatially positioned by a visual inspection device based on a deep learning lightweight recognition network, and a motor drives the clamping plate to clamp the relative position. The riveting of sheets of different thicknesses and widths can be realized, and different punching and riveting positions can be quickly located. The method specifically includes the following steps:

[0004] Step 1: Before riveting, place the test piece in a bottom plate with a fixed groove, and the visual inspection equipment collects the image of the self-piercing riveted plate to obtain the marked riveting position.

[0005] Step 2: Use a bilinear interpolation algorithm to unify the size of the image to create a data set; the data set is divided into three parts: a training data set, a verification data set, and a test data set.

[0006] Step 3, using the YOLO-Mark toolkit to mark the images of the self-piercing riveting defect image dataset; the marking is used to record the boundary information of the riveting position on each image;

[0007] Step 4. After the labeling is completed, a lightweight multi-feature fusion self-piercing riveting point recognition network model is built based on the PyTorch deep learning framework;

[0008] Step 5. After the deep learning network model is built, the training parameters and training strategies are set; the training parameters mainly include the initial learning rate, learning rate decay coefficient, momentum coefficient, weight decay coefficient, number of training times and the number of images extracted for each training; the training strategy is an iterative training strategy, which uses an adaptive momentum estimation gradient descent (Adam) optimizer for training.

[0009] Step 6. After the training is completed, the image of the plate to be tested is identified; when the rivet position is identified in the image to be tested, the coordinates of the rivet position can be obtained through coordinate conversion, and the driving motor drives the clamping plate to clamp the relative position, and then rivet. When it is recognized that the position is wrong, the motor is started to move the plate until it reaches the riveting position, and then riveting is performed; after the riveting is completed, the riveting quality can also be evaluated in real time to determine whether the rivet is accurately riveted at the specified position of the plate (that is, whether there is a defect).

[0010] Preferably, the framework structure of the lightweight multi-feature fusion self-piercing riveting point recognition network model described in the present invention is mainly composed of a shallow network, a backbone network, a feature fusion module and a lightweight multi-head prediction model.

[0011] Preferably, the shallow network of the present invention is mainly composed of one 7×7 large convolutional layer; the backbone network is mainly composed of 13 SSB lightweight feature coding units and 3 GSB lightweight feature coding units.

[0012] Preferably, the backbone network of the present invention comprises three different scales, each scale comprising 3 SSB coding units and 1 GSB coding unit, 7 SSB coding units and 1 GSB coding unit, and 3 SSB coding units and 1 GSB coding unit, respectively; the proposed SSB lightweight feature coding unit consists of 2 1×1 convolutional layers and 1 3×3 convolutional layer, and the proposed GSB lightweight feature coding unit consists of 3 1×1 convolutional layers and 2 3×3 convolutional layers with a step size of 2.

[0013] Preferably, the feature fusion module of the present invention is mainly composed of three high-dimensional feature maps of different scales output from the backbone network; the lightweight multi-head prediction model is mainly composed of 6 convolutional layers.

[0014] Preferably, the 7×7 large convolution layer of the shallow network of the present invention is mainly used to preliminarily extract shallow original information of the image; the SSB lightweight feature coding unit and the GSB lightweight feature coding unit in the backbone network both reduce the parallelism of the network through a two-stage input structure, and set the number of input and output channels of the convolution layer to be the same to reduce memory access consumption during the convolution operation.

[0015] Preferably, the training steps of the lightweight multi-feature fusion self-piercing riveting point recognition network model of the present invention are as follows:

[0016] Step 5.1. Randomly select the number of images set by the training parameters in the training data set as samples for this training.

[0017] Step 5.2. Put each image in the training sample into the Adam optimizer in turn to update the weight parameters; specifically: the Adam optimizer randomly generates a set of initial weights and inputs them into the constructed deep learning network for forward propagation calculation and outputs a set of forward propagation parameters, and then uses the forward propagation parameters to perform back propagation calculation to obtain a new set of weight parameters and corresponding loss function values; the forward propagation and back propagation calculations are recorded as one iterative training; the new weight parameters will be used to iterate the old weight parameters; the loss function of the forward propagation calculation and back propagation calculation is the binary cross entropy (BECLoss) loss function.

[0018] Step 5.3: After completing one iterative training for all images in the training data set, the model is considered to have been trained once. Repeat Step 5.1 to Step 5.2 until the network model is trained for a fixed number of times as set by the training parameters in Step 5, and then terminate the training.

[0019] Step 5.4, after terminating the training of the network model, select the weight parameter with the smallest loss function value in the whole process as the optimal weight parameter for the identification of self-piercing riveting points.

[0020] Another object of the present invention is to provide a self-piercing riveting device, which is provided with an intelligent fixture, and the intelligent fixture includes a fixing groove 1, a screw 2, a rectangular left baffle 3, a camera 5, a front convex baffle 6, a base plate 7, a motor 9, a positioning through hole 10, a rectangular right baffle 11, and a rear convex baffle 12. The base plate 7 is provided with a fixing groove 1, and a positioning through hole 10 is provided in the middle of the fixing groove 1; a rectangular left baffle 3 and a rectangular right baffle 11 are provided on both sides of the fixing groove 1, and a front convex baffle 6 and a rear convex baffle 12 are provided at both ends of the fixing groove 1; the rectangular left baffle 3, the rectangular right baffle 11, the front convex baffle 6, and the rear convex baffle 12 are respectively connected to the motor 9 through the screw 2, and a camera 5 is provided on one side of the base plate 7, and the motor 9 and the camera 5 are respectively connected to the control device.

[0021] Preferably, the rectangular left baffle 3 , the rectangular right baffle 11 , the front convex baffle 6 , and the rear convex baffle 12 of the present invention are all placed on two parallel guide rails 8 , and the guide rails 8 are fixed on the bottom plate 7 .

[0022] Preferably, the camera 5 described in the present invention is fixed to one side of the base plate 7 through a cylindrical bracket 4; the cylindrical bracket and the visual detection equipment are used together for intelligent and accurate positioning, and the motor, the lead screw and the guide rail constitute a transmission system, so that the front and rear convex baffles and the left and right baffles are electrically opened and closed.

[0023] The present invention has the following technical effects:

[0024] The present invention utilizes visual detection and recognition functions to realize accurate positioning of the test pieces to be connected in space and position, and can detect and judge the molding quality of the riveted position after riveting, reduce errors, and further ensure the accuracy of the connection size and corresponding position of the plates; the motor, the lead screw and the guide rail are combined into a transmission system, which can adjust the distance between the front and rear boss baffles and the left and right baffles to realize effective riveting of plates of different lengths and widths; the middle grooved bottom plate can adjust the movable blocks of different thicknesses to achieve different height differences between the movable plate plane and the bottom fixed end plane, so as to realize effective riveting of plates of different thicknesses; the present invention has a simple structure, is easy to install, is intelligent, and improves operational safety and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the overall structure of the self-piercing riveting intelligent fixture;

[0026] Figure 2 Schematic diagram of lightweight coding unit structure;

[0027] Figure 3 Schematic diagram of the deep learning network model structure;

[0028] Figure 4 Schematic diagram of riveting point positioning detection results;

[0029] Figure 5 Schematic diagram of riveting quality judgment test results; a. Correct riveting position; b. Incorrect riveting position.

[0030] Figure 1 1. Bottom plate with fixed groove; 2. Lead screw; 3. Rectangular left baffle; 4. Cylindrical bracket; 5. Camera; 6. Front convex baffle; 7. Bottom plate; 8. Guide rail; 9. Motor; 10. Positioning through hole; 11. Rectangular right baffle; 12. Rear convex baffle. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited to the contents described above.

[0032] Example 1

[0033] A method for realizing effective riveting of plates of different specifications, the method can realize positioning and simultaneously detect whether the rivet is accurately riveted at a specified riveting point (correct identification and incorrect identification), specifically comprising the following steps:

[0034] Step 1. Before riveting, place the test piece in the bottom plate 1 with a fixing groove, and the visual inspection equipment collects the image of the self-piercing riveted plate to obtain the marked riveting position.

[0035] The visual inspection equipment is mainly used for self-made training data sets, and the training data sets are used to train deep learning network models; the information that needs to be collected in the image is mainly the image of the plate with markings on the workbench, and the plate image is accurately positioned within a range using a marker or marking pen to provide an accurate marking point for the final positioning; the coordinates of the rivet point in the image pixel coordinate system can be obtained through the trained deep learning network model, and then the coordinates of the rivet point in the world coordinate system can be obtained through coordinate conversion, and finally positioning is achieved.

[0036] Step 2, using a bilinear interpolation algorithm to unify the size of the image, for making a data set; the data set is divided into three parts: a training data set, a validation data set and a test data set;

[0037] Step 3, using the YOLO-Mark toolkit to mark the images of the self-piercing riveting defect image dataset; the marking is used to record the boundary information of the riveting position on each image;

[0038] Step 4. After the labeling is completed, a lightweight multi-feature fusion self-piercing riveting point recognition network model is built based on the PyTorch deep learning framework.

[0039] Since the computing resources of the riveting device are very limited and the motion control calculations need to be completed at the same time, it is difficult for the riveting device to run a complex visual recognition model; therefore, a lightweight design of the model is required to meet the use requirements of the device. Since the riveting area, size and working background of the plate (considering the complexity of the data set, interference factors, noise factors, etc.) have great differences, and the designed lightweight backbone network has limited feature extraction capabilities (simple structure or lightweight backbone will make it difficult for the model to learn and express complex scene information), a multi-feature fusion enhancement module is designed to improve the recognition accuracy, as follows.

[0040] The framework structure of the lightweight multi-feature fusion self-piercing riveting point recognition network model described in this embodiment is mainly composed of a shallow network, a backbone network, a feature fusion module and a lightweight multi-head prediction model; the shallow network is mainly composed of a 7×7 large convolution layer; the backbone network is mainly composed of 13 SSB lightweight feature coding units and 3 GSB lightweight feature coding units.

[0041] Specifically, the backbone network contains three different scales, each of which contains 3 SSB coding units and 1 GSB coding unit, 7 SSB coding units and 1 GSB coding unit, and 3 SSB coding units and 1 GSB coding unit; the SSB lightweight feature coding unit consists of 2 1×1 convolutional layers and 1 3×3 convolutional layer, and the GSB lightweight feature coding unit consists of 3 1×1 convolutional layers and 2 3×3 convolutional layers with a step size of 2. The feature fusion module is mainly composed of high-dimensional feature maps of three different scales output from the backbone network; the lightweight multi-head prediction model is mainly composed of 6 convolutional layers.

[0042] The 7×7 large convolution layer of the shallow network is mainly used to preliminarily extract the shallow original information of the image; the SSB lightweight feature coding unit and the GSB lightweight feature coding unit in the backbone network both reduce the parallelism of the network through a two-stage input structure, and set the number of input and output channels of the convolution layer to be the same to reduce memory access consumption during the convolution operation.

[0043] Among them, the output features are spliced ​​along the channel direction and then feature scrubbing is performed, which can maximize the channel information exchange between branches and make the distribution of output features more uniform and rich; the feature fusion module adjusts the size of the feature map through average pooling and bilinear interpolation algorithms, and then fuses multi-scale information along the channel dimension to enhance contextual information exchange and expand the model receptive field; the lightweight multi-head detection model is used to simplify the post-processing operations of the model; the lightweight design can improve the forward reasoning speed of the detection model, and the multi-head detector helps the model to identify targets of different sizes.

[0044] Step 5. After the lightweight multi-feature fusion self-piercing riveting point recognition network model is built, the training parameters and training strategies are set; the training parameters mainly include the initial learning rate, the learning rate decay coefficient, the momentum coefficient, the weight decay coefficient, the number of training times and the number of images extracted for each training; the training strategy is an iterative training strategy, which uses the adaptive momentum estimation gradient descent (Adam) optimizer for training. The specific steps are as follows:

[0045] Step 5.1. Randomly select the number of images set by the training parameters in the training data set as samples for this training.

[0046] Step 5.2: Put each image in the training sample described in step 5.1 into the Adam optimizer in turn to update the weight parameters.

[0047] The weight parameter update is specifically as follows: the Adam optimizer randomly generates a set of initial weights, which are input into the constructed deep learning network for forward propagation calculation and output a set of forward propagation parameters, and then the forward propagation parameters are used for back propagation calculation to obtain a new set of weight parameters and corresponding loss function values; the forward propagation and back propagation calculations are recorded as one iterative training; the new weight parameters will be used to iterate the old weight parameters; the loss function of the forward propagation calculation and back propagation calculation is the binary cross entropy (BECLoss) loss function.

[0048] Step 5.3: After completing one iterative training for all images in the training data set, the model is considered to have been trained once. Repeat Step 5.1 to Step 5.2 until the network model is trained for a fixed number of times as set by the training parameters in Step 5, and then terminate the training.

[0049] Step 5.4, after the network model training described in Step 5.3 is terminated, the weight parameter with the smallest loss function value in the whole process is selected as the optimal weight parameter, and the model will output the predicted result for the identification of self-piercing riveting points. The weight parameter is the parameter that needs to be updated each time the deep learning network model is trained; when the training is completed, a set of optimal training parameters, i.e., the training results (weights), can be obtained; the weights can be loaded to predict the marked area of ​​the input image, i.e., the identification of self-piercing riveting points.

[0050] Step 6. After the training is completed, the image of the plate to be tested is identified; when the rivet position is identified in the image to be tested, the coordinates of the rivet position can be obtained through coordinate conversion, and the driving motor drives the clamping plate to clamp the relative position, and then rivet. When it is recognized that the position is incorrect, the motor is started to move the plate until it reaches the riveting position, and then riveting is performed; after the riveting is completed, the riveting quality can also be evaluated in real time to determine whether the rivet is accurately riveted at the specified position of the plate (that is, whether there is a defect). The working status of the intelligent riveting can be monitored in real time through quality evaluation.

[0051] The detection algorithm described in this embodiment can simultaneously detect the riveting position (Location), determine whether it is riveted to the specified position (Location-Correct, Riveting-Fault), and is applied to the entire intelligent riveting process.

[0052] The method of the present invention is used to identify the marked area of ​​the plate in the image (i.e., the position for riveting); by identifying the marked area (riveting point), the coordinates of the area in the pixel coordinate system can be obtained, and by coordinate conversion, the coordinates of the area (riveting point) in the world coordinate system can be obtained, and finally, combined with the riveting equipment, accurate riveting can be achieved. When the recognition accuracy is higher than 95%, the motor is started to rivet, and when the recognition accuracy is lower than 95%, the motor is started and the riveting position is adjusted until the recognition accuracy is higher than 95%, and then riveting is performed. The recognition accuracy is, for example, Figure 4 shown.

[0053] After riveting is completed, the algorithm can be used to detect whether there are riveting defects; if the rivet is correctly riveted at the specified position, it is recognized that the riveting position is correct (such as Figure 5 a); If the riveting position is offset, it is identified as a riveting position error (such as Figure 5 b).

[0054] Example 2

[0055] The self-piercing riveting device of this embodiment is provided with an intelligent fixture, the structure of which is as follows: Figure 1 As shown, it includes a fixed slot 1, a screw 2, a rectangular left baffle 3, a camera 5, a front convex baffle 6, a bottom plate 7, a motor 9, a positioning through hole 10, a rectangular right baffle 11, and a rear convex baffle 12. The bottom plate 7 is provided with a fixed slot 1, and a positioning through hole 10 is provided in the middle of the fixed slot 1; a rectangular left baffle 3 and a rectangular right baffle 11 are provided on both sides of the fixed slot 1, and a front convex baffle 6 and a rear convex baffle 12 are provided at both ends of the fixed slot 1; the rectangular left baffle 3, the rectangular right baffle 11, the front convex baffle 6, and the rear convex baffle 12 are respectively connected to the motor 9 through the screw 2, a camera 5 is provided on one side of the bottom plate 7, and the motor 9 and the camera 5 are respectively connected to the control device.

[0056] The rectangular left baffle 3 , the rectangular right baffle 11 , the front convex baffle 6 , and the rear convex baffle 12 described in this embodiment are all placed on two parallel guide rails 8 , and the guide rails 8 are fixed on the bottom plate 7 .

[0057] The camera 5 described in this embodiment is fixed to one side of the base plate 7 through a cylindrical bracket 4; the cylindrical bracket and the visual detection equipment are used together for intelligent and accurate positioning, and the motor, screw and guide rail form a transmission system to enable the front and rear convex baffles and the left and right baffles to open and close electrically.

[0058] The specific use process of the device described in this embodiment is as follows: before riveting, the test piece is placed in the bottom plate with a fixing groove 1, and the camera 5 with a visual detection function locates the test piece at the corresponding riveting position, and the recognized photo is as follows: Figure 4The riveting position marked by the middle plate, the lead screw 2, the guide rail 8 and the motor 9 are combined into a transmission system, which is controlled by a control device such as a single-chip microcomputer that can generate a pulse signal, and a pulse signal is given to the stepper motor to adjust the distance between the front convex baffle 6, the rear convex baffle 12, the rectangular left baffle 3 and the rectangular right baffle 11. The distance between the front and rear baffles and the left and right baffles can be adjusted according to the required riveting width and thickness changes, so as to achieve riveting of plates of different widths and thicknesses. The guide rail 8 plays the role of a stable plate so that it can press the test piece to be connected; when the visual detection function camera 5 recognizes the plate to the corresponding position, riveting can be carried out. After the riveting process is completed, the transmission system drives the baffle to separate and take out the riveted plate.

Claims

1. A method for achieving effective riveting of plates of different specifications, characterized in that: The visual inspection equipment based on the deep learning lightweight recognition network accurately locates the plate in space, and the motor drives the clamping plate to clamp the relative position, which can realize the riveting of plates of different thickness and width. Specifically, the following steps are included: Step 1: Before riveting, place the sheet into the bottom plate with the fixing groove, and the visual inspection device collects the image of the self-piercing riveted sheet to obtain the marked riveting position; Step 2, using a bilinear interpolation algorithm to unify the size of the image, for making a data set; the data set is divided into three parts: a training data set, a validation data set and a test data set; Step 3, use the YOLO-Mark toolkit to mark the images of the data set; the marking is used to record the boundary information of the riveting position on each image; Step 4. After the labeling is completed, a lightweight multi-feature fusion self-piercing riveting point recognition network model is built based on the PyTorch deep learning framework; Step 5. After the lightweight multi-feature fusion self-piercing riveting point recognition network model based on the PyTorch deep learning framework is built, the training parameters and training strategies are set; the training parameters mainly include the initial learning rate, the learning rate decay coefficient, the momentum coefficient, the weight decay coefficient, the number of training times and the number of images extracted for each training; the training strategy is an iterative training strategy, which uses an adaptive momentum estimation gradient descent optimizer for training; Step 6. After the training is completed, the image of the plate to be tested is identified; when the rivet position is identified in the image to be tested, the coordinates of the rivet position can be obtained through coordinate conversion, and the driving motor drives the clamping plate to clamp the relative position, and then rivet. When the wrong position is identified, the motor is started to move the plate until it reaches the riveting position, and then riveting is performed; after the riveting is completed, the riveting quality is evaluated in real time to determine whether the rivet is accurately riveted at the specified position of the plate.

2. The method for achieving effective riveting of plates of different specifications according to claim 1 is characterized in that: The framework structure of the lightweight multi-feature fusion self-piercing riveting point recognition network model is mainly composed of a shallow network, a backbone network, a feature fusion module and a lightweight multi-head prediction model.

3. The method for achieving effective riveting of plates of different specifications according to claim 2 is characterized in that: The shallow network mainly consists of 1 The backbone network is mainly composed of 13 SSB lightweight feature coding units and 3 GSB lightweight feature coding units.

4. The method for achieving effective riveting of plates of different specifications according to claim 2 is characterized in that: The backbone network contains three different scales, each of which contains 3 SSB coding units and 1 GSB coding unit, 7 SSB coding units and 1 GSB coding unit, and 3 SSB coding units and 1 GSB coding unit; the proposed SSB lightweight feature coding unit consists of 2 Convolutional layer and 1 The proposed GSB lightweight feature encoding unit consists of three convolutional layers. Convolutional layer and 2 strides of 2 Convolutional layers.

5. The method for achieving effective riveting of plates of different specifications according to claim 2 is characterized in that: The feature fusion module is mainly composed of three high-dimensional feature maps of different scales output from the backbone network; the lightweight multi-head prediction model is mainly composed of 6 convolutional layers.

6. The method for achieving effective riveting of plates of different specifications according to claim 3 is characterized in that: 1 of the shallow network The large convolution layer is mainly used to preliminarily extract the shallow original information of the image; the SSB lightweight feature coding unit and the GSB lightweight feature coding unit in the backbone network both reduce the parallelism of the network through a two-stage input structure, and set the number of input and output channels of the convolution layer to be the same to reduce the memory access consumption during the convolution operation.

7. The method for achieving effective riveting of plates of different specifications according to claim 1 is characterized in that: The training steps of the lightweight multi-feature fusion self-piercing riveting point recognition network model are as follows: Step 5.1, randomly select the number of images set by the training parameters in the training data set as samples for this training; Step 5.2, each image in the training sample is sequentially placed into the Adam optimizer to update the weight parameters; specifically, the Adam optimizer randomly generates a set of initial weights and inputs them into the constructed self-piercing riveting point recognition network model to perform forward propagation calculation and output a set of forward propagation parameters, and then uses the forward propagation parameters to perform back propagation calculation to obtain a set of new weight parameters and corresponding loss function values; The forward propagation and back propagation calculations are recorded as one iterative training; the new weight parameters will be used to iterate the old weight parameters; the loss function of the forward propagation calculation and the back propagation calculation is a binary cross entropy loss function (BECLoss); Step 5.3, after completing one iteration of training for all images in the training data set, the model is considered to have been trained once, and steps Step 5.1 to Step 5.2 are repeated until the network model is trained for a fixed number of times set by the training parameters described in Step 5, and then the training is terminated; Step 5.4, after the training of the network model is terminated, the weight parameter with the smallest loss function value in the whole process is selected as the optimal weight parameter for the identification of self-piercing riveting points.

8. The self-piercing riveting device used in the method according to any one of claims 1 to 7, characterized in that: The self-piercing riveting device is provided with an intelligent fixture, which comprises a fixing groove (1), a lead screw (2), a rectangular left baffle (3), a camera (5), a front convex baffle (6), a bottom plate (7), a motor (9), a positioning through hole (10), a rectangular right baffle (11), and a rear convex baffle (12). The bottom plate (7) is provided with a fixing groove (1), and a positioning through hole (10) is provided in the middle of the fixing groove (1); a rectangular left baffle (3) and a rectangular right baffle (11) are provided on both sides of the fixing groove (1), and a front convex baffle (6) and a rear convex baffle (12) are provided at both ends of the fixing groove (1); the rectangular left baffle (3), the rectangular right baffle (11), the front convex baffle (6), and the rear convex baffle (12) are respectively connected to the motor (9) through the lead screw (2); a camera (5) is provided on one side of the bottom plate (7), and the motor (9) and the camera (5) are respectively connected to a control device.

9. The self-piercing riveting device according to claim 8, characterized in that: The rectangular left baffle (3), the rectangular right baffle (11), the front convex baffle (6), and the rear convex baffle (12) are all placed on two parallel guide rails (8), and the guide rails (8) are fixed on the bottom plate (7).

10. The self-piercing riveting device according to claim 8, characterized in that: The camera (5) is fixed to one side of the bottom plate (7) via a cylindrical bracket (4).

Images