CN112589232B - A welding seam tracking method and device based on independent rectification deep learning - Google Patents

Abstract

The invention relates to a welding seam tracking method and device based on independent rectification type deep learning. The device comprises: a working platform, a controller, a vertical guide rail, a horizontal guide rail, a welding torch fixture, a line structured light vision sensor, and a servo motor ; The method includes: S1: collecting the structured light picture of the welding seam; S2: reading the structured light picture of the welding seam; S3: identifying the current welding seam type by using YOLOV3 and locating the center position of the welding seam; S4: positioning the welding seam with S3 The center position is used as the tracking target, and the KCF tracker is initialized; S5: The welding starts, and the KCF‑YOLOV3 algorithm is used to track the welding seam in real time; until the welding ends. The welding device and the welding method of the invention can automatically identify various welding seams and accurately locate the central area of the welding seam, which can effectively improve the welding seam detection speed and recognition accuracy, thereby improving the welding efficiency.

Description

A welding seam tracking method and device based on independent rectification type deep learning

technical field

The invention relates to the technical field of welding, and more particularly, to a welding seam tracking method and device based on independent deviation correction type deep learning.

Background technique

The welding seam tracking technology based on structured light vision combines the advantages of computer vision and laser 3D measurement technology, and has excellent characteristics such as simple data information collection, obvious welding seam characteristics, and strong anti-interference ability. It has the advantages of stability, coherence and ability concentration, so it is widely used in the light source of structured light. At present, the welding seam tracking under the line structured light vision sensing is relatively mature. In order to reduce the interference of strong arc light in the welding process, a filter lens with a certain bandwidth is usually installed in front of the camera lens to keep a clear structured light image, and the center position of the weld is obtained through signal acquisition and image processing.

The current solution for welding seam tracking is basically the integration of welding equipment and tracking device. The disadvantage of this technical solution is that a complex communication protocol needs to be formulated between the tracking device and the welding equipment, and the transmission of sensing signals and control signals needs to be considered. The problem is that the software is difficult to write, the product development cycle is long, and it can only be applied to specific industrial scenarios.

The existing welding seam tracking system based on visual sensing is basically composed of a computer and a visual sensor. In the host computer software, the computer displays the image collected by the visual sensor in real time, and obtains the center position of the welding seam through a specific image processing algorithm, and then calculates the welding seam. seam deviation, and finally issue a control command. Although the performance of the computer is powerful, its shortcomings cannot be ignored: it is large in size and occupies a lot of space, which is not suitable for application in some industrial fields; the cost is high, and the cost performance is not high; the installation is inflexible and inconvenient. Therefore, the computer is not suitable for large-scale industrial application on the job site.

In the welding production process, according to the different groove forms and connection methods, there are mainly the following types of welds: flat butt welds, lap welds, V-shaped welds, fillet welds, and annular welds. The structured light fringe image of each type of weld is different, so it is necessary to develop a specific image feature extraction algorithm according to the type of weld, and at the same time, parameters such as welding speed, welding current and voltage should also be adjusted according to the type of weld in automatic robot welding. . However, conventional seam tracking systems require manual input of the seam type before welding, which severely reduces the automation level of the welding robot. With the significant improvement of computer hardware performance and the large increase of data samples, deep networks based on convolutional neural networks (CNN) have become the mainstream of object detection technology. At present, the target detection technology based on deep learning can simultaneously achieve real-time target classification and target positioning, and achieve good results in terms of accuracy and anti-interference. Applying this technology to weld seam tracking can realize weld type identification and weld center positioning before welding, real-time weld detection during welding, and continuous correction of the tracking results.

In recent years, target tracking algorithm has been applied to weld seam tracking, and many scholars at home and abroad have conducted in-depth research on weld seam tracking based on target tracking algorithm. The target tracking algorithm is divided into two categories: generative model and discriminative model. The generative model describes the characteristics of the tracking target by establishing a model, and the representative algorithms include optical flow method, Kalman filter, etc. However, it is difficult to generate the model of the weld tracking system. , which is generally difficult to achieve in engineering; the discriminant model introduces an online learning classifier in the tracking process, and the positive samples containing the target and the negative samples containing the background constitute a sample set. This kind of algorithm has good tracking performance, and the representative algorithms include TLD, KCF, deep learning methods. Among them, the KCF algorithm uses the relevant filtering technology, which has good performance in tracking accuracy and robust performance, and is suitable for welding seam tracking. The disadvantage of the KCF algorithm is that it is easy to accumulate tracking errors and noise interference during long-term tracking, which leads to model drift. In the tracking process, there are noises such as arc light, scattered laser light, and metal spatter, which seriously affect the KCF algorithm to track the center of the weld, thereby reducing the accuracy of the weld tracking. .

At present, most of the welding seam tracking systems at home and abroad are machine vision systems based on traditional PCs, which mainly include computers, vision sensing systems, image acquisition cards, motion control cards, etc. It is beneficial to be used in some harsh industrial environments; the cost is high, the energy consumption is large, and the cost performance is not high; the installation is difficult, and it is not easy to debug. In addition, the traditional welding seam tracking system needs to formulate complex communication protocols with welding equipment, and needs to consider the transmission of sensing signals and control signals, which limits its application scope to a certain extent.

The traditional welding seam tracking system needs to manually input the welding seam type and adopt the corresponding welding seam image processing method before welding, which seriously reduces the welding automation level.

During the on-site welding process, there is strong noise interference in the image information of the weld seam detected by the line structured light vision sensor, which reduces the accuracy of the weld seam tracking. In recent years, the target tracking algorithm has been applied to the field of welding seam tracking and achieved good results. However, under the interference of strong noise, there is still the problem of tracking drift leading to tracking failure during the tracking process.

SUMMARY OF THE INVENTION

The present invention provides a welding seam tracking method and device based on independent deviation correction type deep learning in order to overcome the defects of low welding seam detection speed and accuracy described in the prior art.

The method includes the following steps:

S1: Collect structured light pictures of welds;

S2: Read the structured light picture of the weld;

S3: According to the requirements of automatic welding seam tracking, the welding seam type needs to be determined before welding starts. Use YOLOV3 to identify the current weld type and locate the center of the weld;

S4: Take the center position of the weld positioned by S3 as the tracking target, and initialize the KCF tracker;

The initialization of the KCF tracker includes selecting the bounding box as a positive sample, establishing a cyclic matrix through cyclic offset, training the classifier, and introducing a Gaussian kernel function to improve the performance of the classifier.

S5: The welding starts, and the KCF-YOLOV3 algorithm is used to track the welding seam in real time; until the welding ends.

Preferably, the positioning method of the center position of the weld is:

YOLOV3 detects the weld in the weld image, and the target bounding box will appear at the center of the weld, so the center coordinate of the bounding box is used as the center of the weld.

Preferably, S5 includes the following steps:

S5.1: Use KCF to track the weld, YOLOV3 detects the weld and continuously corrects the tracker to prevent tracking drift;

S5.2: Determine whether the classification score of YOLOV3 is higher than the fixed value, if so, execute S5.3, if not, execute S5.5;

S5.3: Calculate the offset error rate Po in the x-axis direction of the output weld center position of KCF and _YOLOV3 ; and determine whether Po is greater than the fixed threshold α or equal to zero, if so, execute S5.4, if not, execute S5. 5;

S5.4: Use the detection result of YOLOV3 as the center position of the weld; re-initialize the KCF tracker according to the detection result of YOLOV3;

S5.5: Take the KCF tracking result as the welding center position, and update the KCF tracker in real time;

S5.6: According to the KCF tracker initialized in S5.4 or the KCF tracker updated in S5.5, determine whether it is the last frame of the weld image, if not, return to S5.2; if so, end the welding.

Preferably, the calculation formula of the offset error rate P _o of the weld center position is:

In the formula, x(k) is the x-axis coordinate of the weld center position predicted by the KCF algorithm at time k, and x ^* (k) is the x-axis coordinate of the weld center position detected by the YOLO algorithm at time k.

The welding device of the present invention can realize the welding method, and the device comprises: a working platform, a controller, a vertical guide rail, a horizontal guide rail, a welding torch fixture, a line structured light vision sensor, and a servo motor;

The working platform is used to place the workpiece to be welded;

The controller, vertical guide rail, horizontal guide rail, welding torch fixture, line structured light vision sensor, and servo motor are arranged above the working platform;

The horizontal guide rail is arranged on the vertical guide rail, and can slide up and down along the vertical guide rail;

The motor is arranged at one end of the horizontal guide rail;

The welding torch fixture and the line structured light vision sensor are arranged on the horizontal guide rail;

Through the motor drive, the welding torch fixture and the line structured light vision sensor can be controlled to move in the horizontal direction on the horizontal direction guide rail;

The welding torch fixture is used to hold the welding torch; the line structured light vision sensor is used to collect the welded seam structured light pictures of the welding points, and send the collected welded seam structured light pictures to the controller;

The controller controls the rotation of the motor.

Preferably, the controller includes an upper computer and a lower computer;

The upper computer is connected with the line structured light vision sensor;

The line structured light vision sensor collects the welded seam structured light picture of the welding point and sends it to the host computer;

The upper computer analyzes the weld image and converts the weld deviation information into instructions and sends it to the lower computer;

The lower computer is used to receive the command from the upper computer, analyze the content of the command and convert it into the corresponding control function, and control the servo motor by sending control signals to the servo motor driver.

Preferably, the upper computer is a Raspberry Pi 4b development board, and the lower computer is a MINI-STM32 microcontroller.

Preferably, the motor is a servo motor.

Preferably, the vertical guide rail includes a first fixing bracket, a first lead screw, and a manual pulley;

The first lead screw includes a first screw rod and a first nut;

The first screw rod is rotatably arranged on the first fixing bracket, and one end of the first screw rod is coaxially connected with the manual pulley; the first screw rod can be driven to rotate by rotating the manual pulley;

The first nut is connected with the horizontal guide rail.

Preferably, the horizontal guide rail includes a second fixing bracket and a second lead screw;

The second fixing bracket is connected with the first nut;

The second lead screw includes a second screw rod and a second nut;

The second screw rod is rotatably arranged on the second fixing bracket, and one end of the second screw rod is coaxially connected with the motor drive shaft; the second screw rod can be driven to rotate by the motor;

The second nut is connected with the welding torch fixture and the wire structured light vision sensor.

Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

The welding device and the welding method of the invention can automatically identify various welding seams and precisely locate the central area of the welding seam. In addition, the present invention is based on the welding seam tracking method (KCF-YOLOV3) based on kernel correlation filtering (KCF) combined with deep learning detection (YOLOV3), in the case that the KCF algorithm may be disturbed by noise and cause tracking drift, YOLOV3 can locate the welding seam in real time Center, and continuously correct the KCF tracker, so as to effectively suppress the tracking drift, greatly improve the robust performance of the welding seam tracking algorithm, and further improve the welding seam tracking accuracy. It can effectively improve the welding seam detection speed and recognition accuracy, thereby improving welding efficiency.

Description of drawings

FIG. 1 is a flow chart of a welding seam tracking method based on independent deviation correction deep learning according to Embodiment 1. As shown in FIG.

Figure 2 is the flow chart of the KCF-YOLOV3 algorithm.

FIG. 3 is a schematic diagram of the welding seam tracking device based on the deep learning of the independent deviation correction type according to the first embodiment.

FIG. 4 is a schematic diagram of a vertical guide rail and a horizontal guide rail.

In the picture: 1-controller, 2-servo motor, 3-vertical guide rail, 4-horizontal guide rail, 5-line structured light vision sensor, 6-welding torch fixture, 7-welding torch, 8-arc welding machine, 9 - Working platform, 10 - Workpiece to be welded, 3.1 - First fixing bracket, 2 - First lead screw, 3.3 - Manual pulley, 4.1 - Second lead screw, 4.2 - Second fixing bracket.

Detailed ways

The accompanying drawings are for illustrative purposes only, and should not be construed as limitations on this patent;

In order to better illustrate this embodiment, some parts of the drawings are omitted, enlarged or reduced, which do not represent the size of the actual product;

It will be understood by those skilled in the art that some well-known structures and their descriptions may be omitted from the drawings.

The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

Example 1:

This embodiment provides a welding seam tracking method based on independent deviation correction deep learning.

The welding method described in this embodiment is based on a welding seam tracking method (KCF-YOLOV3) combined with kernel correlation filtering (KCF) and deep learning detection (YOLOV3). In the real-time welding process, the KCF algorithm constructs a large number of positive and negative samples by cyclic shift, uses ridge regression to train the classifier online, and uses the diagonalization property of the cyclic matrix in the Fourier space to convert the operation of the matrix into the Hadamard of the vector. product, which greatly reduces the amount of computation, makes the algorithm meet the real-time requirements, and uses the Gaussian kernel function mapping to improve the tracking accuracy.

The traditional kernel correlation filter tracking is a short-term target tracking method, and its limitation is that it is easy to accumulate tracking errors and noise interference leads to model drift during long-term tracking. During the welding process, strong arc light, splashes and other noises will interfere with the identification of the welding seam. At the same time, the target model is kept updated in a noisy environment. If tracking for a long time will accumulate tracking errors, resulting in tracking drift. Tracking drift occurs in the welding process, which will accumulate tracking errors and even lead to serious consequences of scrapping the workpiece. Therefore, keeping the target model accurate and avoiding tracking drift is very important for welding seam tracking. The center position of the weld seam is detected by the deep learning method, which can periodically correct the target model of the KCF tracker, so that the KCF can always track the weld seam accurately and stably.

YOLOV3 is a deep network based on convolutional neural network, which is currently the mainstream of target detection technology. The algorithm simplifies the classification and localization of targets into a regression problem, and is an end-to-end target detection algorithm. The algorithm of YOLOV3 is characterized by high detection speed and recognition accuracy, so it can realize real-time detection on embedded platforms. In order to realize the weld center position detection based on YOLOV3, it is necessary to collect a large number of weld images and conduct offline training of the network model. In order to improve the performance of the detector, the collected weld image can be the weld image before welding, or it can be the weld image with certain noise during the welding process. Seams, V-welds, fillet welds, annular welds, etc., to increase the diversity and complexity of training samples.

Before welding starts, turn on the welding seam detection function of YOLOV3, so as to identify the current welding seam type and locate the center position of the welding seam; then, use the positioned center position of the welding seam as the tracking target to initialize the KCF tracker; then, welding After the start, KCF is responsible for tracking the weld, and YOLOV3 is responsible for detecting the weld and continuously correcting the target model of the tracker to prevent tracking drift. During the welding process, when the classification module of YOLOV3 detects that a certain classification score of the current weld is high, it is considered that the location accuracy of the detected weld center position is high. Then calculate the offset error rate P _o of the output weld center position of KCF and YOLOV3 in the x-axis direction. Since the weld seam deviation is mainly generated in the x-axis direction, only the offset error in the x-axis direction is considered. Set a fixed threshold α, when P _o >α, it is considered that the KCF tracking drifts, and the target frame of the YOLO algorithm at this time is assigned to the KCF algorithm to re-track; when P _o ≤α, it is considered that the welding seam is tracked, and the KCF algorithm is used. Continue to track; when P _o =0, the KCF algorithm loses the target at this time, and the KCF algorithm is initialized again through the YOLO algorithm to achieve target tracking.

When the tracking drift phenomenon occurs, it is necessary to use the YOLOV3 detection algorithm to re-detect and retrieve the target, use the YOLOV3 detection result at the current moment as the center position of the weld, and re-initialize the KCF tracker. When there is no tracking drift phenomenon, that is, the reliability of the KCF tracking result is high, the KCF tracking result at the current moment is used as the center position of the weld, and the KCF tracker keeps updating online. The welding seam center position tracking at the current moment is completed, and the welding seam tracking of the next frame is continued, and the algorithm process is continuously looped until the last frame, and the welding is ended at the same time.

The method described in this embodiment will be described in detail below with reference to FIG. 1 : the welding method described in this embodiment includes the following steps:

S1: The structured light picture of the weld is collected by the line structured light vision sensor;

S2: Read the pictures collected by the line structured light vision sensor through Raspberry Pi 4b;

S3: Weld seam image processing in Raspberry Pi 4b: (According to the requirements of automatic welding seam tracking, the type of welding seam needs to be determined before welding starts. YOLOV3 detects the welding seam in the welding seam image, and a boundary will appear at the center of the welding seam box, while identifying the current weld type);

S4: Use the center position of the weld positioned by S3 as the tracking target, and initialize the KCF tracker; the initialization of the KCF tracker includes selecting the bounding box as a positive sample, establishing a cyclic matrix through cyclic offset, training the classifier, and introducing a Gaussian kernel function to improve Classifier performance.

S5: The welding starts, and the KCF-YOLOV3 algorithm (as shown in Figure 2) is used to track the welding seam in real time; until the welding ends.

Specifically, the position of the weld feature point is the center coordinate of the tracked bounding box, the weld deviation at the current moment is calculated, and the control command is transmitted to the MINI-STM32 according to the deviation.

In order to prevent the welding torch from moving frequently, MINI-STM32 sets a control dead zone. When the deviation is less than the fixed threshold, the motor will not be controlled; when the deviation is greater than the fixed threshold, MINI-STM32 transmits the corresponding control signal to the servo motor driver, and the motor rotates for a specified distance. , real-time correction, until the end of welding.

Among them, the positioning method of the center of the weld is: YOLOV3 detects the weld in the weld image, and the target bounding box will appear at the center of the weld, so the center coordinate of the bounding box is the center of the weld.

Further, S5 includes the following steps:

S5.1: Use KCF to track the weld, YOLOV3 detects the weld and continuously corrects the tracker to prevent tracking drift;

S5.2: Determine whether the classification score of YOLOV3 is higher than the fixed value, if so, execute S5.3, if not, execute S5.5;

S5.3: Calculate the offset error rate P _o of the output weld center position of KCF and YOLOV3 in the x-axis direction; and judge whether P _o is greater than the fixed threshold α or equal to zero, if so, execute S5.4, if not, implement S5 .5;

S5.4: Use the detection result of YOLOV3 as the center position of the weld; re-initialize the KCF tracker according to the detection result of YOLOV3;

S5.5: Take the KCF tracking result as the welding center position, and update the KCF tracker in real time;

S5.6: According to the KCF tracker initialized in S5.4 or the KCF tracker updated in S5.5, determine whether it is the last frame of the weld image, if not, return to S5.2; if so, end the welding.

Among them, the calculation formula of the offset error rate Po of the weld center position is:

In the formula, x(k) is the x-axis coordinate of the weld center position predicted by the KCF algorithm at time k, and x ^* (k) is the x-axis coordinate of the weld center position detected by the YOLO algorithm at time k.

The welding seam tracking method (KCF-YOLOV3) based on kernel correlation filtering (KCF) combined with deep learning detection (YOLOV3) described in this embodiment is a target tracking algorithm with strong anti-interference ability, which can effectively solve the tracking drift problem.

The welding seam detection based on the deep learning detection algorithm YOLOV3 adopted in this embodiment can automatically identify various welding seams and accurately locate the central area of the welding seam. The algorithm of YOLOV3 is characterized by high detection speed and recognition accuracy, so it can realize real-time detection on embedded platforms.

Example 2:

This embodiment provides a welding seam tracking device based on independent rectification type deep learning, as shown in Figure 3-4, the device includes: a working platform 9, a controller 1, a vertical guide rail 3, a horizontal guide rail 4, Welding torch fixture 6, line structured light vision sensor 5, servo servo motor 2;

The working platform 9 is used to place the workpiece 10 to be welded;

The controller 1, the vertical direction guide rail 3, the horizontal direction guide rail 4, the welding torch fixture 6, the linear structured light vision sensor 5, and the servo servo motor 2 are arranged above the working platform 9;

The horizontal guide rail 4 is arranged on the vertical guide rail 3, and can slide up and down along the vertical guide rail 3;

The servo motor 2 is arranged at one end of the guide rail 4 in the horizontal direction;

The welding torch fixture 6 and the linear structured light vision sensor 5 are arranged on the horizontal guide rail 4;

Driven by the servo motor 2, the welding torch fixture 6 and the linear structured light vision sensor 5 can be controlled to move in the horizontal direction on the horizontal direction guide rail 4;

The welding torch fixture 6 is used to hold the welding torch 7; the line structured light visual sensor 5 is used to collect the structured light picture of the welding seam of the welding point, and send the collected structured light picture of the welding seam to the controller 1;

The controller 1 controls the servo motor 2 to rotate.

Wherein, the controller 1 includes an upper computer and a lower computer;

The upper computer is connected with the line structured light vision sensor 5;

The line structured light vision sensor 5 collects the welded seam structured light picture of the welding point and sends it to the upper computer;

The upper computer analyzes the weld image and converts the weld deviation information into instructions and sends it to the lower computer;

The lower computer is used to receive the instructions from the upper computer, analyze the content of the instructions and convert them into corresponding control functions, and control the servo motor 2 by sending control signals to the servo motor driver.

The upper computer is a Raspberry Pi 4b development board, and the lower computer is a MINI-STM32 microcontroller.

In this embodiment, the host computer uses a Raspberry Pi 4b development board. Its CPU is a 64-bit 1.5GHz quad-core ARM Cortex-A72, its memory is 4GB of LPDDR4 SDRAM, and the storage system uses an SD/Micro SD card; There are four USB ports, one Ethernet port and HDMI high-definition video output port on the Pi motherboard. Through these ports, you can connect with other peripheral devices, such as keyboard, mouse, network cable and monitor, to form a fully functional computer. The operating system is based on the Linux system. The lower computer selects MINI-STM32 single-chip microcomputer, and its CPU is ARM32-bit Cortex-M3, which is a high-performance, low-cost, low-power microcontroller. The host computer development board analyzes the collected seam images and converts the weld seam deviation information into instructions and sends them to the lower computer; the upper computer is connected to the lower computer through serial communication, and the lower computer is used to receive instructions from the upper computer and parse the instructions. The content is converted into the corresponding control function, and the servo motor 2 is controlled by sending a control signal to the servo motor 2 driver.

The vertical guide rail 3 includes a first fixing bracket 3.1, a first lead screw 3.2, and a manual pulley 3.3;

The first lead screw 3.2 includes a first screw rod and a first nut;

The first screw is rotatably arranged on the first fixing bracket 3.1, and one end of the first screw is coaxially connected with the manual pulley 3.3; the first screw can be driven to rotate by rotating the manual pulley 3.3;

The first nut is connected to the horizontal guide rail 4 .

The horizontal guide rail 4 includes a second fixing bracket 4.2 and a second lead screw 4.1;

The second fixing bracket 4.2 is connected with the first nut;

The second screw 4.1 includes a second screw and a second nut;

The second screw is rotatably arranged on the second fixing bracket 4.2, and one end of the second screw is coaxially connected to the drive shaft of the servo motor 2; the second screw can be driven to rotate by the servo motor 2;

The second nut is connected to the welding torch holder 6 and the line structured light vision sensor 5 .

The welding device described in this embodiment can be easily installed and unloaded on an industrial manipulator. During the welding seam tracking process, the manipulator moves along the welding path to complete the relative movement with the workpiece, and the horizontal guide 4 can make the welding torch reciprocate in the horizontal direction to complete the movement. Correction. The welding device described in this embodiment can also be installed on various special welding machines, and the deviation can be corrected by moving on the guide rail 4 in the horizontal direction.

In this embodiment, the Raspberry Pi 4b development board based on the embedded system is used to replace the existing PC industrial computer, and the MINI-STM32 single-chip microcomputer is used to replace the existing motion control card. The system is small in size, cheap in price, easy to load and unload, and improves the flexibility, practicability and popularization of machine vision monitoring equipment.

In addition, Raspberry Pi 4b is a microcomputer based on the Linux operating system, so the algorithm proposed in the present invention is written in python language.

In addition, the welding device described in this embodiment can be installed in most welding equipment. As an independent welding seam deviation correction device, it avoids the problem of formulating complex communication protocols with welding equipment, has a wide range of applications, and is easy to install and debug.

The same or similar reference numbers correspond to the same or similar parts;

The terms describing the positional relationship in the accompanying drawings are only used for exemplary illustration, and should not be construed as a limitation on this patent;

Obviously, the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those of ordinary skill in the art, changes or modifications in other different forms can also be made on the basis of the above description. There is no need and cannot be exhaustive of all implementations here. Any modification, equivalent replacement and improvement made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (8)

1. A welding seam tracking method based on independent deviation correction type deep learning is characterized by comprising the following steps:

s1: collecting a welding seam structure light picture;

s2: reading a welding seam structure light picture;

s3: identifying the current weld type and positioning the center position of the weld by using YOLOV 3; the method for positioning the center position of the welding seam comprises the following steps: the YOLOV3 detects the weld in the weld image, and a target boundary frame appears at the center position of the weld, so the center coordinate of the boundary frame is taken as the center position of the weld;

s4: taking the welding seam center position positioned by S3 as a tracking target, and initializing a KCF tracker;

s5: the welding is started, and the KCF-YOLOV3 algorithm is used for tracking the welding seam in real time; until the welding is finished;

s5 includes the steps of:

s5.1: the KCF is utilized to track the welding seam, and the YOLOV3 detects the welding seam and continuously corrects the tracker to prevent tracking drift;

s5.2: judging whether the classification score of the Yolov3 is higher than a fixed value, if so, executing S5.3, and if not, executing S5.5;

s5.3: calculating offset error rate P of KCF and YOLOV3 in x-axis direction of output weld center position_o(ii) a And judging P_oIf the value is greater than the fixed threshold value alpha or equal to zero, if so, executing S5.4; if not, executing S5.5;

s5.4: taking the detection result of YOLOV3 as the center position of the weld joint; reinitializing the KCF tracker according to a detection result of YOLOV 3;

s5.5: taking a KCF tracking result as a welding center position, and updating a KCF tracker in real time;

s5.6: judging whether the frame is the last frame of the welding seam image according to the KCF tracker initialized in the S5.4 or the KCF tracker updated in the S5.5, and if not, returning to the S5.2; if yes, welding is finished.

2. The weld joint tracking method based on the independent correction type deep learning according to claim 1, characterized in that the error rate P of the deviation of the center position of the weld joint is_oThe calculation formula of (2) is as follows:

wherein x (k) is x-axis coordinate of the weld center position predicted by the KCF algorithm at k moment, x^*(k) And the x-axis coordinate of the weld center position detected by the YOLO algorithm at the moment k.

3. A weld tracking device based on independent deviation correction type deep learning, characterized in that the device comprises: the welding gun comprises a working platform, a controller, a vertical guide rail, a horizontal guide rail, a welding gun clamp, a line structured light vision sensor and a servo motor;

the working platform is used for placing a workpiece to be welded;

the controller, the vertical guide rail, the horizontal guide rail, the welding gun clamp, the line structured light vision sensor and the servo motor are arranged above the working platform;

the horizontal guide rail is arranged on the vertical guide rail and can slide up and down along the vertical guide rail;

the motor is arranged at one end of the guide rail in the horizontal direction;

the welding gun clamp and the line structured light vision sensor are arranged on the horizontal guide rail;

the welding gun clamp and the line structured optical vision sensor can be controlled to move along the horizontal direction on the horizontal direction guide rail through the transmission of the motor;

the welding gun clamp is used for clamping a welding gun; the linear structure light vision sensor is used for collecting a welding seam structure light picture of a welding point and sending the collected welding seam structure light picture to the controller;

the controller tracks the welding seam on the workpiece to be welded by adopting the welding seam tracking method based on the independent correction type deep learning of claim 1 or 2, and controls the motor to rotate according to the welding seam tracking result.

4. The weld joint tracking device based on the independent correction type deep learning according to claim 3, wherein the controller comprises an upper computer and a lower computer;

the upper computer is connected with the line structured light vision sensor;

the line structure light vision sensor collects a welding line structure light picture of a welding point and sends the welding line structure light picture to the upper computer;

analyzing the weld image of the upper computer, converting the weld deviation information into an instruction and sending the instruction to the lower computer;

the lower computer is used for receiving the instruction from the upper computer, analyzing the instruction content, converting the instruction content into a corresponding control function and controlling the servo motor by sending a control signal to the servo motor driver.

5. The weld joint tracking device based on the independent deviation correction type deep learning of claim 4, wherein the upper computer is a Raspberry Pi 4b development board, and the lower computer is a MINI-STM32 single chip microcomputer.

6. The weld tracking device based on the independent correction type deep learning of claim 5, wherein the motor is a servo motor.

7. The weld joint tracking device based on the independent correction type deep learning according to any one of claims 4 to 6, wherein the vertical direction guide rail comprises a first fixed bracket, a first lead screw and a manual pulley;

the first lead screw comprises a first screw rod and a first nut;

the first screw rod is rotatably arranged on the first fixing support, and one end of the first screw rod is coaxially connected with the manual pulley; the first screw rod can be driven to rotate by rotating the manual pulley;

the first nut is connected with the horizontal guide rail.

8. The weld joint tracking device based on the independent correction type deep learning of claim 7, wherein the horizontal guide rail comprises a second fixed bracket and a second lead screw;

the second fixing bracket is connected with the first nut;

the second lead screw comprises a second screw rod and a second nut;

the second screw rod is rotatably arranged on the second fixed support, and one end of the second screw rod is coaxially connected with the motor transmission shaft; the second screw rod can be driven to rotate by the motor;

the second nut is connected with the welding gun clamp and the line structured light vision sensor.

Images