Machine vision-based molten pool behavior visualization method and system

Abstract

The application provides a molten pool behavior visualization method and system based on machine vision, comprising the following steps: S1: acquiring dynamic image data inside the molten pool through synchrotron radiation imaging, and acquiring dynamic image data on the surface of the molten pool through high-speed camera shooting; S2: training and testing a YOLOv5 model by using the preprocessed dynamic image data, and generating a weight file used for detecting tungsten particles in the molten pool; S3: inputting the dynamic image data inside the molten pool and the dynamic image data on the surface of the molten pool into the trained YOLOv5 model, tracking the movement of the tungsten particles in real time, and obtaining a visualized image of the movement track of the tungsten particles. By introducing tungsten particles as tracers, combining synchrotron radiation imaging technology and a high-speed camera, the application can realize multi-level accurate observation of the flow behavior inside and on the surface of the molten pool. The system uses machine vision technology and a deep learning model to automatically detect and track the movement track of the tungsten particles, thereby significantly improving the accuracy and efficiency of observation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser processing, in particular to a molten pool behavior visualization method and system based on machine vision. BACKGROUND

[0002] As one of the core technologies in modern manufacturing, laser processing technology is widely used in high-precision fields such as aerospace, automobile manufacturing, and electronic devices. Due to its advantages of high energy density, precise control, and non-contact processing, laser welding and additive manufacturing technology have become a key means for efficient manufacturing of complex components. In the laser processing process, the behavior of the molten pool directly affects the microstructure and final performance of the material, so the research and characterization of the molten pool are particularly important.

[0003] The flow behavior of the molten pool, including its morphology, flow speed, and flow direction, determines the cooling rate, grain growth, and defect distribution of the processed piece, thereby directly affecting the mechanical properties of the final product. Therefore, studying the flow characteristics of the molten pool is of great significance for optimizing laser processing technology and improving product performance. However, due to the high temperature inside the molten pool, the high-speed flow of molten metal, and the complex thermal-physical reactions, the behavior of the molten pool is difficult to accurately observe and quantitatively analyze through conventional experimental methods.

[0004] Currently, numerical simulation technology is an important tool for studying the behavior of the molten pool. By establishing a physical model of the molten pool, numerical simulation can predict the temperature field, velocity field, and flow behavior of the molten pool. However, numerical simulation has certain limitations, especially under complex actual working conditions, the physical model often needs to be simplified, which leads to a certain deviation between the simulation results and the actual flow behavior of the molten pool. This deviation is due to the incompleteness of the physical model and the variability of the complex processing environment, therefore, relying solely on numerical simulation cannot accurately characterize the real behavior of the molten pool.

[0005] Under this background, there is an urgent need for an experimental method that can accurately characterize the flow of the molten pool to verify the numerical simulation results and further reveal the real dynamic behavior of the molten pool. SUMMARY

[0006] In view of the defects in the prior art, the purpose of the present application is to provide a molten pool behavior visualization method and system based on machine vision.

[0007] According to the molten pool behavior visualization method based on machine vision provided by the present application, the method comprises the following steps:

[0008] Step S1: Obtain dynamic image data inside the molten pool through synchrotron imaging, and obtain dynamic image data on the surface of the molten pool through high-speed camera shooting;

[0009] Step S2: training and testing the YOLOv5 model using the preprocessed dynamic image data to generate a weight file for detecting tungsten particles in the molten pool;

[0010] Step S3: inputting the dynamic image data inside the molten pool and the dynamic image data on the surface of the molten pool into the trained YOLOv5 model to track the motion of the tungsten particles in real time and obtain a visual image of the motion trajectory of the tungsten particles.

[0011] Preferably, the step S1 comprises introducing tungsten particles as tracers for tracking the flow behavior inside the molten pool; obtaining X-ray images through synchrotron radiation imaging to show the dynamic changes inside the molten pool; and recording the flow behavior on the surface of the molten pool through high-speed camera imaging.

[0012] Preferably, the step S2 comprises:

[0013] Step S2.1: manually or automatically labeling the tungsten particles based on the preprocessed dynamic image data to generate label data of the positions of the tungsten particles for model training;

[0014] Step S2.2: training the YOLOv5 model using the generated data set, and iteratively training multiple times until the model can accurately identify the tungsten particles;

[0015] Step S2.3: generating the best weight file after multiple rounds of training and verification, and applying the weight file to the model;

[0016] Step S2.4: testing the trained YOLOv5 model using data from different scenarios to verify the accuracy of the model in detecting tungsten particles.

[0017] Preferably, the step S3 comprises:

[0018] Step S3.1: importing the dynamic image data inside the molten pool and the dynamic image data on the surface of the molten pool into the trained YOLOv5 model, which will automatically identify and label the tungsten particles in the molten pool.

[0019] Step S3.2: the user selects tungsten particles with obvious behavior characteristics through an interactive page, and the system tracks the motion of the tungsten particles in real time;

[0020] Step S3.3: based on the continuous frame data, the system uses a speed and direction calculation unit and the multi-target tracking function of the SORT algorithm to calculate and display the speed and direction of the tungsten particles in real time;

[0021] Step S3.4: the system automatically stores the position information of the selected tungsten particles into a log file and saves the trajectory data of the tungsten particles;

[0022] Step S3.5: generating a visualized image of the motion trajectory of the tungsten particle by reading the tungsten particle position information in the log file.

[0023] According to the present application, a molten pool behavior visualization system based on machine vision is provided, comprising: a laser welding and additive unit, a data acquisition unit, a computer vision unit, and a data processing unit.

[0024] The laser welding and additive unit performs laser welding and additive manufacturing, and introduces tungsten particles as tracer particles during the processing.

[0025] The data acquisition unit is used to capture the flow image of the molten pool during welding or additive manufacturing;

[0026] The computer vision unit performs target detection of tungsten particles based on a YOLOv5 model trained by a deep learning algorithm, and combines a SORT multi-target tracking algorithm to analyze the motion trajectory in real time;

[0027] The data processing unit judges the flow state of the molten pool by calculating the speed, direction and position change of the tungsten particles, and outputs the visualized real-time tracking results.

[0028] Preferably, the data acquisition unit comprises a synchrotron radiation imaging unit and a high-speed camera imaging unit;

[0029] The synchrotron radiation imaging unit tracks the tracer particles inside the molten pool through X-ray imaging, and the high-speed camera imaging unit is used to track the tracer particles on the surface of the molten pool.

[0030] Preferably, the computer vision unit comprises a deep learning algorithm YOLOv5 model and a SORT multi-target tracking algorithm;

[0031] The YOLOv5 model accurately locates and identifies the tungsten particles, generates the bounding box coordinates of their positions, and the SORT multi-target tracking algorithm analyzes the motion trajectory of the identified tungsten particles in multiple frames of images, thereby continuously tracking and analyzing the flow behavior inside or on the surface of the molten pool.

[0032] Preferably, the data processing unit accurately calculates the motion speed and direction of the tungsten particles based on the tungsten particle position information provided by the SORT algorithm;

[0033] By comparing the displacement of the tungsten particles in consecutive frames, calculating the displacement between two frames, and combining the time interval between frames, the motion speed of the tungsten particles is calculated.

[0034] According to the relative position change of the tungsten particles in consecutive frames, the motion direction is determined.

[0035] Preferably, the YOLOv5 model generates multiple candidate bounding boxes through an anchor box mechanism, classifies and regresses these candidate boxes using a convolutional neural network to determine the position and size of the target in the image; when a tungsten particle is detected, the bounding box generated by the system is determined by the upper left corner and the lower right corner coordinates, and the center point coordinates of the target are recorded for motion tracking;

[0036] For tungsten particles of different sizes, the YOLOv5 model dynamically adjusts the size of the bounding box, classifies and outputs the class and its confidence by classifying the candidate box, and uses the non-maximum suppression algorithm to filter out the optimal bounding box through the intersection-over-union ratio to avoid repeated detection.

[0037] Preferably, the SORT multi-target tracking algorithm predicts the position of the tungsten particle in the next frame through Kalman filtering and updates using historical data such as speed and acceleration of the previous frame;

[0038] The bounding box of each frame of tungsten particle is generated by the YOLOv5 model detection, and the SORT multi-target tracking algorithm matches these detection boxes with the results predicted by Kalman filtering through the Hungarian algorithm. For the tungsten particles that are successfully matched, the system updates their trajectory information; if a tungsten particle is not detected in a certain frame, SORT will record the number of times the target is lost, and will remove it when it is not matched for consecutive multiple frames.

[0039] Compared with the prior art, the present application has the following beneficial effects:

[0040] 1. By introducing tungsten particles as tracers, combined with synchronous radiation imaging technology and high-speed cameras, the present application can achieve multi-level accurate observation of the internal and surface flow behavior of the molten pool. The system uses machine vision technology and deep learning models to automatically detect and track the motion trajectory of the tungsten particles, significantly improving the accuracy and efficiency of the observation.

[0041] 2. Compared with the traditional manual observation method, the automatic tracking not only reduces human intervention, but also enhances the stability and consistency of the observation results, providing reliable data support for the study of molten pool behavior.

[0042] 3. The present application uses the YOLOv5 deep learning model and the SORT multi-target tracking algorithm to real-time identify and track the motion of tungsten particles inside and outside the molten pool. The system can automatically calculate the speed and direction of the tungsten particles and generate real-time visual results to help users intuitively analyze the dynamic behavior of the molten pool. Through this method, researchers can more comprehensively understand the flow characteristics of the molten pool, thereby optimizing the laser processing process and improving the final quality and performance of the parts. BRIEF DESCRIPTION OF DRAWINGS

[0043] Other features, objects, and advantages of the application will become more apparent from the following detailed description when read in conjunction with the accompanying drawings:

[0044] Figure 1 For the process flowchart and system configuration diagram of the present application;

[0045] Figure 2 For the system structure schematic diagram of the present application;

[0046] Figure 3 For the synchrotron program interactive interface of the present application;

[0047] Figure 4 For the 20th tracer particle trajectory diagram in the embodiment of the present application;

[0048] Figure 5 For the trajectory diagram of different tracer particles in the embodiment of the present application;

[0049] Figure 6 For the trajectory reproduction of different tracer particles in the embodiment of the present application. DETAILED DESCRIPTION

[0050] The present application will be described in detail below with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These are within the scope of the present application.

[0051] The present application discloses a molten pool behavior visualization method based on machine vision, which automatically identifies tungsten particle positions through deep learning technology, and realizes real-time tracking of tungsten particle motion trajectory and accurate calculation of speed direction by combining multi-target tracking algorithm. The system greatly improves the observation accuracy and efficiency of the molten pool behavior, provides a new experimental method for studying the molten pool behavior in laser processing, and overcomes the limitations of traditional manual observation.

[0052] Specifically, referring to Figure 1 The molten pool behavior visualization method based on machine vision includes:

[0053] Step S1: data acquisition and preprocessing: dynamic image data inside the molten pool is obtained by laser processing and synchrotron imaging, or dynamic image data on the surface of the molten pool is obtained by laser processing and high-speed camera shooting, and contrast adjustment and format conversion processing are performed on these data.

[0054] Step S2: model training and testing: the preprocessed data set is used to train and test the YOLOv5 model, and the best weight file for accurately detecting tungsten particles is generated.

[0055] Step S3: Specific use method of the model: input the pre-processed data into the model, automatically detect the tungsten particles, and select the particles through the interactive interface for real-time tracking, record their coordinates, and finally visualize the particle trajectories.

[0056] In a specific embodiment, the acquisition and preprocessing of data includes:

[0057] Laser processing (welding / additive manufacturing): Introduce tungsten particles as tracers during laser welding or additive manufacturing to track the flow behavior inside the molten pool. During the processing, the system introduces tungsten particles as tracers to track the molten pool behavior. For laser additive manufacturing, the system synchronously feeds tungsten particles with powder materials into the molten pool through the powder feeder, ensuring uniform distribution of tungsten particles without affecting the process. For laser welding processes, the system spreads a small amount of tungsten powder on the welding surface to allow tungsten particles to enter the molten pool.

[0058] Synchrotron imaging or high-speed camera imaging: Through synchrotron imaging technology, real-time capture of the flow process inside the molten pool, high-resolution X-ray images are obtained, showing the dynamic changes inside the molten pool. In addition, by recording the flow behavior on the surface of the molten pool through high-speed camera imaging, the motion trajectory of the surface tracer particles can be captured, providing visual data support for indirectly inferring the overall flow behavior of the molten pool. The combination of these two imaging methods provides multi-level dynamic image data of the molten pool inside and surface.

[0059] Preprocessing: Contrast adjustment: Adjust the contrast of the obtained original image to improve image quality and make the movement of tungsten particles more clearly visible; Image to video: export the image sequence to video format for subsequent model training and analysis.

[0060] In a specific embodiment, model training and testing includes:

[0061] Dataset generation: Based on the pre-processed images and videos, manually or automatically label the tungsten particles to generate label data of the tungsten particle positions for model training.

[0062] YOLOv5 model training: Use the generated dataset to train the YOLOv5 model, adjust the anchor box, learning rate, and other model parameters, and repeatedly iterate the training until the model can accurately identify the tungsten particles.

[0063] Weight file generation: After multiple rounds of training and verification, generate a suitable weight file best.pt for actual use for further application of the model.

[0064] Model testing: The trained YOLOv5 model is tested using data from different scenarios to verify its accuracy in detecting tungsten particles, ensuring that the model can be applied in actual operations.

[0065] In a specific implementation, the specific use of the model includes:

[0066] Data input: The pre-processed video or image data is imported into the trained YOLOv5 model, which automatically identifies and labels the tungsten particles in the molten pool.

[0067] Interactive page operation: The user selects tungsten particles with obvious behavior characteristics through the interactive page, and the system begins to track the motion of these particles in real time.

[0068] Speed and direction calculation: Based on consecutive frame data, the system uses the speed and direction calculation unit, combined with the multi-target tracking function of the SORT algorithm, to calculate and display the speed and direction of the tungsten particles in real time. The SORT algorithm matches and tracks multiple targets through Kalman filtering and the Hungarian algorithm, ensuring the continuity of the tungsten particles in multiple images and predicting their future motion positions to enhance the stability of the tracking.

[0069] Log recording: The system automatically stores the position information of the selected tungsten particles into a log file, saving their trajectory data, which can be used for visual analysis in the future.

[0070] Trajectory visualization: By reading the tungsten particle position information in the log file, a visual image of its motion trajectory is generated, helping to study the flow behavior of the molten pool and providing data support for subsequent process optimization.

[0071] The present application also discloses a molten pool behavior visualization system based on machine vision, comprising:

[0072] Laser welding / additive unit: Laser welding / additive manufacturing is performed, and this unit controls the process parameters of laser welding or additive manufacturing, introducing tungsten particles into the molten pool as tracers during the machining process for tracking the internal flow of the molten pool;

[0073] Data acquisition unit: used to capture the flow images of the molten pool during welding or additive manufacturing, including a synchrotron imaging unit to capture the internal flow of the molten pool using X-ray imaging technology, and a high-speed camera to record the flow behavior on the surface of the molten pool, obtaining different levels of molten pool flow information through multiple imaging means;

[0074] Computer vision unit: YOLOv5 model trained based on deep learning algorithm for target detection of tungsten particles, combined with SORT multi-target tracking algorithm for real-time analysis of its motion trajectory, ensuring accurate identification of tungsten particles in the molten pool and on the surface of the molten pool.

[0075] Data processing unit: By calculating the changes in velocity, direction and position of tungsten particles, it determines the flow state of the molten pool and outputs visualized real-time tracking results.

[0076] Specifically, the laser welding / additive manufacturing unit includes a series of industrial-grade equipment such as a laser, laser head, powder feeder, water chiller, and electrical control box, which can perform additive manufacturing or welding processes on the substrate surface.

[0077] The laser transmits the laser beam to the laser head via an optical fiber, concentrating energy on the substrate surface to heat and melt the powder material, thereby forming a molten pool. The powder feeder uniformly feeds the additive material (such as AlSi7Mg powder) and a certain proportion of tungsten tracer particles into the molten pool, ensuring that the tungsten particles are distributed in the molten pool for subsequent flow tracking.

[0078] Specifically, the data acquisition unit includes synchrotron radiation imaging and high-speed camera imaging, wherein synchrotron radiation imaging is mainly used to track tracer particles inside the molten pool, while high-speed camera imaging is mainly used to track tracer particles on the surface of the molten pool.

[0079] The synchrotron radiation imaging employed multicolor X-rays to provide optimal penetration and signal-to-noise ratio. The X-ray energy was 20 keV, enabling high-precision observation and recording of the flow behavior within the molten pool. As the laser moved across the top surface of the substrate, X-rays penetrated the sample for imaging. The imaging system consisted of a 200 μm thick LuAg scintillator and a 4x magnifying glass. After passing through the sample, the X-rays were converted into visible light by the scintillator and finally recorded by a high-resolution camera. The imaging field of view was 3 mm × 1.6 mm (width × height). The entire observation process was recorded at a rate of 20,000 frames per second, with exposure times ranging from 1 to 40 μs, lasting several seconds, completely capturing the entire process of laser activation, movement, and deactivation.

[0080] When the high-speed camera captures images during the processing, the frame rate ranges from 5 to 200 fps.

[0081] By incorporating a small amount of tungsten particles into the powder feed material, allowing them to enter the molten pool along with the powder during additive manufacturing, the movement trajectory of the tungsten particles within the molten pool can clearly reflect changes in flow direction, velocity, and morphology. The high density of tungsten particles can be clearly displayed in X-ray imaging, providing excellent markers for observing fluid movement within the molten pool.

[0082] Specifically, the computer vision unit includes a deep learning algorithm YOLOv5 model and a SORT multi-target tracking algorithm, which are used together for real-time analysis of the motion of tungsten particles. YOLOv5 is used to accurately locate and identify tungsten particles, generating the coordinates of the bounding box of their position; the SORT algorithm analyzes the motion trajectory of the identified tungsten particles in multiple frames of images by tracking them, thereby realizing continuous tracking and analysis of the flow behavior inside or on the surface of the molten pool.

[0083] The YOLOv5 target detection model is characterized in that YOLOv5 generates multiple candidate bounding boxes (bbox) through an anchor box mechanism, as shown in Figure 2 The image is divided into multiple sub-regions, and each grid is responsible for predicting the target in its region. When a tungsten particle is detected, the bounding box generated by the system is determined by the coordinates of the upper left corner (x1, y1) and the lower right corner (x2, y2), and the center point (x_center, y_center) of the target is recorded to help subsequent motion tracking. For tungsten particles of different sizes, YOLOv5 dynamically adjusts the size of the bounding box to ensure accurate coverage of the target. On this basis, YOLOv5 classifies the candidate boxes and outputs the class (tungsten particle) and its confidence to achieve detection. In addition, YOLOv5 uses the non-maximum suppression (NMS) algorithm to filter out the optimal bounding box through the intersection over union (IoU) to avoid repeated detection.

[0084] The SORT multi-target tracking algorithm is characterized in that the SORT algorithm predicts the position of the tungsten particle in the next frame through Kalman filtering and updates it using historical data such as velocity and acceleration from the previous frame. The bounding box of the tungsten particle in each frame is generated by YOLOv5 detection, and SORT matches these detection boxes with the results of Kalman filter prediction through the Hungarian algorithm to ensure that the trajectory of the tungsten particle in consecutive frames remains consistent. For tungsten particles that are successfully matched, the system updates their trajectory information; if a tungsten particle is not detected in a certain frame, SORT will record the number of times the target is lost and remove it if it is not matched for consecutive multiple frames.

[0085] Specifically, the data processing unit realizes a velocity and direction real-time calculation unit that accurately calculates the motion velocity and direction of the tungsten particle based on the position information (x1, y1, t) of the tungsten particle provided by the SORT algorithm through mathematical formulas. The system compares the displacement of the tungsten particle in consecutive frames using the Euclidean distance formula:

[0086]

[0087] Where (x1, y1) and (x2, y2) are the coordinates of the tungsten particle in two consecutive frames, the displacement between the two frames is calculated, and the velocity of the tungsten particle is calculated in real time by combining the time interval t between the frames.

[0088] Furthermore, the system determines the direction of motion of tungsten particles using an azimuth calculation formula based on the changes in their relative positions across consecutive frames:

[0089] θ = atan 2 (y2-y1, x2-x1)

[0090] The results are updated in real time and stored in a data structure for further analysis of the molten pool flow behavior.

[0091] The data processing unit allows users to select tungsten particles to be observed within the program and save their position information to a log file. Through the interactive interface, users can double-click to select a tungsten particle; its coordinates will be recorded in real time and stored in the log file. This log file contains time-series data of the tracer particles throughout the tracking process, which can then be used for trajectory calculation and visualization.

[0092] The following example further illustrates the machine vision-based visualization method for melt pool behavior provided by this invention.

[0093] 1. Laser Additive Manufacturing: A DED (Dry-Edge Electrode) powder feeding additive manufacturing mode was employed. The substrate was a 1060 aluminum substrate, and the powder used was AlSi7Mg powder with a particle size of approximately 50 micrometers, incorporating 5% tungsten tracer particles. During the laser additive manufacturing process, the laser parameters were selected as follows: laser power 1500W, scanning speed 1200 mm / min, powder feeding speed 0.7 r / min, powder gas flow rate 20 L / min, and shielding gas flow rate 12 L / min. The raw data was obtained by recording the laser welding process using a synchrotron radiation rapid processing device.

[0094] 2. Data Preprocessing: The acquired raw synchrotron radiation imaging data is imported into the image processing software ImageJ. Brightness and contrast are adjusted to optimize image quality, making the tungsten particle trajectories more clearly visible. After processing, the image sequence is exported as a video.

[0095] 3. Program Preprocessing: Import the preprocessed video file into the pre-trained program folder and rename the file to "testvideo" so that the program can recognize and test it (the program only performs tests on files named "testvideo"). Then, configure the program's runtime environment in the terminal, ensuring that the required dependencies and settings are correctly loaded, and the automatic detection and analysis process can begin.

[0096] 4、Program running: After the program starts, a demo interface (interactive interface) will be displayed, in which multiple tracer particles have been automatically identified by the system (as shown in the attached Figure 3 ). In this interface, the user can select the tracer particle of interest, such as tracer particle No. 20, and by double-clicking on the particle, its real-time motion trajectory will be displayed in green (see the attached Figure 4 ), making it easy to observe. At the same time, the speed and direction information of the selected particle will be displayed in real time in the upper left corner of the interface. After a certain period of observation, the complete motion trajectory of tracer particle No. 20 will be gradually plotted (see the attached Figure 5 ), providing intuitive data support for the analysis of the flow behavior of the molten pool.

[0097] 5、Data post-processing. During the program running process, the user can select multiple tracer particles for tracking analysis, such as particles No. 2, 62, and 99 (see the attached Figure 5 ). After selecting these particles, the system will automatically store the coordinate data of their motion trajectories in the log folder of the program folder. By reading the data in the log file, the user can accurately reproduce and analyze the motion trajectories of the selected particles (see the attached Figure 6 ), thereby further understanding the flow behavior characteristics of the molten pool and providing a basis for subsequent optimization of the additive process.

[0098] Those skilled in the art know that, in addition to implementing the system provided by the present application and each device, module, unit thereof in the form of pure computer readable program code, the same function can also be achieved by logically programming the method steps in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system provided by the present application and each device, module, unit thereof can be considered as a hardware component, and the devices, modules, units included therein for achieving various functions can also be considered as structures within the hardware component; the devices, modules, units for achieving various functions can also be considered as both software modules implementing methods and structures within hardware components.

[0099] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0100] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be combined with each other at will without conflict.

Claims

1. A machine vision-based method for visualizing molten pool behavior, characterized in that, include: Step S1: Obtain dynamic image data of the inside of the molten pool through synchrotron radiation imaging, and obtain dynamic image data of the surface of the molten pool through high-speed camera imaging; Step S2: Train and test the YOLOv5 model using preprocessed dynamic image data to generate a weight file for detecting tungsten particles in the molten pool; Step S3: Input the dynamic image data inside the molten pool and the dynamic image data on the surface of the molten pool into the trained YOLOv5 model to track the motion of tungsten particles in real time and obtain a visualized image of the tungsten particle motion trajectory. Step S1 includes: introducing tungsten particles as tracers to track the flow behavior inside the molten pool; acquiring X-ray images through synchrotron radiation imaging to show the dynamic changes inside the molten pool; and recording the flow behavior on the surface of the molten pool using a high-speed camera. For laser additive manufacturing, tungsten particles and powder materials are simultaneously fed into the molten pool using a powder feeder; for laser welding, tungsten powder is laid on the welding surface to allow tungsten particles to enter the molten pool. Step S3 includes: Step S3.1: Import the dynamic image data of the molten pool into the trained YOLOv5 model. The model will automatically identify and label the tungsten particles in the molten pool. Step S3.2: The user selects tungsten particles with obvious behavioral characteristics through the interactive page, and the system tracks and responds to the movement of the tungsten particles in real time; Step S3.3: Based on continuous frame data, the system uses the velocity and direction calculation unit, combined with the multi-target tracking function of the SORT algorithm, to calculate and display the velocity and direction of the tungsten particles in real time; Step S3.4: The system automatically stores the position information of the selected tungsten particles in the log file, saving their trajectory data; Step S3.5: Generate a visual image of the tungsten particle's trajectory by reading the tungsten particle position information from the log file.

2. The machine vision-based molten pool behavior visualization method according to claim 1, characterized in that, Step S2 includes: Step S2.1: Based on the preprocessed dynamic image data, label the tungsten particles to generate label data of the tungsten particle positions for model training; Step S2.2: Train the YOLOv5 model using the generated dataset, iterating multiple times until the model can accurately identify tungsten particles; Step S2.3: After multiple rounds of training and validation, generate the optimal weight file and use it for model application; Step S2.4: Test the trained YOLOv5 model using data from different scenarios to verify its accuracy in detecting tungsten particles.

3. A machine vision-based molten pool behavior visualization system, characterized in that, It includes: a laser welding and additive manufacturing unit, a data acquisition unit, a computer vision unit, and a data processing unit; The laser welding and additive manufacturing unit performs laser welding and additive manufacturing, and introduces tungsten particles as tracer particles during the processing. The data acquisition unit is used to capture images of the flow of the molten pool during welding or additive manufacturing. The computer vision unit performs target detection of tungsten particles based on the YOLOv5 model trained by deep learning algorithm, and analyzes their motion trajectory in real time by combining the SORT multi-target tracking algorithm. The data processing unit determines the flow state of the molten pool by calculating the changes in the velocity, direction, and position of the tungsten particles, and outputs visualized real-time tracking results. The data acquisition unit includes a synchrotron radiation imaging unit and a high-speed camera imaging unit; The synchrotron radiation imaging unit tracks tracer particles inside the molten pool using X-ray imaging, while the high-speed camera imaging unit tracks tracer particles on the surface of the molten pool.

4. The machine vision-based molten pool behavior visualization system according to claim 3, characterized in that, The computer vision unit includes the YOLOv5 deep learning algorithm model and the SORT multi-object tracking algorithm; The YOLOv5 model accurately locates and identifies tungsten particles, generating bounding box coordinates of their positions. The SORT multi-target tracking algorithm tracks the identified tungsten particles and analyzes their motion trajectories in multiple frames of images, thereby achieving continuous tracking and analysis of the flow behavior inside or on the surface of the molten pool.

5. The machine vision-based molten pool behavior visualization system according to claim 3, characterized in that, The data processing unit accurately calculates the velocity and direction of tungsten particles based on the position information of the tungsten particles provided by the SORT algorithm. The velocity of tungsten particles is calculated by comparing the displacement of tungsten particles in consecutive frames, calculating the displacement between two frames, and combining the time interval between frames. The direction of motion of tungsten particles is determined by the change in their relative positions in consecutive frames.

6. The machine vision-based molten pool behavior visualization system according to claim 3, characterized in that, The YOLOv5 model generates multiple candidate bounding boxes through the anchor box mechanism, and uses a convolutional neural network to classify and regress these candidate boxes to determine the position and size of the target in the image. When tungsten particles are detected, the bounding box generated by the system is determined by the coordinates of the upper left and lower right corners, and the coordinates of the center point of the target are recorded for motion tracking. For tungsten particles of different scales, the YOLOv5 model dynamically adjusts the size of the bounding boxes, classifies candidate boxes and outputs the categories and their confidence scores, uses the non-maximum suppression algorithm, and selects the optimal bounding boxes by cross-union ratio to avoid duplicate detection.

7. The machine vision-based molten pool behavior visualization system according to claim 3, characterized in that, The SORT multi-target tracking algorithm predicts the position of tungsten particles in the next frame using Kalman filtering and updates it using historical data such as velocity and acceleration from the previous frame. The bounding boxes of tungsten particles in each frame are generated by the YOLOv5 model. The SORT multi-object tracking algorithm matches these bounding boxes with the Kalman filter prediction results using the Hungarian algorithm. For tungsten particles that are successfully matched, the system updates their trajectory information. If a tungsten particle is not detected in a frame, SORT will record the number of times the target is missed and remove it if it fails to match for several consecutive frames.

Images