Welding robot control system based on ai vision

Abstract

The application relates to the field of AI vision technology, and particularly discloses a welding robot control system based on AI vision, which comprises a collection module, an analysis module, an optimization module and a control module, in particular: the collection module collects a molten pool image, pre-processes the molten pool image to obtain a target image, and extracts the edge of the molten pool in the target image based on AI vision; the analysis module obtains the actual width A1 of the molten pool based on the edge, obtains the actual depth A2 of the molten pool, calculates the width-depth ratio, and marks as an abnormal situation when the width-depth ratio does not belong to a preset safety range; the optimization module collects features in real time within a preset collection cycle when an abnormal situation occurs, constructs two sample sets based on the features, and obtains the importance score of the features based on the sample sets; and the control module executes a preset control strategy based on the importance score. The application can reduce energy and welding material consumption and help green and efficient welding manufacturing.

Description

TECHNICAL FIELD

[0001] The application relates to the field of AI vision, in particular to a welding robot control system based on AI vision. BACKGROUND

[0002] AI vision is an important branch of artificial intelligence, which enables machines to have human-like visual understanding through deep learning. Convolutional neural networks extract multi-scale features, attention mechanisms focus on key areas, and complete tasks such as object detection, segmentation, and pose estimation.

[0003] With the continuous development of industrial automation technology, robots are widely used in manufacturing, medical treatment, service and other fields. Welding robots are widely used in automobile manufacturing, shipbuilding, aerospace and other industries due to their high precision, high efficiency, strong repeatability and safety. In production, welding robots effectively improve the quality of welding points and production efficiency through advanced motion control systems and visual recognition technology, reduce the danger of manual work, and help enterprises realize automation upgrade.

[0004] Existing welding technology research focuses on weld forming quality and strength, and pays little attention to the control of energy and material loss, resulting in serious resource waste in the production process, which is not conducive to green, low-carbon manufacturing and sustainable development. SUMMARY

[0005] The purpose of the present application is to provide a welding robot control system based on AI vision, which solves the following technical problems:

[0006] Existing welding technology research focuses on weld forming quality and strength, and pays little attention to the control of energy and material loss, resulting in serious resource waste in the production process, which is not conducive to green, low-carbon manufacturing and sustainable development.

[0007] The purpose of the present application can be achieved by the following technical solutions:

[0008] The welding robot control system based on AI vision comprises a collection module, an analysis module, an optimization module and a control module, specifically:

[0009] The collection module collects molten pool images, pre-processes the molten pool images to obtain target images, the pre-processing includes denoising and enhancement, and extracts the edge of the molten pool in the target image based on AI vision;

[0010] The analysis module obtains the actual width A1 of the molten pool based on the edge, and obtains the actual depth A2 of the molten pool, calculates the width-depth ratio D1=A1 / A2, and marks it as an abnormal situation when the width-depth ratio does not belong to the preset safe range;

[0011] An optimization module: when an abnormal situation occurs, real-time collection of characteristics in a preset collection period, the characteristics including a width-depth ratio, a welding speed, a welding current, an energy consumption and a welding rod material consumption, construction of two sample sets based on the characteristics, acquisition of an importance score of the characteristics based on the sample sets;

[0012] A control module: execution of a preset control strategy based on the importance score, the control strategy including an energy control strategy and a material control strategy, the energy control strategy being used for controlling a current of a welding robot to reduce the energy consumption, the material control strategy being used for controlling a welding rod moving speed of the welding robot to reduce the welding rod material consumption.

[0013] As a further scheme of the present application: the construction of the two sample sets based on the characteristics includes:

[0014] The welding speed, the welding current and the energy consumption are taken as first initial samples, and the welding speed, the welding current and the welding rod material consumption are taken as second initial samples;

[0015] The width-depth ratio is taken as a label of the first initial samples to obtain a first sample, the first samples corresponding to consecutive n time points are acquired to generate a first sample set, and n is a preset number;

[0016] The width-depth ratio is taken as a label of the second initial samples to obtain a second sample, the second samples corresponding to consecutive n time points are acquired to generate a second sample set.

[0017] As a further scheme of the present application: the acquisition of the importance score of the characteristics based on the sample sets includes:

[0018] The width-depth ratio is taken as a regression target, a random forest model is trained by using the first sample set, a purity improvement value when the characteristics are taken as nodes for splitting is calculated, the purity improvement values of the same characteristics are summed to obtain a purity improvement total sum, the purity improvement total sum is normalized to obtain the importance score;

[0019] The width-depth ratio is taken as a regression target, a random forest model is trained by using the second sample set, a purity improvement value when the characteristics are taken as nodes for splitting is calculated, the purity improvement values of the same characteristics are summed to obtain a purity improvement total sum, the purity improvement total sum is normalized to obtain the importance score.

[0020] As a further scheme of the present application: the execution of the preset control strategy based on the importance score includes:

[0021] The importance score P1 corresponding to the energy consumption and the importance score P2 corresponding to the welding rod material consumption are acquired, and the preset control strategy is executed based on P1 and P2.

[0022] As a further scheme of the present application: the edge of the molten pool in the target image is extracted based on AI vision, which comprises:

[0023] The target image is input into an edge detection network, which directly outputs a pixel-level edge probability map through multi-scale feature fusion and deep supervision;

[0024] The pixel-level edge probability map is processed, which comprises thresholding and non-maximum suppression, and removes low confidence and redundant responses;

[0025] Based on the morphological closing operation, the broken edges in the processed pixel-level edge probability map are filled, and the molten pool edge pixel set is screened out using connected component analysis, and the molten pool edge pixel set is smoothed and simplified by a polygon fitting algorithm to obtain the edge of the molten pool.

[0026] As a further scheme of the present application: the actual width A1 of the molten pool is obtained based on the edge, which comprises:

[0027] The actual width Z1 of a predetermined reference object is obtained, and the width Z2 of the reference object in the target image is obtained;

[0028] The maximum width z1 of the molten pool in the target image is obtained, and the actual width A1=(Z1 / Z2)*z1 is calculated.

[0029] The present application has the following advantages:

[0030] 1) The present application introduces a multi-scale visual edge detection network, thresholding and non-maximum suppression, and combines morphological closing operation and connected component analysis to realize millisecond-level molten pool edge capture and width-depth ratio calculation in the welding process; the analysis module compares the real-time results with the safety interval, and deviates from the normal position to trigger an early warning, avoiding the lag and error of traditional offline detection or manual visual inspection, significantly improving the consistency of weld formation, process stability and quality traceability, reducing the rework and scrap rate, and adapting to different material and posture welding scenes without frequent calibration, shortening the production line switching time, and laying a high-credibility data foundation for subsequent energy and material optimization;

[0031] 2) In the optimization module, this invention employs a hierarchical sample set and random forest regression to perform multidimensional correlation learning with process parameters such as aspect ratio, welding speed, welding current, energy consumption, and electrode material consumption, and dynamically calculate the importance of features. This mechanism can update the model weights in real time with the production cycle, automatically identify key variables affecting weld stability and energy consumption, and drive the control module to prioritize adjusting the welding current or electrode movement speed, avoiding trial and error caused by empirical adjustments. Through a continuous iterative machine learning process, the system simultaneously suppresses energy and material waste while ensuring weld quality, and provides transparent decision-making basis for operation and maintenance personnel with a visual score, facilitating rapid optimization of the process window for different batches of materials.

[0032] 3) Based on the threshold relationship of energy and material importance scores, this invention constructs a hierarchical closed-loop control strategy: when energy consumption is dominant, the welding current is automatically fine-tuned; when material consumption is prominent, the electrode movement speed is automatically fine-tuned; if both are in extreme states, the management personnel are notified immediately. This adaptive adjustment process forms a closed loop with real-time width-to-depth ratio monitoring, continuously iterating until the width and depth of the molten pool fall back into the safe range, achieving a dynamic balance between energy consumption and material usage. Compared with fixed formulas or manual intervention schemes, this invention can improve energy utilization and extend the service life of welding materials without sacrificing weld strength and formation. It also ensures timely handling of abnormal working conditions through digital alarms, supporting long-term stable operation in flexible production and intelligent manufacturing scenarios. Attached Figure Description

[0033] The invention will now be further described with reference to the accompanying drawings.

[0034] Figure 1 This is a schematic diagram of the structure of the welding robot control system based on AI vision of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Please see Figure 1 As shown, this invention is an AI vision-based welding robot control system, comprising: a data acquisition module, an analysis module, an optimization module, and a control module, specifically:

[0037] Acquisition module: Acquires molten pool image, preprocesses the molten pool image to obtain target image, the preprocessing includes denoising and enhancement, and extracts the edge of molten pool in the target image based on AI vision;

[0038] In a preferred embodiment of the present application, the edge of the molten pool in the target image is extracted based on AI vision, which comprises:

[0039] The target image is input into an edge detection network, which directly outputs a pixel-level edge probability map through multi-scale feature fusion and deep supervision;

[0040] The pixel-level edge probability map is processed, and the processing includes thresholding and non-maximum suppression to remove low confidence and redundant responses;

[0041] Based on a morphological closing operation, the broken edges in the processed pixel-level edge probability map are filled, and a molten pool edge pixel set is screened out using connected component analysis, and the molten pool edge pixel set is smoothed and simplified through a polygon fitting algorithm to obtain the edge of the molten pool;

[0042] It should be noted that the molten pool image was acquired in real time by a high-temperature resistant high-speed industrial camera at an angle of approximately 45 degrees between the welding torch and the weld seam. A narrow-band filter and quartz protective glass were added to the front of the camera to shield against arc spatter, and a hardware trigger signal synchronized with the welding machine ensured that a clear frame was obtained between each arc pulse. The raw BayerRAW data was first de-mosaiced and linearly de-gammaed, then converted into a 16-bit grayscale image with a timestamp for subsequent traceability. In the preprocessing stage, a 3x3 median filter was first used to reduce the salt-and-pepper noise generated by spatter particles, and then a moving average was applied to five adjacent frames to suppress random brightness jitter. If stripe noise appeared, band-stop filtering was performed on the corresponding peak in the frequency domain. After noise reduction, contrast-limited adaptive histogram equalization was performed in the brightness channel to improve details, and gamma correction (parameter less than one) was applied to the grayscale values ​​to enhance the dark area level, finally obtaining a target image with balanced contrast and clear edges. Then, the target image was fed into an improved overall nested edge detection network. The network's encoding end inserted dilated convolutions in shallow layers to maintain resolution, and the decoding end used pyramid feature fusion to converge the edges. Texture and semantic information at different scales are processed, and depth supervision is applied to each level of output to prevent gradient vanishing, ultimately generating a pixel-level edge probability map of the same size as the input. An adaptive threshold segmentation method is then employed, such as using the Oswald method to automatically determine the threshold based on the bimodal position of the histogram, binarizing the probability map to retain responses with high confidence, and then sliding a three-pixel window along the gradient direction to implement non-maximum suppression, retaining only the pixels with local maxima to form a one-pixel-wide thin edge. The resulting thin edge often breaks due to high-frequency arc flicker, so a morphological closing operation of dilation and erosion is performed on a 3x3 elliptical structuring element to seal the gaps. Next, connected component analysis is used to statistically analyze the area and centroid position of each connected component, selecting the connected component with the largest area located in the lower half of the image as the candidate region for the melt pool. For the edge pixel set of the candidate region, the Lamer-Douglas-Pike algorithm is used to simplify the point set according to a set tolerance, smoothing continuous polylines into a small number of control vertices, and using Bezier interpolation between vertices to generate smooth polygonal curves, ultimately obtaining the melt pool edge, providing an accurate geometric contour for subsequent size calibration and aspect ratio calculation.

[0043] Analysis module: Based on the edge, obtain the actual width A1 of the molten pool and the actual depth A2 of the molten pool, calculate the width-to-depth ratio D1 = A1 / A2, and mark it as an abnormal situation when the width-to-depth ratio does not belong to the preset safety range;

[0044] And it is worth noting that by geometrically quantifying the edge of the molten pool to obtain the actual width and depth and calculate the width-depth ratio, the welding transient heat input and the metal melting state can be condensed into a single dimensionless index, so that workpieces of different thicknesses and different groove forms can be evaluated in real time on a unified scale; when the width-depth ratio deviates from the preset safe interval, it is immediately marked as abnormal, which can prompt process instability before defects are formed, thereby providing a clear trigger point for subsequent energy or material regulation; this approach not only avoids misjudgment caused by focusing on a single size, but also reduces the dependence on complex acoustic and optical detection equipment, helping the system to closely couple weld seam forming quality monitoring, energy consumption optimization and material saving with simple and intuitive criteria, and achieve the unified goal of high quality and green manufacturing.

[0045] In another preferred embodiment of the present application, the actual width A1 of the molten pool is obtained based on the edge, which includes:

[0046] The actual width Z1 of a preset reference object is obtained, and the width Z2 of the reference object in the target image is obtained;

[0047] The maximum width z1 of the molten pool in the target image is obtained, and the actual width A1 is calculated as (Z1 / Z2)*z1;

[0048] It should be noted that after the camera is installed, a high-temperature-resistant calibration piece with a known width is first fixed near the weld seam and on the same plane, and the true width is recorded as Z1. After the image enters the processing link, it is corrected for distortion and sub-pixel interpolation. After gray threshold segmentation, the edges of the two sides of the calibration piece are detected using Hough line transformation, and the horizontal distance of the two edges in the pixel coordinate system is calculated to obtain Z2. Then, the polygon fitting result is called in the same frame to extract all pixel points of the molten pool edge, and the maximum horizontal span z1 is obtained by traversing the horizontal coordinates. The sub-pixel error at the endpoints is corrected using spline interpolation. Finally, the actual pixel width is enlarged by the proportional coefficient Z1 divided by Z2, and this value will be written into the database for continuous updating by the analysis module.

[0049] Optimization module: when an abnormal situation occurs, real-time collection of features is performed within a preset collection period, the features including the width-depth ratio, welding speed, welding current, energy consumption, and electrode material consumption, two sample sets are constructed based on the features, and importance scores of the features are obtained based on the sample sets;

[0050] In another preferred embodiment of the present application, constructing two sample sets based on the features includes:

[0051] The welding speed, welding current and energy consumption are taken as the first initial sample, and the welding speed, welding current and electrode material consumption are taken as the second initial sample;

[0052] A first sample is obtained by taking the width-depth ratio as the label of the first initial sample, and a first sample set is generated by obtaining first samples corresponding to consecutive n time points, where n is a preset number;

[0053] A second sample is obtained by taking the width-depth ratio as the label of the second initial sample, and a second sample set is generated by obtaining second samples corresponding to consecutive n time points;

[0054] It can be understood that first, the welding speed, welding current and energy consumption measured in the same frame are spliced into a group of three-dimensional vectors in chronological order, which is recorded as a first initial sample, and the welding speed, welding current and electrode material consumption at the same time stamp are spliced into another group of three-dimensional vectors, which is recorded as a second initial sample; then the width-depth ratio calculated in the same frame is appended to the end of the first initial sample as a supervised label of the sample, forming a complete first sample, and similarly, the same width-depth ratio is appended to the second initial sample to obtain a second sample; the data acquisition thread continuously writes to the ring buffer at a fixed interval, and the algorithm thread opens a sliding window, which traces back to consecutive n time points from the latest data, and sequentially extracts n groups of first samples to form a first sample set, and n groups of second samples to form a second sample set, if there are missing values in the window, they are completed by linear interpolation to ensure the consistency of the sample dimension, finally the two sample sets are strictly aligned on the time axis, which can be directly input into the subsequent random forest model for training and feature weight calculation;

[0055] In a preferred case of the embodiment, obtaining the importance score of the feature based on the sample set includes:

[0056] Taking the width-depth ratio as the regression target, the random forest model is trained using the first sample set, the purity improvement value when the feature is used as a node for splitting is calculated, the purity improvement values of the same feature are summed to obtain a purity improvement sum, and the purity improvement sum is normalized to obtain the importance score;

[0057] Taking the width-depth ratio as the regression target, the random forest model is trained using the second sample set, the purity improvement value when the feature is used as a node for splitting is calculated, the purity improvement values of the same feature are summed to obtain a purity improvement sum, and the purity improvement sum is normalized to obtain the importance score;

[0058] Exemplarily, the first sample set is randomly divided into an eight-to-two training set and a validation set, a random forest containing one hundred sub-trees is constructed on the training set, the maximum depth of each sub-tree is set to ten, and the minimum leaf node sample number is set to two, the error is estimated by out-of-bag data, and overfitting is automatically stopped; during the model training process, when a certain feature is used for node splitting, the reduction of the square sum of the prediction residual width-depth ratio before and after the splitting can be calculated, which is the purity improvement value, and the purity improvement values of the same feature on all nodes are accumulated to obtain the purity improvement total; after all the sub-trees of the forest are trained, the purity improvement totals of the three features are normalized, and the sum of the totals is equal to one, i.e., the relative importance scores of the welding speed, the welding current, and the energy consumption under the energy perspective are obtained; then, the above process is repeated for the second sample set, only the energy consumption feature is replaced by the electrode material consumption, and other parameters remain the same, and the purity improvement totals are also summarized and normalized, so as to obtain the importance scores of the features under the material perspective; the system stores the two sets of scores in the buffer area for real-time calling and driving the closed-loop adjustment of the welding current or the electrode moving speed by the threshold strategy.

[0059] The control module: based on the importance score, a preset control strategy is executed, the control strategy includes an energy control strategy and a material control strategy, the energy control strategy is used to control the current of the welding robot to reduce the energy consumption, and the material control strategy is used to control the electrode moving speed of the welding robot to reduce the electrode material consumption.

[0060] It should be noted that dynamically selecting the energy control strategy or the material control strategy according to the real-time calculated feature importance score can make the system always accurately adjust the most critical resource waste point in the current production stage; when the importance of energy consumption is higher than the threshold, the welding current is finely adjusted to change the heat input with the smallest amplitude, so as to significantly inhibit the power surplus without sacrificing the weld quality; when the electrode material consumption is dominant, the electrode moving speed is finely adjusted to reduce the excess deposition and avoid material accumulation causing subsequent polishing and rework; if both are in atypical state, it is immediately reported to the human, so that the abnormal scene can be intervened in time; this shunt closed-loop strategy closely combines data-driven judgment with process-level fine adjustment, so as to form a unified logic chain of energy and consumable optimization and weld forming quality monitoring, thereby ensuring that the production process meets the green and low-carbon requirements and maintains high consistency of weld performance, and achieving the comprehensive goal of intelligent manufacturing.

[0061] It should be noted that executing a preset control strategy based on the importance score includes:

[0062] The importance score P1 corresponding to the energy consumption and the importance score P2 corresponding to the electrode material consumption are obtained, and a preset control strategy is executed based on P1 and P2.

[0063] The specific process includes:

[0064] When P1>0.6>P2, the energy control strategy is executed:

[0065] The welding current I1at this time is obtained, a new welding current I1'=I1-ΔI is calculated, ΔI is a preset welding current adjustment value, the welding current of the welding robot is adjusted to I1', and the width-depth ratio is obtained again, and the above steps are repeated until the width-depth ratio after adjusting the welding current at a certain time belongs to the safe range;

[0066] When P1<0.6<P2, the material control strategy is executed:

[0067] The electrode moving speed V1at this time is obtained, a new electrode moving speed V1'=V1-ΔV is calculated, ΔV is a preset electrode moving speed adjustment value, the electrode moving speed of the welding robot is adjusted to V1', and the width-depth ratio is obtained again, and the above steps are repeated until the width-depth ratio after adjusting the electrode moving speed at a certain time belongs to the safe range;

[0068] In addition to the above two cases, a prompt message is sent to a preset manager.

[0069] It is worth noting that by setting the importance score threshold, the influence degree of energy consumption and electrode material consumption is intuitively quantified, and the system can immediately focus on the most effective adjustment means when a single resource dominates: when the energy importance is prominent, the welding current is slightly reduced to quickly reduce the excessive heat input, thereby reducing the energy consumption of the welding robot, and when the material importance is high, the electrode moving speed is finely adjusted to reduce the excessive deposition, and the electrode material consumption of the welding robot, both strategies gradually converge with the width-depth ratio monitored in real time as a feedback signal, ensuring that the process parameters do not change excessively and eventually return to the safe range; if the two scores are at the critical or extreme state, the system automatically switches to the manual intervention mode to avoid misjudgment. This hierarchical closed-loop mechanism ensures the stability of the weld forming quality while controlling energy consumption and material loss in a unified logic chain, providing reliable protection for the intelligent and low-carbon welding production goal.

[0070] The above describes one embodiment of the present application in detail, but the content described is only a preferred embodiment of the present application and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made within the scope of the present application should still belong to the patent coverage of the present application.

Claims

1. An AI vision-based welding robot control system, characterized by, The system comprises a collection module, an analysis module, an optimization module and a control module, specifically: The collection module collects a molten pool image, pre-processes the molten pool image to obtain a target image, the pre-processing includes denoising and enhancement, extracts the edge of the molten pool in the target image based on AI vision; The analysis module obtains the actual width A1 of the molten pool based on the edge, and obtains the actual depth A2 of the molten pool, calculates the width-depth ratio D1=A1 / A2, and marks it as an abnormal situation when the width-depth ratio is not in the preset safe range; The optimization module: when an abnormal situation occurs, real-time collection of features within a preset collection period, the features include width-depth ratio, welding speed, welding current, energy consumption and electrode material consumption, based on the features to construct two sample sets, based on the sample set to obtain the importance score of the features; The control module: based on the importance score to execute the preset control strategy, the control strategy includes energy control strategy and material control strategy, the energy control strategy is used to control the current of the welding robot to reduce the energy consumption, and the material control strategy is used to control the electrode moving speed of the welding robot to reduce the electrode material consumption; Based on the features to construct two sample sets, including: Taking the welding speed, welding current and energy consumption as the first initial sample, and taking the welding speed, welding current and electrode material consumption as the second initial sample; Taking the width-depth ratio as the label of the first initial sample to obtain a first sample, obtaining the first sample corresponding to the continuous n time points to generate a first sample set, n is a preset number; Taking the width-depth ratio as the label of the second initial sample to obtain a second sample, obtaining the second sample corresponding to the continuous n time points to generate a second sample set; Based on the sample set to obtain the importance score of the features, including: Taking the width-depth ratio as the regression target, training a random forest model using the first sample set, calculating the purity improvement value when the feature is used as a node to split, summing the purity improvement values of the same feature to obtain a purity improvement total sum, normalizing the purity improvement total sum to obtain the importance score; Taking the width-depth ratio as the regression target, training a random forest model using the second sample set, calculating the purity improvement value when the feature is used as a node to split, summing the purity improvement values of the same feature to obtain a purity improvement total sum, normalizing the purity improvement total sum to obtain the importance score; Based on the importance score to execute the preset control strategy, including: Obtaining the importance score P1 corresponding to the energy consumption and the importance score P2 corresponding to the electrode material consumption, based on P1 and P2 to execute the preset control strategy; The specific process includes: When P1>0.6>P2, execute the energy control strategy: Obtain the welding current I1 at this time, calculate the new welding current I1'=I1-ΔI, ΔI is a preset welding current adjustment value, adjust the welding current of the welding robot to I1', and obtain the width-depth ratio again, repeat the above steps until the width-depth ratio after adjusting the welding current at a certain time belongs to the safe range; When P1 < 0.6 < P2, the material control strategy is executed: The current electrode moving speed V1 is obtained, a new electrode moving speed V1'=V1-ΔV is calculated, ΔV is a preset electrode moving speed adjustment value, the electrode moving speed of the welding robot is adjusted to V1', and the width-depth ratio is obtained again, and the above steps are repeated until the width-depth ratio after adjusting the electrode moving speed in a certain time belongs to the safety range; In addition to the above two cases, a prompt message is sent to a preset manager.

2. The AI vision-based welding robot control system of claim 1, wherein, The edge of the molten pool in the target image is extracted based on AI vision, which includes: The target image is input into an edge detection network, the edge detection network directly outputs a pixel-level edge probability map through multi-scale feature fusion and deep supervision; The pixel-level edge probability map is processed, and the processing includes thresholding and non-maximum suppression to remove low confidence and redundant responses; Based on the morphological closing operation, the broken edges in the processed pixel-level edge probability map are filled, the molten pool edge pixel set is screened out by using connected component analysis, and the molten pool edge pixel set is smoothed and simplified by using a polygon fitting algorithm to obtain the edge of the molten pool.

3. The AI vision-based welding robot control system of claim 1, wherein, The actual width A1 of the molten pool is obtained based on the edge, which includes: The actual width Z1 of a preset reference object is obtained, and the width Z2 of the reference object in the target image is obtained; The maximum width z1 of the molten pool in the target image is obtained, and the actual width A1= (Z1 / Z2)*z1 is calculated.

Images