A welding robot welding spot image processing method based on a vision sensor

Abstract

The application discloses a welding robot welding spot image processing method based on a visual sensor, relates to the technical field of machine vision and welding automation, and comprises four steps: laser wavelength dynamic matching imaging, wavelength adaptive switching is realized through slag thickness and welding layer number correlation, and image definition is improved in combination with double-domain enhancement; layered feature accurate extraction and fusion, region is dynamically divided based on welding current, interlayer features are correlated through forming consistency coefficients, and cross-layer defect accurate tracking is realized; three-dimensional defect reconstruction and quantization, layered feature constraints are introduced, and binocular vision is combined to restore defect stereoscopic morphology; multi-parameter fusion quality evaluation, a hazard degree model is constructed through a position factor and material performance. The method breaks through the limitations of traditional fixed wavelength, fixed region division and single-dimensional evaluation, improves defect detection rate, positioning accuracy and evaluation scientificity, and provides support for high-pressure container welding quality control.

Description

Technical Field

[0001] The present invention relates to the technical field of machine vision and welding automation, and in particular to a welding robot welding spot image processing method based on a vision sensor. Background Art

[0002] Welding robots are widely used in modern industrial manufacturing, and weld spot image processing is a key component in automated welding quality monitoring. For thick plate welding, such as steel pressure vessels, traditional image processing methods based on vision sensors face numerous challenges due to the numerous weld layers, thick slag layers, and hidden defects. These challenges include image quality degradation due to arc interference, and difficulty accurately extracting multi-layer weld feature aliasing, which directly impacts the accuracy and efficiency of defect detection.

[0003] Existing welding image processing techniques often use single-wavelength imaging, which is unable to dynamically adjust penetration based on slag thickness. This results in low defect detection rates in mid- and deep welds. Furthermore, layered feature extraction often relies on fixed region partitioning, failing to incorporate dynamic parameters such as welding current. This leads to large errors in cross-layer defect location. Furthermore, traditional quality assessment relies solely on defect size, ignoring the impact of location and material properties on hazard risk. This can easily lead to over-repair or under-assessment, making it difficult to meet quality control requirements for equipment such as high-pressure vessels.

[0004] In view of this, this application is hereby filed. Summary of the Invention

[0005] The purpose of the present invention is to provide a welding robot welding spot image processing method based on a visual sensor to solve the problems mentioned in the above background technology.

[0006] To solve the above technical problems, the present invention provides a method for processing welding spot images of a welding robot based on a visual sensor, comprising the following steps:

[0007] Step 1: Dynamic matching of laser wavelength and imaging based on slag thickness estimation and welding layers The corresponding relationship dynamically selects the laser wavelength ,in , is the slag formation rate coefficient, is the single-layer welding time; the acquired image is subjected to dual-domain enhancement processing, and adaptive histogram equalization is used in the spatial domain. The formula is:

[0008] ;

[0009] in is the transformation function, Grayscale The number of pixels, is the total number of pixels in the image; the frequency domain uses a Butterworth high-pass filter, and the transfer function is:

[0010] ;

[0011] in is the cutoff frequency, for point The distance to the origin of the frequency plane, is the order;

[0012] Step 2: Accurately extract and fuse layered features. Divide the layered areas according to the welding current signal, extract the basic features, defect features and inter-layer correlation features of each layer, and the inter-layer correlation features are calculated through the forming consistency coefficient:

[0013] calculate, is the melt width, For Yu Gao;

[0014] Step 3: 3D defect reconstruction and quantification. A binocular vision system is used for calibration and matching. The spatial coordinates are calculated using the disparity map. The 3D defect model is constructed and the volume is calculated by combining hierarchical feature constraints.

[0015] Step 4: Multi-parameter fusion quality assessment and calculation of defect location factors , combined with the defect volume , material yield strength , through the hazard index :

[0016] ;

[0017] Assess the quality level, Critical volume, is the base yield strength, 、 is the correction coefficient; through the coordination of dynamic matching of laser wavelength, layered feature extraction, three-dimensional reconstruction and multi-parameter evaluation, high-precision detection and evaluation of thick plate multi-layer welding defects are achieved. Compared with the existing technology, the detection rate of deep defects is improved, and the accuracy of quality assessment is improved.

[0018] Furthermore, in step 1, the laser wavelength The selection rule is: and Layer, ;when and Layer, ;when and Layer, , wavelength switching delay time <7ms; ensuring image clarity under different slag thicknesses, improving the detection rate of middle and deep defects, and providing high-quality data for subsequent processing.

[0019] Furthermore, in step 1, the multi-wavelength imaging system includes 4 laser emitting units, a high-resolution camera, a spectrometer and a synchronization control unit. The angle between the laser emitting unit and the camera optical axis ranges from 36° to 55°. It adapts to the imaging requirements of different welding scenarios, improves the signal-to-noise ratio of each layer of weld image, and improves the retention of defect details.

[0020] Furthermore, in step 2, the dynamic division of the layered area is specifically as follows: collecting the welding current signal , determine the starting time of each layer and end time , according to welding speed Calculate the Layer vertical range , which divides the stratified areas ; Achieve accurate separation of defects in each layer, reduce cross-layer defect positioning errors, and improve positioning accuracy.

[0021] Furthermore, in step 2, the defect characteristics include the defect area ,perimeter , minimum enclosing rectangle aspect ratio , For the long side, is the short side; cross-layer defects are matched by feature matrix track, For the Layer and The similarity of layer defect features is determined by using the Hungarian algorithm to find the optimal matching path; the continuity of cross-layer defect tracking reaches 100%, the recognition rate of forming anomalies is improved, and accurate constraints are provided for 3D reconstruction.

[0022] Furthermore, in step 2, the penetration depth of the basic feature Measured by laser displacement sensor, , is the laser spot offset, The laser incident angle and the penetration depth measurement accuracy are improved, providing reliable data for the calculation of interlayer correlation features and improving the accuracy of the forming consistency coefficient evaluation.

[0023] Furthermore, in step 3, the stereo matching adopts the feature-based SGBM algorithm, and the matching cost function is C(x, y) = w1·C color +w2·C grad +w3·C laye , w1+w2+w3=1; where C coloris the color cost, C grad is the gradient cost, is the hierarchical constraint cost, w1, w2, and w3 are all weight coefficients; the point cloud matching error rate is reduced, the 3D reconstruction accuracy is improved, and the defect volume assessment error is reduced.

[0024] Furthermore, in step 4, the defect location factor The stress level at the defect location is reflected, the fusion line is determined through the weld formation model, and the stress concentration area is marked according to the finite element simulation; considering the impact of the defect location on the hazard, the accuracy of the hazard determination is improved, and the detection rate of small defects in high-risk areas is improved.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The laser wavelength dynamic matching imaging technology breaks through the limitations of traditional fixed wavelengths. The wavelength adaptive switching is achieved by correlating the slag thickness with the number of welding layers. Combined with dual-domain enhancement processing, the clarity of the middle and deep weld images is significantly improved, solving the problem of defect detection caused by slag obstruction in thick plate welding, and laying a high-quality data foundation for subsequent feature extraction.

[0027] 2. The precise extraction and fusion technology of layered features breaks the traditional thinking of fixed area division. It dynamically divides the layered areas based on the welding current signal and associates the inter-layer features through the forming consistency coefficient, thus achieving precise tracking and positioning of cross-layer defects, overcoming the positioning error problem caused by the aliasing of multi-layer features, and providing precise feature constraints for three-dimensional reconstruction.

[0028] 3. The 3D defect reconstruction and quantification technology introduces hierarchical feature constraints and combines them with the high-precision calibration and matching of the binocular vision system. This technology breaks through the limitations of large errors in monocular vision volume assessment and achieves accurate restoration of the spatial morphology of defects. It provides reliable 3D quantitative data for quality assessment and solves the problem that 2D images cannot reflect the 3D characteristics of defects.

[0029] 4. Multi-parameter fusion quality assessment technology overcomes the technical bias of evaluating defects based solely on defect size. It quantifies the stress levels in different areas through position factors and constructs a hazard model based on material properties, thus achieving scientific determination of quality levels, reducing over-repair or under-assessment, effectively reducing container operation risks, and improving production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The figure is a flow chart of a welding robot welding spot image processing method based on a visual sensor. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] See also Figure 1 This invention provides a technical solution: a method for processing weld spot images for a welding robot based on a visual sensor. In this embodiment, the quality inspection requirement for multi-layer, multi-pass welding of steel pressure vessels (plate thickness 12mm-55mm, 6-22 layers) is met. The slag layer formed by thick plate welding in this scenario can be as thick as 0.6mm-2.8mm. Traditional single-wavelength imaging technology cannot penetrate the slag beyond 8 layers, resulting in an internal defect detection rate below 68%. By combining dynamic laser wavelength matching, precise layered feature extraction, 3D reconstruction quantification, and multi-parameter quality assessment, high-precision monitoring of the entire welding process, from the welding process to quality assessment, is achieved.

[0033] Step 1: Dynamic matching of laser wavelength for imaging: When welding thick steel pressure vessel plates, each weld produces a 0.4mm-0.9mm slag layer. As the number of weld layers increases, the cumulative slag thickness increases, and the penetration capabilities of lasers of different wavelengths vary significantly. Based on the correspondence between slag thickness and the number of weld layers, dynamic selection of laser wavelength can ensure the clarity of each weld image and provide reliable data for subsequent feature extraction. We can use the following methods:

[0034] 1. The multi-wavelength imaging system consists of four laser emission units (532nm green, 650nm red, 980nm near-infrared, and 1310nm infrared), a high-resolution camera (3840×2160 pixels, 50fps), a spectrometer, and a synchronization control unit. The angle between the laser emission unit and the camera's optical axis ranges from 36° to 55°, and the spectrometer monitors the spectral absorption characteristics of the slag layer in real time.

[0035] 2. Laser wavelength dynamic switching rules: define the estimated slag thickness ,in is the slag generation rate coefficient (0.08mm / s-0.25mm / s), This is the single layer welding time. Combined with the number of welding layers and The corresponding relationship determines the laser wavelength :

[0036] when ( layer), select (green light), suitable for penetrating shallow thin slag;

[0037] when ( layer), select (Near infrared), using its low absorption rate to metal oxides to penetrate the middle layer of slag;

[0038] when ( layer), select (Infrared), deep slag penetration is achieved through differences in molecular vibration absorption.

[0039] The wavelength switching is triggered by the synchronization control unit, and the switching delay time is <7ms, ensuring synchronization with the number of welding layers.

[0040] 3. Dual-domain image enhancement processing

[0041] 3.1 Spatial domain enhancement: Adaptive histogram equalization is used, the formula is:

[0042] ;

[0043] in is the transformation function, Grayscale The number of pixels, is the total number of pixels in the image, and details are enhanced by adjusting the local grayscale distribution;

[0044] Frequency domain enhancement: After Fourier transform of the image, a Butterworth high-pass filter is used to suppress low-frequency noise. Its transfer function is:

[0045] ;

[0046] in is the cutoff frequency, for point The distance to the origin of the frequency plane, is the order, and the cut-off frequency is dynamically adjusted with the arc intensity.

[0047] Here we give an example: In the Q345R steel pressure vessel cylinder girth welding, welding to the 10th layer ( ), the system automatically switches to a 980nm near-infrared laser. After dual-domain enhancement processing, the grayscale gradient at the melt pool boundary increases from 18 grayscale levels per pixel to 62 grayscale levels per pixel. The edge continuity of an interlayer lack of fusion defect (0.28mm long, 0.14mm wide) reaches 88%. Conventional single-wavelength 650nm imaging only reveals 28% of the edge of this defect.

[0048] Existing public documents use fixed 650nm laser imaging. When the slag thickness exceeds 1mm, the penetration rate drops below 28%, and the deep defect detection rate is only 53%. This technical solution dynamically matches the slag thickness and wavelength. The core difference is that it breaks through the technical limitation of "laser wavelength is independent of the number of welding layers" and establishes The mapping model solves the challenge of detecting medium- and deep-seated defects caused by slag layers (>0.9mm) in welds thicker than 12mm. The deep-seated defect detection rate is 38% higher than existing technologies, and the maximum slag penetration depth can reach 3.8mm.

[0049] Compared with the single-wavelength laser imaging used by existing technical means, the signal-to-noise ratio of each layer of weld images is improved by 45%-65%, and the detection rate of middle and deep layer defects is increased from 53% to 94%, providing clear image data for layered feature extraction, and the defect detail retention is improved by 58%.

[0050] Step 2: Accurate extraction and fusion of layered features: In multi-layer, multi-pass welding, the forming parameters of each weld layer (weld width, excess height) fluctuate by 2%-7%. Traditional overall feature extraction will confuse the defect information of different layers, resulting in a positioning error of 0.5mm-1.1mm for cross-layer defects. Layered feature extraction based on the welding current-layer number correlation model can achieve accurate separation and feature fusion of each layer's defects, improving defect positioning accuracy. We can use the following technical means:

[0051] 1. Dynamic division of hierarchical areas

[0052] Collect welding current signals , determine the starting time of each layer by detecting the current peak and end time , No. The time interval of the layer is . According to the welding speed , calculate the The longitudinal extent of the layer weld in the image: ;

[0053] in For the The y coordinate interval of the layer in the image coordinate system, which is used to divide the layer area .

[0054] 2. Multi-layer feature extraction and fusion

[0055] Basic features: Internal extraction melt width , Yu Gao , penetration (Measured by laser displacement sensor, , is the laser spot offset, is the laser incident angle);

[0056] Defect characteristics: Use the improved Canny operator to extract defect edges and calculate defect area ,perimeter , minimum enclosing rectangle aspect ratio ( For the long side, is the short side);

[0057] Inter-layer correlation features: Calculate the Layer and Layer forming consistency coefficient:

[0058] ;

[0059] The closer to 1, the better the consistency. It is judged as forming abnormality.

[0060] 3. Cross-layer defect tracking

[0061] Cross layer( ) defects, establish a feature matching matrix ,in For the Layer and The similarity of layer defect features is determined, and the optimal matching path is found through the Hungarian algorithm to achieve defect trajectory tracking between multiple layers.

[0062] Here is an example: In the welding of the connection weld between the head and the shell of a steel pressure vessel, the weld width of the 8th and 9th layers are 13.5mm and 13.3mm respectively, and the residual height is 2.4mm and 2.3mm respectively. The calculated value is (Good consistency). A cross-layer pore was detected in the 7th to 9th layers. The defect area of ​​each layer was 0.038mm through feature matching matrix tracking. 2 , 0.068mm 2 , 0.028mm 2 The positioning error is reduced from 0.85mm in traditional overall extraction to 0.075mm.

[0063] Existing technologies use fixed region division to extract features, failing to consider the correlation between welding current and number of layers. This can lead to positioning errors exceeding 0.75mm for cross-layer defects. This technology dynamically divides regions based on current and number of layers. Its key difference lies in breaking away from the traditional assumption that layered regions are fixed. Instead, it uses the welding current signal as the basis for layering, resolving the issue of inaccurate tracking and positioning of cross-layer defects (spanning two or more layers) in multi-layer, multi-pass welds. It is important to note that this technology reduces the positioning error of cross-layer defects by 84% compared to existing technologies, and achieves a 91% accuracy rate for identifying forming anomalies.

[0064] We can clearly see that compared with the overall feature extraction method, the positioning accuracy of layered defects is improved to 0.09mm, the tracking continuity of cross-layer defects reaches 100%, and the recognition rate of forming anomalies is increased from 54% to 91%, providing accurate layered feature constraints for three-dimensional reconstruction.

[0065] Step 3: 3D Defect Reconstruction and Quantification: Steel pressure vessel welding defects are mostly three-dimensional structures. Two-dimensional images can only reflect their surface projections, resulting in volumetric assessment errors of 33%-58%, which cannot meet the stringent defect quantification requirements for high-pressure vessels. Using binocular vision and hierarchical feature-constrained 3D reconstruction technology, we can accurately restore the spatial morphology of defects and calculate their volume. Therefore, we designed the following:

[0066] 1. Binocular vision system calibration and matching: Using the improved Zhang Zhengyou calibration method, circular marking points are added to the calibration plate, and the calibration accuracy is improved through sub-pixel corner detection. The error of the intrinsic parameter matrix is ​​controlled within 0.45%, and the external parameter matrix is The rotation angle error is <0.09° and the translation error is <0.045mm.

[0067] Stereo matching uses the feature-based SGBM algorithm and introduces hierarchical feature constraints: Layer Area The pixels within are only matched with the corresponding areas of the adjacent layers, and the matching cost function is:

[0068] C(x,y)=w1·C color +w2·C grad +w3·C laye ;

[0069] Among them C color is the color cost, C grad is the gradient cost, is the hierarchical constraint cost, w1+w2+w3=1, w1, w2, w3 are all weight coefficients, and the matching robustness is enhanced by weight adjustment.

[0070] 2. 3D coordinate transformation and point cloud optimization

[0071] According to the disparity map Calculate spatial coordinates:

[0072] ;

[0073] in is the binocular baseline distance, is the camera focal length, are image coordinates, is the spatial coordinate.

[0074] The generated point cloud is filtered and optimized: statistical filtering is used to remove outliers, and radius filtering is used to smooth the point cloud surface and retain the detailed features of the defect edge.

[0075] 3. Layered Constrained 3D Reconstruction

[0076] The delamination defect features extracted in step 2 ( , , ), as a constraint condition, construct the defect three-dimensional model:

[0077] For the first Layer defects, according to Determine the point cloud range within this layer;

[0078] Combine (Distance between defect and interlayer interface) Determine the point cloud at Position in the axial direction;

[0079] Generate defect 3D mesh by Poisson surface reconstruction algorithm and calculate volume :

[0080] ;

[0081] in For the The spatial region of layer defects.

[0082] Let's take an example as an example: In the welding of the fillet weld of a 20MnMo steel pressure vessel, a span of four layers of internal slag was detected. After 3D reconstruction, the volume of each layer was 0.032mm 3 , 0.078mm 3 , 0.058mm 3 , 0.022mm 3 , total volume , the spatial coordinates are The pore volume of the traditional two-dimensional evaluation is 0.108mm 3 , the error is 43%, the reconstruction result of this technology is different from the actual anatomical measurement value (0.186mm 3 ) with an error of only 2.1%.

[0083] Therefore, based on existing publicly available documentation of monocular vision-based 3D reconstruction without the use of layered feature constraints, the point cloud matching error rate reached 24%, and the volume estimation error exceeded 48%. This technology, through binocular vision and layered feature constraints, differs from existing technologies by using layered features as reconstruction constraints, addressing the challenge of 3D shape restoration for cross-layer defects (those spanning four or more layers). This technology reduces the volume estimation error by 48% compared to existing technologies, and improves point cloud matching accuracy to 0.045mm.

[0084] Compared with monocular vision 3D reconstruction, the defect volume assessment error in this scheme is reduced from 48% to below 5%, the matching error rate of the 3D point cloud is reduced by 78%, and the spatial morphological restoration degree of cross-layer defects reaches 94%, providing accurate 3D quantitative data for quality assessment.

[0085] Step 4: Multi-parameter fusion quality assessment: Steel pressure vessels are pressure-bearing equipment, and the harmfulness of their welding defects is not only related to their size, but also closely related to the defect location (such as whether it is located at the fusion line or stress concentration area) and material properties. Traditional assessment methods based solely on size lead to 24%-38% misjudgment. An assessment model that integrates the three-dimensional characteristics of defects, location factors, and material mechanical properties can achieve scientific determination of quality grades. The following example illustrates this:

[0086] 1. Defect location factor calculation

[0087] Define the position factor , reflecting the stress level at the defect location:

[0088] ;

[0089] in, The defect is located at the fusion line or the preset stress concentration area. The defect is located in the center of the weld. The defect is located in the heat-affected zone. The fusion line position is determined by the weld formation model, and the stress concentration area is pre-marked according to the finite element simulation results.

[0090] 2. Defect severity assessment model

[0091] Comprehensive defect volume , location factor , material yield strength , calculate the hazard index :

[0092] ;

[0093] in is the critical volume of this type of defect (determined according to the vessel design pressure), is the benchmark yield strength (235MPa), the correction factor , , Larger values ​​indicate more severe hazards.

[0094] Quality grade determination and repair plan

[0095] according to Value determines the quality level:

[0096] Level Ⅰ (qualified): ;

[0097] Level II (for monitoring purposes): ;

[0098] Level III (needs repair): .

[0099] For level III defects, generate repair parameters: repair welding current , number of repair welds ,in is the normal welding current, is the filling volume of each repair weld, is the ceiling function.

[0100] Here is another example to illustrate that in the circumferential weld of a 15CrMoR steel pressure vessel, the volume of the pores at the fusion line is ( ), material yield strength , calculated as:

[0101] 0.293;

[0102] It is judged as qualified for level I. Another slag inclusion is located in the stress concentration area. , generate a repair plan: the repair welding current is increased from 195A to 212A, and the number of repair welding passes is 3.

[0103] This shows that existing public documents assess quality solely based on defect size (e.g., diameter > 0.55mm), without considering location and material properties. This results in 34% of qualified defects being mistakenly identified as requiring repair. This technology, through a multi-parameter fusion criticality model, breaks the technical bias that "defect hazard is determined solely by size" by incorporating location factors and material properties into the assessment system, resolving the challenge of determining significant differences in hazard for defects of the same size at different locations. This demonstrates that its quality assessment accuracy is 43% higher than existing technologies, while reducing the over-repair rate by 58%.

[0104] Therefore, compared with traditional dimensional assessment methods, the quality assessment accuracy of this technical means has been increased from 64% to 95%, the over-repair rate has been reduced from 34% to 13%, and the repair qualification rate of Level III defects has reached 100%, effectively reducing the risk of container operation, while reducing unnecessary repair welding operations and improving production efficiency.

[0105] To sum up: We can know that the four steps of this technical solution form a closed-loop system of "dynamic imaging-layered extraction-3D reconstruction-quality assessment". The collaborative mechanism of each step is as follows: the dynamic matching of laser wavelength in step 1 provides clear images of each layer for layered feature extraction; the layered feature extraction in step 2 provides precise constraints for 3D reconstruction; the 3D quantification in step 3 provides accurate defect parameters for quality assessment; the multi-parameter evaluation in step 4 provides a scientific basis for repair. The four steps support each other to achieve precise control of the quality of thick plate multi-layer welding.

[0106] Compared with the existing technology, the core of this method is reflected in:

[0107] Overcoming technological biases: Breaking through traditional perceptions such as "laser wavelength is fixed," "stratification is unrelated to current," and "evaluation is based solely on size," achieving technological breakthroughs through dynamic adaptation, precise extraction, and multi-parameter evaluation.

[0108] Solve long-standing challenges: Overcome three major industry pain points: slag penetration (>0.9mm) in welding plates thicker than 12mm, cross-layer defect location (error <0.09mm), and unreasonable quality assessment. These problems have existed in the pressure vessel manufacturing field for many years and have not been effectively solved;

[0109] The comprehensive defect detection rate is improved, the quality assessment accuracy is improved, the first-time pass rate of container welding is improved, and the annual rework cost is reduced, generating significant technical and economic value. This method provides key technical support for the safe manufacturing of steel pressure vessels and has broad engineering application prospects.

Claims

1. A welding robot welding point image processing method based on a visual sensor, characterized by: The following steps are involved: Step 1: Dynamic matching of laser wavelength and imaging based on slag thickness estimation and welding layers The corresponding relationship dynamically selects the laser wavelength ,in , is the slag formation rate coefficient, is the single-layer welding time; the acquired image is subjected to dual-domain enhancement processing, and adaptive histogram equalization is used in the spatial domain. The formula is: ； in is the transformation function, Grayscale The number of pixels, is the total number of pixels in the image; the frequency domain uses a Butterworth high-pass filter, and the transfer function is: ； in is the cutoff frequency, for point The distance to the origin of the frequency plane, is the order; Step 2: Accurately extract and fuse layered features. Divide the layered areas according to the welding current signal, extract the basic features, defect features and inter-layer correlation features of each layer, and the inter-layer correlation features are calculated through the forming consistency coefficient: calculate, is the melt width, For Yu Gao; Step 3: 3D defect reconstruction and quantification. A binocular vision system is used for calibration and matching. The spatial coordinates are calculated using the disparity map. The 3D defect model is constructed and the volume is calculated by combining hierarchical feature constraints. Step 4: Multi-parameter fusion quality assessment and calculation of defect location factors , combined with the defect volume , material yield strength , through the hazard index : ； Assess the quality level, Critical volume, is the base yield strength, 、 is the correction factor.

2. The method for processing welding spot images of a welding robot based on a visual sensor according to claim 1, characterized in that: In step 1, the laser wavelength The selection rule is: and Layer, ;when and Layer, ;when and Layer, , wavelength switching delay time is <7ms.

3. The method for processing welding spot images of a welding robot based on a visual sensor according to claim 1, wherein: In step 1, the multi-wavelength imaging system includes four laser emitting units, a high-resolution camera, a spectrometer and a synchronization control unit. The angle between the laser emitting unit and the camera optical axis ranges from 36° to 55°.

4. The method for processing welding spot images of a welding robot based on a visual sensor according to claim 1, wherein: In step 2, the dynamic division of the layered area is specifically as follows: collecting the welding current signal , determine the starting time of each layer and end time , according to welding speed Calculate the Layer vertical range , which divides the stratified areas .

5. The method for processing welding spot images of a welding robot based on a visual sensor according to claim 1, wherein: In step 2, the defect characteristics include the defect area ,perimeter , minimum enclosing rectangle aspect ratio , For the long side, is the short side; Cross-layer defects through feature matching matrix track, For the Layer and The similarity of layer defect features is determined and the Hungarian algorithm is used to find the optimal matching path.

6. The method for processing welding spot images of a welding robot based on a visual sensor according to claim 1, wherein: In step 2, the penetration depth of the basic feature Measured by laser displacement sensor, , is the laser spot offset, is the laser incident angle.

7. The method for processing welding spot images of a welding robot based on a visual sensor according to claim 1, wherein: In step 3, stereo matching uses the feature-based SGBM algorithm, and the matching cost function is C(x, y) = w1·C color +w2·C grad +w3·C laye , w1+w2+w3=1; where C color is the color cost, C grad is the gradient cost, is the hierarchical constraint cost, w1, w2, and w3 are all weight coefficients.

8. The method for processing welding spot images of a welding robot based on a visual sensor according to claim 1, wherein: In step 4, the defect location factor The stress level at the defect location is reflected, the fusion line is determined by the weld formation model, and the stress concentration area is marked according to the finite element simulation.

Images