Vision-based welding quality intelligent evaluation method and system

Abstract

The invention discloses an intelligent welding quality evaluation method and system based on vision, and relates to the technical field of welding quality evaluation.The method comprises the steps that the welding process is monitored in real time through multi-mode imaging equipment, weld joint surface texture, molten pool heat distribution and three-dimensional shape information are synchronously obtained, and a three-dimensional voxel model is constructed; establishing a two-stage multi-channel architecture for executing dynamic resource scheduling, and extracting geometric features and deep semantic features in the three-dimensional voxel model; and on the basis of the defect confidence coefficient vector, in combination with the real-time process parameters and the molten pool state, a dynamic evaluation model of the welding quality index is constructed, and time-phased quality scores are calculated and output through stage weighting. According to the method, the welding seam surface texture, molten pool heat distribution and three-dimensional shape information are synchronously collected through multi-mode imaging equipment, a texture-temperature-space coordinate bound three-dimensional voxel model is constructed, and multi-source heterogeneous data fusion analysis of the welding quality is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of welding quality evaluation, in particular to a visual-based welding quality intelligent evaluation method and system. BACKGROUND

[0002] As one of the core processes of modern manufacturing, welding technology has been widely used in key fields such as automobile body manufacturing, aerospace precision part connection, new energy battery module packaging, and pressure vessel welding. With the increasing demand for industrial automation, the quality stability of the welding process is directly related to the structural strength and safety of the product. Intelligent welding and assembly technology has been applied on a large scale in many industrial fields, and has shown significant value in the manufacturing of large structural parts. In the field of water conservancy engineering, such as the pressure steel pipe installation scene of Huangjinping Hydropower Station, intelligent welding and assembly machines achieve high-precision welding through digital control systems, solving the problems of low efficiency and unstable quality of traditional manual welding.

[0003] However, traditional welding quality assurance mainly relies on post-weld destructive testing (such as tension testing) or non-destructive sampling inspection (such as X-ray detection and ultrasonic testing). These methods have obvious limitations: first, the detection is lagging, and defects are difficult to correct in real time after they are formed, leading to high scrap rates; second, the cost is high and the efficiency is low, for example, full inspection of LNG tank welds requires a large amount of manpower and equipment resources.

[0004] Existing technologies have gradually shifted to online monitoring of the welding process, mainly relying on multi-source sensing information analysis. For example, by collecting welding current / voltage signal statistics to predict droplet transfer stability; using time-frequency features of arc sound signals combined with SVM models to identify penetration states; or monitoring molten pool temperature field gradients based on infrared thermal imaging to capture local temperature drops caused by porosity. However, in the process of realizing real-time monitoring of welding, there are limitations of single-mode monitoring, such as electromagnetic interference of electrical signals, low signal-to-noise ratio of arc sound in noisy workshops, and difficulty in quantifying microscopic defects such as cracks with infrared thermal imaging; and poor adaptability to complex working conditions, such as visual misjudgment of three-dimensional curvature changes or overlapping areas of automobile body welds, which can misdiagnose the design form as a defect; insufficient dynamic process response, existing algorithms mostly rely on static expert database comparison, lack real-time modeling capability for transient molten pool deformation, material thermal deformation, etc. during the welding starting / ending stage, and changes in lighting conditions will significantly reduce the reliability of gray level and gradient features. SUMMARY

[0005] Therefore, it is necessary to provide a visual-based welding quality intelligent evaluation method and system to solve the above technical problems.

[0006] In a first aspect, the present application provides a visual-based welding quality intelligent evaluation method, comprising: Real-time monitoring of the welding process is performed using a multi-modal imaging device, and surface texture, molten pool heat distribution, and three-dimensional topography information of the weld are synchronously acquired to construct a three-dimensional voxel model. A two-stage multi-channel architecture for performing dynamic resource scheduling is built, geometric features and deep semantic features in the three-dimensional voxel model are extracted, a cross-modal attention mechanism is combined, and a defect type classification result and a defect confidence vector of the weld are output. Based on the defect confidence vector, a dynamic evaluation model of the welding quality index is constructed by combining real-time process parameters and the molten pool state, and a time-period quality score is output by stage weighting calculation. The two-stage multi-channel architecture for performing dynamic resource scheduling includes: A multi-scale feature pyramid is used to output feature maps, a two-stage decoupling framework of coarse detection and fine classification is constructed, and a preliminary screening result of the coarse detection stage and a corresponding parallel processing channel of the fine classification stage are set. Based on the feature maps output by the multi-scale feature pyramid, a variable convolution is used to adapt to the weld deformation, and a preliminary screening result containing spatial position and risk level is output. The input, processing, and output constraints of the parallel processing channel of the fine classification stage are established, a dynamic resource allocation strategy is set based on task priority, and the computing power of different parallel processing channels is isolated and shared through thread parallel control.

[0007] Further, real-time monitoring of the welding process is performed using a multi-modal imaging device, and surface texture, internal molten pool structure, and three-dimensional topography information of the weld are synchronously acquired to construct a three-dimensional voxel model, which includes: A three-spectrum imaging device combining visible light, short-wave infrared, and coded structured light is deployed to synchronously capture surface texture, molten pool heat distribution, and three-dimensional topography information of the weld. Frequency domain noise separation is performed on the surface texture image of the weld, spatial filtering is performed on the three-dimensional topography image, and the thermal radiation features of the molten pool heat distribution image are enhanced. Based on preset calibration parameters, the surface texture, molten pool heat distribution, and three-dimensional topography point cloud of the weld are mapped to a unified coordinate system, sub-pixel level spatial alignment is achieved through feature point matching and affine transformation, and a registration dataset bound by texture, temperature, and coordinates is generated. The point cloud is converted into a discrete three-dimensional grid, and the texture color value, temperature gradient value, and spatial coordinate value are fused in each voxel to construct a three-dimensional voxel model containing multi-channel attributes.

[0008] Further, a multi-scale feature pyramid is used to output feature maps, a two-stage decoupling framework of coarse detection and fine classification is constructed, and a preliminary screening result of the coarse detection stage and a corresponding parallel processing channel of the fine classification stage are set, which includes: A unified feature extraction backbone network is constructed, and feature maps of different paths are input into a multi-scale feature pyramid. Multi-scale feature fusion is achieved through top-down and horizontal connection to generate three groups of outputs. Based on the multi-scale feature pyramid feature map, three types of preliminary screening results of spatial position and risk label are generated through deformable convolution to realize deformation adaptive region proposal and risk grading. The preliminary screening results include regular welds, abnormal defects and process abnormalities. According to the spatial position and risk label output by the coarse detection, the local feature map is cropped from the multi-scale feature pyramid, and the corresponding fine classification parallel processing channel is activated and the related feature data is routed. A phased training mechanism is used to freeze the backbone network parameters, and a gradient blocking layer is added to the output layer of the backbone network to limit the interference of fine classification task back propagation on coarse detection feature extraction.

[0009] Further, the geometric features and deep semantic features in the three-dimensional voxel model are extracted, combined with the cross-modal attention mechanism, and the defect type classification results and defect confidence vectors of the weld are output, including: The coordinate system is reconstructed with the weld pool center as the origin, the spatial alignment and normalization of the three-dimensional voxel model are performed, and the histogram equalization of the texture and temperature channels is performed to generate a standardized voxel model. Geometric features and deep semantic features are extracted from the standardized voxel model in parallel, and visible light texture and infrared gradient features are fused in the abnormal defect area. Through channel interaction attention, the residual feature of small sample defect is enhanced. According to the preliminary screening results of the coarse detection stage, the corresponding parallel processing channel of the fine classification stage is allocated, the channel exclusive classification task is executed, and the defect type label and defect confidence vector are output. A feature data cache pool is established, and memory resources are dynamically released according to the priority of the classification results, and the load status of each channel is fed back to trigger dynamic resource allocation adjustment.

[0010] Further, based on the defect confidence vector, combined with real-time process parameters and pool state, a dynamic evaluation model of welding quality index is constructed, and a time-period quality score is output through stage weighting calculation, including: Based on the welding process characteristics, different dynamic time periods are divided, and the defect confidence vector, real-time process parameters and pool state are time-stamped synchronized and spatial coordinate mapped to achieve accurate alignment. Based on the characteristics of the welding stage, the contribution weight of different input data is dynamically adjusted, and the feature vector of the three types of input data is dynamically weighted to generate an adaptive and optimized comprehensive feature vector. A real-time dynamic evaluation model is constructed to map the comprehensive feature vector to the welding quality index in the standard interval, and the time period score is calculated to mark the key defect period position.

[0011] Further, based on the welding stage characteristics, the contribution weights of different input data are dynamically adjusted, and the adaptive optimized comprehensive feature vector is generated by dynamically weighting the feature vectors of the three types of input data, including: Based on the molten pool state time sequence mode and the process parameter change rate, the current welding stage is identified, and the stage label and stage confidence are outputted; For the three types of input data, set the quality sensitive factor, call the preset weight according to the stage label, evaluate the contribution of each type of input data to the welding quality index through iteration, and generate the contribution weight; Based on the output contribution weight after iteration, the feature vectors of the three types of input data are gate weighted and fused to generate an anti-interference comprehensive feature vector; The prediction deviation of the welding quality index is fused to correct the quality sensitive factor in the iteration process, and the closed loop optimization of the contribution weight is realized.

[0012] Further, a real-time dynamic evaluation model is constructed to map the comprehensive feature vector to the welding quality index in the standard interval, and the time period score is calculated to mark the key defect period position, including: The input comprehensive feature vector is adaptively standardized, the welding physical rule constraint is embedded, and the time sequence feature and spatial feature are fused through the attention enhanced gate recurrent unit model to dynamically map the normalized feature vector to the welding quality index in the standard interval; Based on the time sequence of the welding quality index, the time period score is dynamically calculated according to the welding stage, and the defect high incidence window is marked through statistical process; According to the actual flaw detection result and the welding quality index prediction deviation, the normalized parameters of the attention enhanced gate recurrent unit model are automatically adjusted to generate a multi-dimensional quality evaluation board, real-time locate the defect window position, and trigger the graded warning signal.

[0013] Further, the calculation formula of the time period score is: ; In the formula, Q T Indicates the current time period score; WQI t Indicates t The welding quality score at the moment; R c Indicates the compliance rate score; S m Indicates the state index; D d Indicates the defect density factor; T Indicates the welding period division; α 、 β 、 γ 、 δThe preset weight coefficients are represented by a, b and c.

[0014] Further, a real-time dynamic evaluation model is constructed to map the comprehensive feature vector into a welding quality index in a standard interval, and after the time period score is counted and the key defect period position is marked, the following steps are further included: According to the distribution characteristics of the defect confidence vectors, a plurality of confidence labels are divided, combined with the welding quality index, the target sensitive area to be rechecked is identified, and rechecking resources are allocated; Based on the defect type and quality evaluation result, process optimization parameters are generated and input into the welding digital twin to pre-act the heat deformation trend after parameter optimization and drive the welding equipment to dynamically adjust parameters; According to the distribution characteristics of the defect confidence vectors, a plurality of confidence labels are divided, combined with the welding quality index, the target sensitive area to be rechecked is identified, and rechecking resources are allocated, including: The defect confidence vectors are divided into a plurality of confidence labels, the confidence labels, the welding quality index and the process fluctuation are fused, the rechecking sensitive area cluster is identified through the spatiotemporal constraint density clustering, and the center coordinates of the sensitive area cluster are located; According to the label density of the sensitive area cluster, the welding quality index mean and the area ratio, the risk score is calculated through weighting, and when the risk score is greater than a set threshold, local rechecking is triggered; According to the rechecking result, the hierarchical judgment threshold of the confidence label is corrected.

[0015] In a second aspect, the present application also provides a visual-based welding quality intelligent evaluation system, which comprises: A synchronous acquisition module is used for monitoring the welding process in real time by using a multi-modal imaging device, synchronously acquiring weld surface texture, molten pool heat distribution and three-dimensional topographic information, and constructing a three-dimensional voxel model; A multi-channel processing module is used for building a two-stage multi-channel architecture for executing dynamic resource scheduling, extracting geometric features and deep semantic features in the three-dimensional voxel model, combining a cross-modal attention mechanism, and outputting a defect type classification result and a defect confidence vector of the weld; A quality evaluation module is used for constructing a dynamic evaluation model of a welding quality index based on the defect confidence vector, combining real-time process parameters and molten pool states, and outputting a time period quality score through stage weighting calculation; The synchronous acquisition module, the multi-channel processing module and the quality evaluation module are sequentially connected.

[0016] The present application has the following advantages: 1. By synchronously collecting the weld surface texture, molten pool heat distribution and three-dimensional topography information through a multi-modal imaging device, a three-dimensional voxel model bound by texture-temperature-space coordinates is constructed to realize multi-source heterogeneous data fusion analysis of the welding quality; combined with real-time process parameters and molten pool state, a dynamic evaluation model of the welding quality index is constructed to output time-period quality scores through stage weighting calculation, significantly improving the comprehensiveness and real-time performance of the evaluation; and a defect confidence grading mechanism and a dynamic allocation strategy of review resources are introduced to form a whole-process monitoring and evaluation of welding detection, evaluation and optimization, driving dynamic adjustment of welding parameters and accurate prevention and control of defect risks, effectively avoiding the evaluation deviation caused by ignoring the dynamic nature of welding in traditional methods.

[0017] 2. According to the quality change characteristics of different stages in the welding process, a stage weight self-adaptive adjustment mechanism is introduced to identify the current welding stage through the molten pool state time sequence mode and the process parameter change rate, dynamically allocate the contribution weights of the defect confidence vector, process parameters and molten pool state three types of data, and generate an anti-interference comprehensive feature vector, solving the problem that the traditional static evaluation model cannot adapt to the non-uniformity of the welding process, especially strengthening the sensitivity to the unstable state of the arc striking segment and the molten pool shrinkage effect of the arc collecting segment, ensuring that the high-risk defect period is accurately marked, and significantly improving the detection robustness of key defects.

[0018] 3. By constructing a two-stage processing framework decoupled from coarse detection and fine classification, combining a cross-modal attention mechanism and dynamic resource allocation of parallel processing channels, intelligent scheduling of computing resources and optimization of detection efficiency are realized; in the coarse detection stage, a multi-scale feature pyramid and deformable convolution are used to quickly screen high-risk areas and route them to the fine classification channel; in the fine classification stage, the dedicated classification channel is activated according to the task priority, the small sample defect features are enhanced through channel interaction attention, and resources are dynamically released based on the cache pool load state, thereby meeting the real-time requirements of high-speed welding. BRIEF DESCRIPTION OF DRAWINGS

[0019] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 is a flowchart of a visual-based welding quality intelligent evaluation method according to an embodiment of the present application; Figure 2 is a system principle block diagram of a visual-based welding quality intelligent evaluation system according to an embodiment of the present application.

[0020] Reference numerals: 1, synchronous acquisition module; 2, multi-channel processing module; 3, quality evaluation module. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and not to limit the present application.

[0022] Please refer to Figure 1 , provides a visual-based welding quality intelligent evaluation method, comprising: S1, using a multi-modal imaging device to monitor the welding process in real time, synchronously acquiring weld surface texture, molten pool heat distribution and three-dimensional topography information, and constructing a three-dimensional voxel model.

[0023] In the description of the present application, using a multi-modal imaging device to monitor the welding process in real time, synchronously acquiring weld surface texture, internal molten pool structure and three-dimensional topography information, and constructing a three-dimensional voxel model comprises: S11, deploying a three-spectrum imaging device combined with visible light, short-wave infrared and coded structured light, synchronously capturing weld surface texture, molten pool heat distribution and three-dimensional topography information.

[0024] Specifically, the visible light imaging device is equipped with a 650nm narrowband filter to suppress arc light interference and focus on surface texture details; the short-wave infrared imaging device (wavelength range 1.4-3um) penetrates the oxide layer to capture internal heat radiation signals of the molten pool; the structured light imaging device (DLP projector) projects a red and black dual-color pseudo-random dot array + single-color sinusoidal fringe combined coding pattern, and compensates for smoke shielding through phase calculation.

[0025] The three-spectrum imaging device generates a synchronous pulse signal based on FPGA, which simultaneously triggers the exposure of the three-mode devices; the collection time of each mode data is recorded by hardware timestamp to ensure timing alignment.

[0026] S12, performing frequency domain noise separation on the weld surface texture image, spatial filtering on the three-dimensional topography image, and enhancing the heat radiation features of the molten pool heat distribution image.

[0027] Specifically, the visible light image (weld surface texture image) is subjected to frequency domain noise separation, the structured light point cloud (three-dimensional topography image) is subjected to spatial filtering, and the heat radiation features of the short-wave infrared image (molten pool heat distribution image) are enhanced, solving the problems of arc light overexposure, smoke scattering and heat signal attenuation, and outputting a set of enhanced multispectral images with high signal-to-noise ratio.

[0028] Visible light frequency domain processing needs to perform Fourier transform on the original visible light image to separate high-frequency arc light noise and low-frequency weld texture, and to retain low-frequency components to reconstruct a clear surface image.

[0029] The structured light point cloud optimization adopts a moving least squares method (MLS) to smooth the point cloud and filter out outlier noise caused by splashing particles; and the phase unwrapping algorithm is used to reconstruct the topographic data of the area blocked by smoke and dust.

[0030] The short-wave infrared feature enhancement requires non-uniformity correction (NUC) of the original thermal image to eliminate the response difference of the detector; and histogram equalization is used to improve the contrast between the molten pool area and the background.

[0031] S13, based on the preset calibration parameters, the weld surface texture, the molten pool heat distribution and the three-dimensional topographic point cloud are mapped to a unified coordinate system, the sub-pixel level spatial alignment is realized through feature point matching and affine transformation, and a registration data set bound with texture-temperature-coordinate is generated.

[0032] Specifically, based on the preset calibration parameters, the visible light texture, the short-wave infrared thermal image and the structured light three-dimensional point cloud are mapped to a unified coordinate system, the sub-pixel level spatial alignment is realized through feature point matching and affine transformation, and a registration data set bound with “texture-temperature-coordinate” is generated, laying a foundation for spatial consistency for three-dimensional voxel fusion.

[0033] The checkerboard calibration board is used to calculate the intrinsic matrix of each camera and the extrinsic matrix relative to the world coordinate system; the calibration parameters are optimized through bundle adjustment to reduce the re-projection error; the SIFT feature points are extracted from the visible light and infrared images; the RANSAC algorithm is used to remove the mismatched points; and the affine transformation matrix is solved to map the 2D image pixel coordinates to the 3D point cloud spatial coordinates.

[0034] S14, the point cloud is converted into a discrete three-dimensional grid, the texture color value, the temperature gradient value and the spatial coordinate value are fused in each voxel, and a three-dimensional voxel model containing multi-channel attributes is constructed.

[0035] Specifically, first, the point cloud is voxelized, the AABB bounding box is calculated according to the spatial distribution of the structured light point cloud, and the voxel resolution is set; the GPU accelerated octree index is used to discretize the point cloud into a voxel grid, and the spatial coordinate mean value of each voxel is stored.

[0036] The three-dimensional voxel model contains multi-channel attributes, and for multi-channel attribute fusion, the following contents are included: 1. Texture channel: the RGB value of the registered visible light image is interpolated and mapped to the corresponding voxel; 2. Temperature channel: the radiation intensity value of the short-wave infrared thermal image is converted into a temperature value and written into the voxel; 3. Topographic channel: the three-dimensional coordinates and normal vector of the voxel center point are recorded.

[0037] S2, a two-stage multi-channel architecture for performing dynamic resource scheduling is built, geometric features and deep semantic features in the three-dimensional voxel model are extracted, a cross-modal attention mechanism is combined, and a defect type classification result and a defect confidence vector of the weld are output.

[0038] In the description of the application, the two-stage multi-channel architecture for performing dynamic resource scheduling is built, the geometric features and deep semantic features in the three-dimensional voxel model are extracted, the cross-modal attention mechanism is combined, and the defect type classification result and the defect confidence vector of the weld are output, which includes: S21, a multi-scale feature pyramid is used to output feature maps, a two-stage decoupling framework of coarse detection and fine classification is constructed, and the preliminary screening result of the coarse detection stage and the corresponding parallel processing channel of the fine classification stage are set.

[0039] In the description of the application, the multi-scale feature pyramid is used to output feature maps, the two-stage decoupling framework of coarse detection and fine classification is constructed, and the preliminary screening result of the coarse detection stage and the corresponding parallel processing channel of the fine classification stage are set, which includes: S211, a unified feature extraction backbone network is constructed, feature maps of different paths are input into a multi-scale feature pyramid, multi-scale feature fusion is realized through top-down and horizontal connection, and three groups of outputs are generated.

[0040] Specifically, by designing a unified feature extraction backbone network, a multi-scale feature pyramid (FPN) is used to synchronously output high-resolution geometric features and low-resolution semantic features, to provide a shared feature base for coarse detection and fine classification, and to avoid double-stage repeated calculation.

[0041] The hierarchical convolution path includes a high-resolution path and a low-resolution path.

[0042] The high-resolution path stacks 3 layers of depth separable convolution using a lightweight MobileNetV3 module, outputs a 1 / 8 resolution feature map (256x256 pixels), and focuses on geometric details such as weld contour edges and misalignment amounts.

[0043] The low-resolution path deploys 4 layers of ResNet-18 residual modules, expands the receptive field by combining dilated convolution (dilation=2), outputs a 1 / 32 resolution feature map (64x64 pixels), and captures semantic information such as molten pool temperature gradient field and texture mutation.

[0044] The feature maps of the high / low resolution path are input into the FPN, multi-scale feature fusion is realized through top-down and horizontal connection, and three groups of outputs (1 / 8, 1 / 16, 1 / 32 resolution) are generated, covering different scale detection requirements.

[0045] S212, based on the multi-scale feature pyramid feature map, generate the spatial position and risk label of the three types of preliminary screening results through deformable convolution, realize deformation adaptive region proposal and risk grading. Among them, the preliminary screening results include regular weld, abnormal defect and process anomaly.

[0046] Specifically, in the YOLOv8 detection head, 3 groups of deformable convolution (DCNv2) are embedded to dynamically learn the weld deformation offset and improve the robustness to irregular mutations of the molten pool boundary.

[0047] For the preliminary screening results including regular weld, abnormal defect and process anomaly, the determination logic and output form are as follows: 1. Regular weld, the determination logic is that the geometric parameters meet the standard (width error < threshold) and the contour smoothness is greater than or equal to the threshold, and the output form is a high confidence label (Confidence > 0.95); 2. Abnormal defect, the determination logic is that the local texture mutation area and the temperature gradient anomaly area are spatially overlapped, and the output form is a defect candidate box (IoU > 0.7); 3. Process anomaly, the determination logic is that the current fluctuation frequency is greater than the threshold and the correlation coefficient with the temperature gradient is less than the threshold, and the output form is an abnormal mark + associated sensor ID.

[0048] S213, according to the spatial position and risk label output by the coarse detection, crop the local feature map from the multi-scale feature pyramid, activate the corresponding fine classification parallel processing channel and route the related feature data.

[0049] Specifically, fine classification needs to establish and bind corresponding image processing channels, and the channel binding strategy is as follows: 1. Regular weld channel: activate FPN1 / 8 high resolution feature map routing, only receive geometric feature vectors of high resolution feature map, and force lightweight verification (MobileNetV3 classification head + B-spline contour fitting); 2. Abnormal defect channel: requires dual-mode input (visible light texture + infrared thermal image), forces fusion of visible light texture (1 / 8 layer) and infrared thermal image (1 / 32 layer) features, and enables cross-modal attention mechanism; 3. Process anomaly channel: bind the current sensor time series data stream, input the molten pool temperature sudden change point coordinate sequence, and force to use time series convolution network (TCN) to analyze the root cause.

[0050] S214, adopt a phased training mechanism to freeze the backbone network parameters, add a gradient blocking layer at the output layer of the backbone network, and limit the reverse propagation interference of fine classification task to coarse detection feature extraction.

[0051] S22, based on the feature map output by the multi-scale feature pyramid, the deformable convolution is used to adapt to the weld deformation, and the initial screening result containing the spatial position and risk level is output.

[0052] Specifically, based on the feature map output by the multi-scale feature pyramid (including 1 / 8, 1 / 16, 1 / 32 resolution feature map), a deformable convolution module (DCNv2) is deployed to handle the geometric deformation of the weld in the coarse detection stage. The specific implementation process is as follows: 1. Deformation offset learning: for the multi-scale feature map output by the FPN, the sampling point offset (Δ x y ) is dynamically predicted by the deformable convolution kernel, so that the convolution kernel can adapt to the irregular deformation of the molten pool boundary (such as weld distortion caused by thermal deformation). The offset is generated by an additional convolution layer, and its input is the geometric context information of the current feature map.

[0053] 2. Region proposal generation: use deformable convolution to slide and scan on three scale feature maps, output the spatial position of three types of initial screening results, as follows: 2.1, regular weld, when the weld width error < preset threshold (such as ± 0.1mm) and the contour smoothness ≥ 0.9, output high confidence bounding box (Confidence > 0.95); 2.2, abnormal defect, detect the spatial coincidence area of local texture mutation area (Laplacian response > threshold) and temperature gradient abnormal area , output defect candidate box (IoU > 0.7); 2.3, process anomaly, analyze the area where the current fluctuation frequency > threshold (such as 10Hz) and the correlation coefficient < 0.3 with the temperature gradient, output abnormal marker point and associated sensor ID.

[0054] 3. Risk classification mechanism: calculate the risk score according to the spatial position feature, as follows: 3.1, position confidence, based on the matching degree (Hausdorff distance ≤ 0.05mm) of the deformable convolution sampling point and the real weld edge; 3.2, risk level, fuse the geometric deformation amplitude (curvature change rate) and the thermodynamic parameter (temperature gradient entropy), output low (L1), medium (L2) and high (L3) risk labels.

[0055] S23, establish the input, processing and output constraints of the parallel processing channel in the fine classification stage, set a dynamic resource allocation strategy based on task priority, and realize the calculation power isolation and sharing of different parallel processing channels through thread parallel control.

[0056] ​Specifically, in the fine classification stage, the pre-designed proportion of computing resources is 70% for abnormal defects, 20% for process abnormalities, and 10% for regular welds. The GPU core occupancy rate is controlled through CUDA stream. Real-time monitoring of the task queue state of each channel, if a channel is idle (such as the regular weld queue is empty), then dynamically schedule its computing power to other high-load channels (such as abnormal defect channel), and avoid resource conflicts through thread locking.

[0057] The parallel control mechanism creates independent CUDA streams for each type of channel, schedules tasks through a priority queue (abnormal defect task > process abnormality > regular weld), and designs an idle resource redistributor that automatically transfers the computing power quota of a channel to a high-priority channel when the task queue of the channel is empty.

[0058] S24, the center of the weld pool is taken as the origin to reconstruct the coordinate system, the three-dimensional voxel model is spatially aligned and normalized, and the histogram equalization is performed on the texture and temperature channels to generate a standardized voxel model.

[0059] Specifically, the center of the weld pool is taken as the origin to reconstruct the coordinate system, and the voxel model is translated to the center of the standard coordinate system through affine transformation; the principal component analysis (PCA) is used to calculate the principal axis direction of the weld, and the model is rotated to align the length direction of the weld with the Z axis. The original voxel is resampled to 0.1mm³ resolution using trilinear interpolation method to ensure the comparability of the geometric dimensions of different welds, and the histogram equalization is performed on the texture (RGB) and temperature (infrared) channels respectively to enhance the consistency of the features.

[0060] S25, geometric features and deep semantic features are extracted from the standardized voxel model in parallel, and visible light texture and infrared gradient features are fused in the abnormal defect area to realize residual feature enhancement of small sample defects through channel interaction attention.

[0061] Specifically, geometric features (contour curvature, misalignment amount) and deep semantic features (pool temperature gradient field, texture mutation pattern) are extracted from the standardized voxel model in parallel to avoid feature interference through a double-path isolation.

[0062] For the geometric feature path, first, the contour curvature is calculated, the weld edge point set is extracted in the X-Y cross section, the second derivative is calculated through B-spline curve fitting, and the maximum curvature value and position are output; then the misalignment amount is quantified, the Hausdorff distance of the height difference between the two sides of the weld is calculated, and the absolute value of the misalignment amount is output.

[0063] For the semantic feature path, first, the temperature gradient field is generated, a 3x3x3 Sobel operator is used to scan the pool area voxel, and the temperature change rate vector Then, the 3D Laplacian operator is applied to the visible light channel voxels to enhance the edge response of defects such as pores and cracks.

[0064] In addition, based on the preset channel protocol, the visible light texture and the infrared thermal gradient features of the abnormal defect area are fused, and the cross-material generalization ability of the small sample defect is enhanced through channel interaction attention (CI-Attention). In the implementation process, the infrared thermal image needs to be mapped to the visible light texture coordinate system through bilinear interpolation to realize pixel-level spatial alignment; the visible light branch generates a Query vector through 1x1 convolution, and the infrared branch generates a Key-Value vector, and a cross-modal attention weight matrix is calculated; the weighted features are added to the original visible light features to retain the original texture details while strengthening the thermal gradient associated area (such as the temperature drop area around the pores).

[0065] S26, according to the preliminary screening result of the coarse detection stage, assign to the corresponding parallel processing channel of the fine classification stage, execute the channel exclusive classification task, output the defect type label and the defect confidence vector.

[0066] Specifically, according to the preliminary screening result (including the spatial position and risk label of the three types of areas of regular weld, abnormal defect and process anomaly) output by the coarse detection stage, the following fine classification task allocation and processing process is executed: 1, channel activation and feature routing: activate the corresponding fine classification parallel processing channel according to the risk label.

[0067] 1.1, regular weld area: activate FPN1 / 8 high-resolution feature map routing and only receive geometric feature vectors (such as contour curvature and edge error) to verify the geometric parameter compliance through a lightweight MobileNetV3 classification head to perform B-spline contour fitting verification.

[0068] 1.2, abnormal defect area: forcibly fuse visible light texture (FPN1 / 8 layer) and infrared thermal image (FPN1 / 32 layer) features, and enable cross-modal attention mechanism (CI-Attention). In the specific implementation, the infrared thermal image is aligned to the visible light coordinate system through bilinear interpolation, the visible light branch generates a Query vector, the infrared branch generates a Key-Value vector, and an attention weight matrix is calculated to weight the fused features, and the texture-thermal gradient correlation of the defect area (such as the temperature drop area around the pores) is strengthened.

[0069] 1.3, process anomaly area: bind the current sensor time series data stream, input the melt pool temperature sudden change point coordinate sequence, and use a time convolution network (TCN) to analyze the correlation between current fluctuation and temperature gradient.

[0070] 2、Dynamic resource scheduling: allocate computing resources according to preset priority (abnormal defect > process anomaly > regular weld): abnormal defect channel occupies 70% GPU computing power (FP16 precision), process anomaly channel occupies 20% (INT8 quantization), and regular weld channel occupies 10%. Through independent CUDA stream, the computing power is isolated, and the task queue state is monitored in real time. If a channel is idle (such as the regular weld queue is empty), the computing power is dynamically scheduled to a high-load channel (such as the abnormal defect channel) through a thread lock mechanism, and the idle resource allocator is used to transfer the computing power quota.

[0071] 3、Feature standardization and classification output: standardize the input features of each channel: reconstruct the coordinate system with the molten pool center as the origin, align the weld main axis direction through PCA, and histogram equalize the texture (RGB) and temperature (infrared) channels respectively to ensure feature consistency.

[0072] Among them, for channel-specific task execution: the abnormal defect channel extracts the continuity edge features of the micro-defects through the residual feature enhancement module, and outputs the defect type label (porosity / crack / under-melt) combined with the cross-modal attention. The process anomaly channel uses the time convolution network (TCN) to analyze the molten pool temperature sudden change point sequence, and outputs the abnormal root cause label (such as current overload or uneven cooling).

[0073] 4、Result generation and resource release: each channel outputs the defect type label and multi-dimensional confidence vector, and the vector elements are respectively represented, which are embodied in the following aspects: 4.1、Class probability: the likelihood of the defect belonging to the target class (0~1); 4.2、Positioning reliability: spatial position error score based on Hausdorff distance; 4.3、Morphological integrity: defect contour closure and thermal gradient continuity score.

[0074] S27、Establish a feature data cache pool, dynamically release memory resources according to the priority of the classification results, and feed back the channel load state to trigger dynamic resource allocation adjustment.

[0075] Specifically, a feature data cache pool is established, memory resources are dynamically released according to the priority of the classification results, and the channel load state is fed back to the resource scheduling engine to realize computing power rebalancing.

[0076] For example, the hierarchical caching strategy can refer to the following content: 1、High confidence results (confidence > 95%) only save feature vectors and classification labels, and original voxel data are immediately released; 2、Low confidence results (confidence ≤ 80%) retain original voxel data and intermediate features for S24 channel iterative analysis.

[0077] Real-time statistics of each channel processing delay, if the average delay of abnormal defect channel > 20ms, trigger S214 resource scheduling engine to increase GPU core quota.

[0078] S3, based on the defect confidence vector, combined with real-time process parameters and molten pool state, a dynamic evaluation model of welding quality index is constructed, the quality score of each period is calculated by stage weighting, and a confidence grading mechanism is introduced to trigger the allocation of re-inspection resources.

[0079] In the description of the application, based on the defect confidence vector, combined with real-time process parameters and molten pool state, a dynamic evaluation model of welding quality index is constructed, the quality score of each period is calculated by stage weighting, and a confidence grading mechanism is introduced to trigger the allocation of re-inspection resources, including: S31, based on the welding process characteristics, different dynamic periods are divided, the defect confidence vector, real-time process parameters and molten pool state are time-stamped and spatially mapped, and accurate alignment is realized.

[0080] Specifically, according to the welding process characteristics (arc striking→stable→arc extinguishing), the process can be divided into three dynamic periods, and the time axis of the defect confidence vector, process parameters and molten pool state is accurately aligned, laying a time sequence foundation for periodical evaluation.

[0081] For example, the period division rule can refer to the following content: 1, arc striking segment (0-10% welding time length): current climbing period, monitoring the initial state of molten pool formation; 2, stable segment (10%-85% welding time length): process parameter fluctuation tolerance period, focusing on defect-prone areas; 3, arc extinguishing segment (85%-100% welding time length): current decay period, detecting shrinkage defects (such as arc pit cracks).

[0082] S32, based on the characteristics of welding stage, dynamically adjust the contribution weight of different input data, generate adaptive optimized comprehensive feature vector by dynamically weighting the feature vectors of three types of input data.

[0083] In the description of the application, based on the characteristics of welding stage, dynamically adjust the contribution weight of different input data, generate adaptive optimized comprehensive feature vector by dynamically weighting the feature vectors of three types of input data, including: S321, based on the time sequence mode of molten pool state and the change rate of process parameters, identify the current welding stage, output stage label and stage confidence, realize accurate staging of welding process, and provide stage basis for subsequent weight allocation.

[0084] Specifically, based on the time sequence mode of molten pool state (such as temperature change rate, molten pool oscillation frequency) and the change rate of process parameters (current slope, voltage fluctuation rate), the dynamic characteristics of the welding process are analyzed in real time.

[0085] Model the molten pool temperature time series curve through the LSTM network, combine the first derivative change characteristics of the process parameters (such as the current change rate > 10A / ms to determine the arc stage), output the stage label (arc / stable / arc) and the stage confidence (0-1 probability value).

[0086] S322, for three types of input data, set the quality sensitive factor, call the preset weight according to the stage label, and generate the contribution weight by iteratively evaluating the contribution of each type of input data to the welding quality index.

[0087] Specifically, for the three types of input data (molten pool state, process parameters, and defect confidence), the quality sensitive factors are set respectively: molten pool temperature gradient entropy (thermal stability), current fluctuation variance (process stability), and defect density (quality risk).

[0088] According to the stage label output by S321, the preset weight base value is called, and the partial derivative of each type of data to the welding quality index ( WQI ) is iteratively evaluated by the Gauss-Newton method to dynamically correct the contribution weight. For example, in the arc stage, the weight of the molten pool state data is increased from the initial 0.25 to 0.42 after 3 iterations, because it has a higher contribution to cold crack prediction.

[0089] S323, based on the contribution weight output after iteration, the feature vectors of the three types of input data are gated and weighted to generate an anti-interference comprehensive feature vector.

[0090] Specifically, based on the contribution weight generated by iteration, the gating mechanism (Gating Mechanism) is used to weight and fuse the three types of feature vectors: the molten pool state vector is multiplied by the Sigmoid gating coefficient (activated when the weight > 0.4), the process parameter vector is bound with the Tanh gating function, and the defect confidence vector is filtered by the ReLU gating mechanism. Low confidence features.

[0091] The weighted vector is normalized by channel and then spliced into a comprehensive feature vector, which has anti-interference performance: when the arc is suddenly changed, the gating mechanism reduces the molten pool state weight from 0.35 to 0.18, suppressing abnormal heat distribution noise.

[0092] S324, fuse the prediction deviation of the welding quality index, and reverse correct the quality sensitive factor in the iteration process to realize the closed-loop optimization of the contribution weight.

[0093] Specifically, the predicted value is compared with the actual detection result (such as the defect area deviation of X-ray detection) in real time. WQI When the prediction deviation > 15%, reverse correction is triggered: the partial derivative of the quality sensitive factor to the deviation is calculated by the gradient descent method, and the learning rate η=0.01 update the sensitive factor in S322.

[0094] For example, when the arc collection stage gas hole detection rate rises, the system corrects the sensitive coefficient of the defect density factor from 0.32 to 0.48, and finally forms a “prediction-measurement-correction” closed loop.

[0095] S33, a real-time dynamic evaluation model is constructed, the comprehensive feature vector is mapped to the welding quality index in the standard interval, and the time period score is counted, and the key defect period position is marked.

[0096] In the description of the present application, the real-time dynamic evaluation model is constructed, the comprehensive feature vector is mapped to the welding quality index in the standard interval, and the time period score is counted, and the key defect period position is marked. S331, adaptive standardization processing is performed on the input comprehensive feature vector, embedding welding physical rule constraints, and through attention enhanced gate recurrent unit model fusion time sequence feature and space feature, the normalized feature vector is dynamically mapped to the welding quality index in the standard interval.

[0097] S332, based on the time sequence of the welding quality index, the time period score is dynamically calculated according to the welding stage, and the defect high incidence window is marked through statistical process.

[0098] In the description of the present application, the calculation formula of the time period score is: .

[0099] In the formula, Q T The current time period score is represented by S(t); WQI t The welding quality score at the moment is represented by S(t); t R c The compliance rate score is represented by C(t); S m The state index is represented by S(t); D d The defect density factor is represented by D(t); T The welding time period division is represented by T(t); α 、 β 、 γ 、 δ All represent the preset weight coefficient.

[0100] It should be noted that, Q T The comprehensive evaluation result of the welding process in a specific period (such as the arc starting segment and the stable segment) is obtained by weighted calculation. Since all input items in the formula are normalized, Q T ​It is a dimensionless value, usually falling in the range of [0,1]. The higher the value, the better the welding quality during that period. It serves as the core output indicator for dynamic evaluation of welding quality.

[0101] WQI t ( t (Welding quality score at any time) reflects the welding process t The instantaneous quality state at any given moment is generated from real-time sensor data (current, voltage, temperature) through feature fusion and normalization. For example, the current volatility index is transformed into... WQI t Components of.

[0102] R c (Compliance score) represents the degree of conformity between actual welding parameters and process standards, and its formula is: It is defined as the deviation ratio, and the data source depends on the threshold preset by the process specification. I target , I threshold ) and the current value collected by the sensor in real time ( I actual ).

[0103] S m (State index) describes the dynamic stability of the welding process, such as the aspect ratio of the weld pool and temperature gradient. It needs to be standardized to eliminate differences in physical dimensions, and its sources include images of the weld pool morphology captured by high-speed industrial cameras and temperature gradient matrices from infrared thermal imagers.

[0104] D d (Defect density factor) reflects the probability density of defects occurring in the current period. It is obtained through historical statistics or real-time detection, and is derived from the historical defect distribution statistics of X-ray flaw detection reports or the defect count of real-time online detection systems.

[0105] S333. Based on the deviation between the actual flaw detection results and the predicted welding quality index, the normalization parameters of the attention-enhanced gating loop unit model are automatically adjusted to generate a multi-dimensional quality assessment dashboard, locate the defect window position in real time, and trigger graded early warning signals.

[0106] S34. Divide the defect confidence vector into multiple confidence labels, combine them with the welding quality index, identify the target sensitive area to be re-inspected, and allocate re-inspection resources.

[0107] In the description of this invention, multiple confidence labels are divided based on the distribution characteristics of the defect confidence vector. Combined with the welding quality index, the target sensitive area to be re-inspected is identified, and re-inspection resources are allocated, including: S341. Divide the defect confidence vector into multi-level confidence labels, integrate the confidence labels, welding quality index and process fluctuations, identify the sensitive area clusters for re-inspection through spatiotemporal constraint density clustering, and locate the center coordinates of the sensitive area clusters.

[0108] Specifically, the defect confidence vector (which includes three dimensions: category probability, positioning reliability, and morphological integrity) is divided into three levels of confidence labels according to percentiles: high confidence label (category probability ≥ 0.9 and positioning error ≤ 0.05 mm), medium confidence label (category probability ∈ [0.7, 0.9] or positioning error ∈ (0.05 mm, 0.1 mm]), and low confidence label (category probability < 0.7 or positioning error > 0.1 mm).

[0109] By integrating welding quality index and real-time process fluctuation parameters, an improved DBSCAN clustering algorithm is used to scan the weld area, setting a spatial constraint radius ≤5mm (with the center of the molten pool as the origin) and a density threshold. Isolated noise points are eliminated through spatiotemporal constraints, generating clusters of sensitive areas for re-inspection, and calculating the weighted centroid coordinates of each cluster to achieve precise location of high-probability defect areas.

[0110] S342. Based on the label density, average welding quality index, and area ratio of the sensitive area cluster, a risk score is calculated by weighting. When the risk score is greater than the set threshold, a local re-inspection is triggered.

[0111] Specifically, a comprehensive risk score is calculated for each sensitive region cluster. R This includes the following aspects: 1. Label density factor ( D label ): The ratio of the number of low / medium confidence labels within a cluster to the cluster area, weighted. α =0.5; 2. Quality Index Factor ( WQI norm ): within cluster WQI Normalized mean (if) WQI If the value is less than 0.6, the reciprocal is used to amplify the risk. β =0.3; 3. Area proportion factor ( A ratio ): The proportion of cluster area to the total weld area, weight γ =0.2.

[0112] The comprehensive risk score is calculated by weighted summation. R ,when R When the value exceeds 0.65 (preset threshold), a partial re-inspection process is automatically triggered. Re-inspection resources are dynamically allocated according to priority: high-risk clusters ( RThe top 30% of the values in descending order preferentially occupy 70% of the detection device computing power, and the remaining resources are sequentially allocated to low-risk clusters to avoid detection bottlenecks through real-time load balancing.

[0113] S343, according to the re-inspection result, the hierarchical decision threshold of the confidence label is corrected.

[0114] Specifically, if the high-confidence label region re-inspection finds missed detection, the category probability threshold is lowered, and the positioning reliability weight is increased; if the low-confidence label region re-inspection is correct, the shape integrity score threshold is raised, and the false positive rate is reduced.

[0115] S35, based on the defect type and quality evaluation results, process optimization parameters are generated and input into the welding digital twin to simulate the heat distortion trend after parameter optimization and drive the welding equipment to dynamically adjust parameters.

[0116] Specifically, based on the defect type classification results and quality evaluation indicators, key defect cause features are extracted: if gas hole is detected, the process parameters related to abnormal molten pool temperature gradient and insufficient protective gas flow are associated; if the crack confidence is high, the process parameter combination of the amount of misalignment and the cooling rate exceeding the limit is combined. Reinforcement learning strategy (such as Q-learning algorithm) is used to search for the optimal solution in the process parameter space with defect rate reduction and heat distortion inhibition as reward targets to generate optimization parameter combinations including current, voltage, and welding speed.

[0117] The optimization parameters are input into the welding digital twin, which simulates the welding temperature field, stress field and deformation field through finite element modeling (such as Perfect-Welding software). Based on the thermal elastic-plastic principle, real-time calculation of time-varying parameters such as elastic modulus, yield strength, and thermal expansion coefficient under temperature change is performed to simulate the heat distortion trend after parameter optimization.

[0118] For example, the reduction amplitude of the welding part warping deformation and the change of the thermal stress concentration area are predicted. y 1. Through virtual welding environment dynamic demonstration of heat distortion process and comparison with real-time DIC (digital image correlation method) monitoring data of actual welding, if the predicted heat distortion exceeds the safety threshold (such as residual stress > 70% of material yield strength), the reverse correction mechanism is triggered: the partial derivative of parameter deviation to heat distortion is calculated through gradient descent method, and the optimization parameters are updated with learning rate η =0.01.

[0119] The modified optimization parameters are transmitted to the welding equipment control system through an industrial bus. Based on the adaptive control strategy, the equipment output is adjusted in real time: for example, when the molten pool temperature gradient is predicted to be abnormal, the current is dynamically reduced and the protective gas flow is increased; when the cooling rate is detected to be out of limits, the heating device is triggered to reduce the cooling gradient. The system synchronously monitors the deviation of the actual welding parameters from the target value, and if the deviation persists >5%, the digital twin pre-play process is restarted.

[0120] See Figure 2 A visual-based intelligent welding quality evaluation system is also provided, comprising: A synchronous acquisition module 1 is used to monitor the welding process in real time using a multi-modal imaging device, synchronously acquire the weld surface texture, molten pool heat distribution and three-dimensional topography information, and construct a three-dimensional voxel model.

[0121] A multi-channel processing module 2 is used to build a two-stage multi-channel architecture for dynamic resource scheduling, extract geometric features and deep semantic features from the three-dimensional voxel model, combine a cross-modal attention mechanism, and output the defect type classification result and defect confidence vector of the weld.

[0122] A quality evaluation module 3 is used to construct a dynamic evaluation model of the welding quality index based on the defect confidence vector, combine real-time process parameters and molten pool state, output a time-period quality score through stage weighting calculation, and introduce a confidence grading mechanism to trigger re-inspection resource allocation.

[0123] The synchronous acquisition module 1, multi-channel processing module 2 and quality evaluation module 3 are sequentially connected.

[0124] In summary, by means of the above technical solutions of the present application, the weld surface texture, molten pool heat distribution and three-dimensional topography information are synchronously acquired by a multi-modal imaging device, a three-dimensional voxel model is constructed by binding texture-temperature-space coordinates, multi-source heterogeneous data fusion analysis of welding quality is realized, a dynamic evaluation model of the welding quality index is constructed based on real-time process parameters and molten pool state, a time-period quality score is output through stage weighting calculation, the comprehensiveness and real-time performance of the evaluation are significantly improved, a defect confidence grading mechanism and re-inspection resource dynamic allocation strategy are introduced, a full-process monitoring and evaluation of welding detection, evaluation and optimization is formed, welding parameter dynamic adjustment and defect risk precise prevention and control are driven, and evaluation deviation caused by neglecting welding dynamics in traditional methods is effectively avoided.

[0125] It should be understood that although the steps in the flowcharts of the drawings are shown in sequential order, such that each step depends on completion of the previous step before execution of the next step, the steps are not necessarily performed in the order indicated by the arrows. Unless specifically stated otherwise, the steps can be performed in other orders. Moreover, at least some of the steps in the flowcharts of the drawings can include multiple sub-steps or stages, which are not necessarily performed at the same time, but can be performed at different times, and which are not necessarily performed sequentially, but can be performed in rotation or alternation with other steps or sub-steps or stages of other steps.

Claims

1. A vision-based intelligent assessment method for welding quality, characterized in that, include: S1. Utilize multimodal imaging equipment to monitor the welding process in real time, simultaneously acquire information on weld surface texture, molten pool heat distribution, and three-dimensional morphology, and construct a three-dimensional voxel model; S2. Build a two-stage, multi-channel architecture for dynamic resource scheduling, extract geometric features and deep semantic features from the three-dimensional voxel model, and combine cross-modal attention mechanism to output the defect type classification results and defect confidence vector of the weld. S3. Based on the defect confidence vector, combined with real-time process parameters and molten pool status, a dynamic evaluation model for welding quality index is constructed, and the quality score for each time period is output through stage weighted calculation. The two-stage, multi-channel architecture for dynamic resource scheduling includes: S21. Use a multi-scale feature pyramid to output feature maps, construct a two-stage decoupled framework for coarse detection and fine classification, and set the initial screening results of the coarse detection stage and the corresponding parallel processing channels for the fine classification stage. S22. Based on the feature map output by the multi-scale feature pyramid, the weld deformation is adapted through variable convolution, and the initial screening results containing spatial location and risk level are output. S23. Establish input, processing and output constraints for parallel processing channels in the fine classification stage, set dynamic resource allocation strategies based on task priorities, and realize the isolation and sharing of computing power for different parallel processing channels through thread parallel control.

2. The vision-based intelligent welding quality assessment method according to claim 1, characterized in that, The process of using multimodal imaging equipment to monitor the welding process in real time, simultaneously acquiring weld surface texture, internal molten pool structure, and three-dimensional morphology information, and constructing a three-dimensional voxel model includes: S11. Deploy a tri-spectral imaging device that combines visible light, short-wave infrared, and coded structured light to simultaneously capture weld surface texture, molten pool heat distribution, and three-dimensional morphology information. S12. Perform frequency domain noise separation on the weld surface texture image, perform spatial filtering on the three-dimensional topography image, and enhance the thermal radiation characteristics of the molten pool heat distribution image. S13. Based on the preset calibration parameters, the surface texture of the weld, the heat distribution of the molten pool and the three-dimensional topography point cloud are mapped to a unified coordinate system. Sub-pixel level spatial alignment is achieved through feature point matching and affine transformation to generate a registration dataset with texture-temperature-coordinate binding. S14. Convert the point cloud into a discrete 3D mesh, and fuse texture color values, temperature gradient values ​​and spatial coordinate values ​​within each voxel to construct a 3D voxel model containing multi-channel attributes.

3. The vision-based intelligent welding quality assessment method according to claim 1, characterized in that, The method employs a multi-scale feature pyramid to output feature maps, constructing a two-stage decoupled framework for coarse detection and fine classification. The parallel processing channels for the initial screening results in the coarse detection stage and the corresponding fine classification stage are defined as follows: S211. Construct a unified feature extraction backbone network, input feature maps from different paths into a multi-scale feature pyramid, and achieve multi-scale feature fusion through top-down and horizontal connections to generate three sets of outputs. S212. Based on the feature map of the multi-scale feature pyramid, the spatial location and risk label of three types of preliminary screening results are generated by deformable convolution to realize deformation adaptive region proposal and risk classification; wherein, the preliminary screening results include regular welds, abnormal defects and process abnormalities. S213. Based on the spatial location and risk label of the coarse detection output, crop the local feature map from the multi-scale feature pyramid, activate the corresponding fine classification parallel processing channel and route the relevant feature data. S214. A phased training mechanism is adopted to freeze the parameters of the backbone network. A gradient blocking layer is added to the output layer of the backbone network to limit the backpropagation interference of the fine classification task with the coarse detection feature extraction.

4. The vision-based intelligent welding quality assessment method according to claim 1, characterized in that, The extracted geometric and deep semantic features from the 3D voxel model, combined with a cross-modal attention mechanism, output the defect type classification result and defect confidence vector of the weld, including: S24. Reconstruct the coordinate system with the center of the weld pool as the origin, perform spatial alignment and normalization on the three-dimensional voxel model, perform histogram equalization on the texture and temperature channels, and generate a normalized voxel model. S25. Extract geometric features and deep semantic features in parallel from the standardized voxel model, and fuse visible light texture and infrared gradient features for abnormal defect regions. Enhance the residual features of small sample defects through channel interaction attention. S26. Based on the initial screening results of the coarse detection stage, assign them to the parallel processing channels corresponding to the fine classification stage, execute the channel-specific classification task, and output the defect type label and defect confidence vector. S27. Establish a feature data cache pool, dynamically release memory resources according to the priority of classification results, and provide feedback on the load status of each channel to trigger dynamic resource allocation adjustment.

5. The vision-based intelligent welding quality assessment method according to claim 1, characterized in that, The dynamic evaluation model for welding quality index, based on defect confidence vectors and combined with real-time process parameters and molten pool status, outputs time-segmented quality scores through stage-weighted calculation, including: S31. Based on welding process characteristics, different dynamic time periods are divided, and the defect confidence vector, real-time process parameters and molten pool status are time-stamped and mapped to spatial coordinates to achieve precise alignment. S32. Based on the characteristics of the welding stage, dynamically adjust the contribution weights of different input data, and generate an adaptively optimized comprehensive feature vector by dynamically weighting the feature vectors of the three types of input data. S33. Construct a real-time dynamic evaluation model, map the comprehensive feature vector to the welding quality index of the standard interval, and calculate the time period score and mark the location of the critical defect time period.

6. The vision-based intelligent welding quality assessment method according to claim 5, characterized in that, The process of dynamically adjusting the contribution weights of different input data based on the characteristics of the welding stage, and generating an adaptively optimized comprehensive feature vector by dynamically weighting the feature vectors of the three types of input data, includes: S321. Based on the molten pool state timing pattern and process parameter change rate, identify the current welding stage and output the stage label and stage confidence level. S322. For the three types of input data, set quality sensitivity factors, call preset weights according to stage labels, and iteratively evaluate the contribution of each type of input data to the welding quality index to generate contribution weights. S323. Based on the contribution weights of the output after iteration, the feature vectors of the three types of input data are gated and weighted to generate a comprehensive feature vector that is resistant to interference. S324. By adjusting the prediction bias of the fusion welding quality index, the quality-sensitive factors in the iterative process are corrected in reverse, thereby achieving closed-loop optimization of the contribution weight.

7. The vision-based intelligent welding quality assessment method according to claim 5, characterized in that, The construction of a real-time dynamic evaluation model maps the comprehensive feature vector to a welding quality index for a standard interval, and statistically analyzes time-period scores, marking the location of key defect time periods, including: S331. Adaptively standardize the input comprehensive feature vector, embed welding physical rule constraints, and fuse temporal and spatial features through attention-enhanced gated recurrent unit model to dynamically map the normalized feature vector into the welding quality index of the standard interval. S332. Based on the time series of welding quality index, dynamically calculate the time period score according to the welding stage, and mark the high defect incidence window through statistical process; S333. Based on the deviation between the actual flaw detection results and the predicted welding quality index, the normalization parameters of the attention-enhanced gating loop unit model are automatically adjusted to generate a multi-dimensional quality assessment dashboard, locate the defect window position in real time, and trigger graded early warning signals.

8. The vision-based intelligent welding quality assessment method according to claim 7, characterized in that, The formula for calculating the time period score is as follows: ； In the formula, Q T Indicates the score for the current time period; WQI t express t Welding quality score at any given moment; R c This indicates the compliance rate score; S m Indicates the state index; D d Indicates the defect density factor; T Indicates the division of welding time periods; α , β , γ , δ All of these represent preset weighting coefficients.

9. The vision-based intelligent welding quality assessment method according to claim 5, characterized in that, The construction of a real-time dynamic evaluation model, which maps the comprehensive feature vector to a welding quality index for a standard interval, and statistically analyzes time-period scores and marks the location of key defect time periods, also includes: S34. Divide the defect confidence vector into multiple confidence labels, combine them with the welding quality index, identify the target sensitive area to be re-inspected, and allocate re-inspection resources. S35. Generate process optimization parameters based on defect type and quality assessment results, input them into the welding digital twin, simulate the thermal deformation trend after parameter optimization, and drive the welding equipment to dynamically adjust parameters. The process of dividing multiple confidence labels based on the distribution characteristics of the defect confidence vector, combining this with the welding quality index, identifying the target sensitive area to be re-inspected, and allocating re-inspection resources includes: S341. Divide the defect confidence vector into multi-level confidence labels, integrate the confidence labels, welding quality index and process fluctuations, identify the sensitive area clusters for re-inspection through spatiotemporal constraint density clustering, and locate the center coordinates of the sensitive area clusters. S342. Based on the label density, average welding quality index, and area ratio of the sensitive area cluster, a risk score is calculated by weighting. When the risk score is greater than a set threshold, a local re-inspection is triggered. S343. Adjust the grading threshold of the confidence label based on the re-inspection results.

10. A vision-based intelligent welding quality assessment system, used to implement the vision-based intelligent welding quality assessment method according to any one of claims 1-9, characterized in that, include: The synchronous acquisition module is used to monitor the welding process in real time using multimodal imaging equipment, and simultaneously acquire information on weld surface texture, molten pool heat distribution and three-dimensional morphology to construct a three-dimensional voxel model; The multi-channel processing module is used to build a two-stage multi-channel architecture for executing dynamic resource scheduling, extract geometric features and deep semantic features from the three-dimensional voxel model, and combine cross-modal attention mechanism to output the defect type classification results and defect confidence vector of the weld. The quality assessment module is used to construct a dynamic assessment model of welding quality index based on defect confidence vector, combined with real-time process parameters and molten pool status, and output time-segmented quality scores through stage weighted calculation. The synchronous acquisition module, the multi-channel processing module, and the quality assessment module are sequentially connected.

Images