A vision-based intelligent assessment method and system for welding quality

Abstract

This invention discloses a vision-based intelligent welding quality assessment method and system, belonging to the field of welding quality assessment technology. The method includes: using multimodal imaging equipment to monitor the welding process in real time, simultaneously acquiring weld surface texture, molten pool heat distribution, and three-dimensional morphology information, and constructing a three-dimensional voxel model; building a two-stage, multi-channel architecture for dynamic resource scheduling, extracting geometric features and deep semantic features from the three-dimensional voxel model; and constructing a dynamic assessment model of the welding quality index based on defect confidence vectors, combined with real-time process parameters and molten pool status, and outputting time-segmented quality scores through stage-weighted calculation. This invention synchronously acquires weld surface texture, molten pool heat distribution, and three-dimensional morphology information using multimodal imaging equipment, constructing a texture-temperature-spatial coordinate bound three-dimensional voxel model, and realizing multi-source heterogeneous data fusion analysis of welding quality.

Description

Technical Field

[0001] This invention relates to the field of welding quality assessment technology, and in particular to a vision-based intelligent welding quality assessment method and system. Background Technology

[0002] Welding technology, as one of the core processes in modern manufacturing, has been widely applied in key areas such as automotive body-in-white manufacturing, aerospace precision component joining, new energy battery module packaging, and pressure vessel welding. With the increasing demand for industrial automation, the quality and stability of the welding process directly affect the structural strength and safety of products. Intelligent welding and assembly technologies have achieved large-scale application in multiple industrial sectors, demonstrating significant value, especially in the manufacturing of large structural components. In the field of water conservancy engineering, such as the installation of pressure steel pipes at the Huangjinping Hydropower Station, intelligent assembly welding machines achieve high-precision welding through a digital control system, solving the problems of low efficiency and unstable quality associated with traditional manual welding.

[0003] However, traditional welding quality assurance mainly relies on post-weld destructive testing (such as tensile testing) or non-destructive sampling inspection (such as X-ray flaw detection and ultrasonic testing). These methods have obvious limitations: first, the detection is lagging, and it is difficult to correct defects in real time after they are formed, resulting in a high scrap rate; second, they are costly and inefficient. For example, full inspection of LNG storage tank welds requires a lot of manpower and equipment resources.

[0004] Existing technologies are gradually shifting towards online monitoring of the welding process, primarily relying on multi-source sensor information analysis. For example, statistical characteristics of welding current / voltage signals are collected to predict droplet transfer stability; the time-frequency characteristics of arc sound signals are combined with SVM models to identify penetration status; or infrared thermal imaging is used to monitor the temperature field gradient of the molten pool and capture local temperature drops caused by porosity. However, in achieving real-time monitoring of welding, there are significant limitations of single-mode monitoring. For instance, electrical signals are susceptible to electromagnetic interference, arc sound has a low signal-to-noise ratio in noisy workshops, and infrared thermal imaging is difficult to quantify microscopic defects such as cracks. Furthermore, it has poor adaptability to complex working conditions; changes in the three-dimensional curvature or overlapping areas of automotive body welds can easily lead to visual misjudgments, misdiagnosing design features as defects. Dynamic process response is insufficient; existing algorithms mostly rely on static expert database comparisons and lack real-time modeling capabilities for transient molten pool deformation and material thermal deformation during the arc initiation / outgoing phases of welding. Moreover, changes in lighting conditions significantly reduce the reliability of grayscale and gradient features. Summary of the Invention

[0005] Therefore, it is necessary to provide a vision-based intelligent evaluation method and system for welding quality to address the aforementioned technical problems.

[0006] In a first aspect, the present invention provides a vision-based intelligent assessment method for welding quality, comprising:

[0007] The welding process is monitored in real time using multimodal imaging equipment, and the surface texture, heat distribution of the weld pool and three-dimensional morphology information of the weld are acquired simultaneously to construct a three-dimensional voxel model.

[0008] A two-stage, multi-channel architecture for dynamic resource scheduling is constructed, geometric features and deep semantic features are extracted from the three-dimensional voxel model, and a cross-modal attention mechanism is combined to output the defect type classification results and defect confidence vector of the weld.

[0009] 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.

[0010] Building a two-stage, multi-channel architecture for dynamic resource scheduling includes:

[0011] A multi-scale feature pyramid is used to output feature maps, and a two-stage decoupled framework for coarse detection and fine classification is constructed. The initial screening results of the coarse detection stage and the corresponding parallel processing channels for the fine classification stage are set.

[0012] 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.

[0013] 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 achieve computational power isolation and sharing of different parallel processing channels through thread parallel control.

[0014] Furthermore, multimodal imaging equipment is used to monitor the welding process in real time, simultaneously acquiring information on weld surface texture, internal molten pool structure, and three-dimensional morphology, and constructing a three-dimensional voxel model, including:

[0015] 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;

[0016] Frequency domain noise separation is performed on the weld surface texture image, spatial filtering is performed on the three-dimensional topography image, and the thermal radiation characteristics of the molten pool heat distribution image are enhanced.

[0017] Based on preset calibration parameters, the surface texture of the weld, the heat distribution of the molten pool and the point cloud of the three-dimensional topography are mapped to a unified coordinate system. Sub-pixel level spatial alignment is achieved through feature point matching and affine transformation, generating a registration dataset with texture-temperature-coordinate binding.

[0018] The point cloud is converted into a discrete 3D mesh, and texture color values, temperature gradient values ​​and spatial coordinate values ​​are fused within each voxel to construct a 3D voxel model containing multi-channel attributes.

[0019] Furthermore, a multi-scale feature pyramid is used to output feature maps, constructing a two-stage decoupled framework for coarse detection and fine classification. The initial screening results of the coarse detection stage and the corresponding parallel processing channels for the fine classification stage are defined as follows:

[0020] A unified feature extraction backbone network is constructed, and feature maps from different paths are input into a multi-scale feature pyramid. Multi-scale feature fusion is achieved through top-down and horizontal connections to generate three sets of outputs.

[0021] 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, realizing deformation adaptive region proposal and risk classification; among them, the preliminary screening results include regular welds, abnormal defects and process abnormalities.

[0022] Based on the spatial location and risk label of the coarse detection output, local feature maps are cropped from the multi-scale feature pyramid, the corresponding fine classification parallel processing channels are activated, and the relevant feature data is routed.

[0023] A phased training mechanism is adopted to freeze the parameters of the backbone network, and a gradient blocking layer is added to the output layer of the backbone network to limit the backpropagation interference of the fine classification task to the coarse detection feature extraction.

[0024] Furthermore, geometric features and deep semantic features are extracted from the 3D voxel model, and combined with a cross-modal attention mechanism, the defect type classification results and defect confidence vectors of the weld are output, including:

[0025] The coordinate system is reconstructed with the center of the weld pool as the origin. The three-dimensional voxel model is spatially aligned and normalized. The texture and temperature channels are histogram equalized to generate a normalized voxel model.

[0026] Geometric features and deep semantic features are extracted in parallel from the standardized voxel model, and visible light texture and infrared gradient features are fused to abnormal defect regions. Residual feature enhancement of small sample defects is achieved through channel interaction attention.

[0027] Based on the initial screening results of the coarse detection stage, the results are assigned to the parallel processing channel corresponding to the fine classification stage, and the channel-specific classification task is executed to output the defect type label and defect confidence vector.

[0028] 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.

[0029] Furthermore, based on the defect confidence vector and combined with real-time process parameters and molten pool status, a dynamic evaluation model for the welding quality index is constructed. The model outputs time-segmented quality scores through stage-weighted calculation, including:

[0030] Based on the characteristics of welding process, 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.

[0031] Based on the characteristics of the welding stage, the contribution weights of different input data are dynamically adjusted, and an adaptively optimized comprehensive feature vector is generated by dynamically weighting the feature vectors of the three types of input data.

[0032] A real-time dynamic evaluation model is constructed, which maps the comprehensive feature vector to the welding quality index of the standard interval, and calculates the time period score and marks the time period location of key defects.

[0033] Furthermore, based on the characteristics of the welding stage, the contribution weights of different input data are dynamically adjusted. By dynamically weighting the feature vectors of the three types of input data, an adaptively optimized comprehensive feature vector is generated, including:

[0034] Based on the molten pool state timing pattern and process parameter change rate, the current welding stage is identified, and the stage label and stage confidence level are output.

[0035] For the three types of input data, quality sensitivity factors are set, preset weights are called according to the stage labels, and the contribution of each type of input data to the welding quality index is evaluated iteratively to generate contribution weights;

[0036] 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.

[0037] By integrating the prediction bias of the welding quality index, the quality-sensitive factors in the iterative process are corrected in reverse, thereby achieving closed-loop optimization of the contribution weights.

[0038] Furthermore, a real-time dynamic evaluation model is constructed, mapping the comprehensive feature vector to a welding quality index for a standard interval, and statistically analyzing time-period scores, marking the locations of key defect time periods, including:

[0039] The input comprehensive feature vector is adaptively standardized, and welding physical rules are embedded. Then, the temporal and spatial features are fused through the attention-enhanced gated recurrent unit model, and the normalized feature vector is dynamically mapped to the welding quality index of the standard interval.

[0040] Based on the time series of welding quality index, the period score is dynamically calculated according to the welding stage, and the high-incidence window of defects is marked by statistical process.

[0041] 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.

[0042] Furthermore, the formula for calculating the time period score is as follows:

[0043] ;

[0044] 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.

[0045] Furthermore, a real-time dynamic evaluation model is constructed, mapping the comprehensive feature vector to a welding quality index for a standard interval, and statistically analyzing time-period scores. After marking the location of key defect time periods, the model also includes:

[0046] Based on the distribution characteristics of the defect confidence vector, multiple confidence labels are divided, and combined with the welding quality index, the target sensitive areas to be re-inspected are identified, and re-inspection resources are allocated.

[0047] Based on the defect type and quality assessment results, process optimization parameters are generated and input into the welding digital twin. The thermal deformation trend after parameter optimization is simulated, and the welding equipment is driven to dynamically adjust the parameters.

[0048] This involves dividing the defect confidence vector into multiple confidence labels, combining them with the welding quality index, identifying the target sensitive areas to be re-inspected, and allocating re-inspection resources, including:

[0049] The defect confidence vector is divided into multi-level confidence labels. The confidence labels, welding quality index and process fluctuations are integrated. The density clustering with spatiotemporal constraints is used to identify the clusters of sensitive areas for re-inspection and to locate the center coordinates of the clusters of sensitive areas.

[0050] Based on the label density, average welding quality index, and area ratio of sensitive area clusters, a risk score is calculated using weighted average. When the risk score exceeds a set threshold, a local re-inspection is triggered.

[0051] The grading threshold for confidence labels was adjusted based on the re-inspection results.

[0052] Secondly, the present invention also provides a vision-based intelligent welding quality assessment system, comprising:

[0053] 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;

[0054] 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.

[0055] 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.

[0056] The synchronous acquisition module, multi-channel processing module, and quality assessment module are connected sequentially.

[0057] The beneficial effects of this invention are as follows:

[0058] 1. By simultaneously acquiring weld surface texture, molten pool heat distribution, and three-dimensional morphology information through multimodal imaging equipment, a three-dimensional voxel model with texture-temperature-spatial coordinate binding is constructed to achieve multi-source heterogeneous data fusion analysis of welding quality. Combining real-time process parameters and molten pool status, a dynamic evaluation model of welding quality index is constructed. Through stage-weighted calculation, time-segmented quality scores are output, significantly improving the comprehensiveness and real-time performance of the evaluation. Furthermore, a defect confidence grading mechanism and a dynamic allocation strategy for re-inspection resources are introduced to form a full-process monitoring and evaluation of welding inspection, evaluation, and optimization. This drives dynamic adjustment of welding parameters and precise prevention and control of defect risks, effectively avoiding evaluation biases caused by neglecting the dynamics of welding in traditional methods.

[0059] 2. To address the quality change characteristics at different stages of the welding process, an adaptive adjustment mechanism for stage weights is introduced. The current welding stage is identified by the molten pool state time sequence pattern and the rate of change of process parameters. The contribution weights of three types of data—defect confidence vector, process parameters, and molten pool state—are dynamically allocated to generate a comprehensive feature vector that resists interference. This solves the problem that traditional static evaluation models cannot adapt to the non-uniformity of the welding process. In particular, it enhances the sensitivity to the unstable state at the start of the arc and the molten pool contraction effect at the end of the arc, ensuring that high-incidence defect periods are accurately marked and significantly improving the robustness of key defect detection.

[0060] 3. By constructing a two-stage processing framework that decouples coarse detection and fine classification, and combining cross-modal attention mechanism with dynamic resource allocation of parallel processing channels, intelligent scheduling of computing resources and optimization of detection efficiency are achieved. In the coarse detection stage, high-risk areas are quickly screened using multi-scale feature pyramids and deformable convolutions and routed to the fine classification channel. In the fine classification stage, dedicated classification channels are activated according to task priority, small sample defect features are enhanced through channel interaction attention, and resources are dynamically released based on the load status of the cache pool, thereby meeting the real-time requirements of high-speed welding. Attached Figure Description

[0061] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0062] Figure 1 This is a flowchart of a vision-based intelligent welding quality assessment method according to an embodiment of the present invention;

[0063] Figure 2 This is a system principle block diagram of a vision-based intelligent welding quality assessment system according to an embodiment of the present invention.

[0064] The reference numerals are: 1. Synchronous acquisition module; 2. Multi-channel processing module; 3. Quality assessment module. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] Please see Figure 1 This paper provides a vision-based intelligent assessment method for welding quality, including:

[0067] 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.

[0068] In the description of this invention, the welding process is monitored in real time using a multimodal imaging device, and the surface texture, internal molten pool structure, and three-dimensional morphology information of the weld are acquired simultaneously to construct a three-dimensional voxel model, including:

[0069] 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.

[0070] 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 1.4-3μm) penetrates the oxide layer to capture the thermal radiation signal inside the molten pool; the structured light imaging device (DLP projector) projects a combination of red and black pseudo-random dot matrix and monochrome sinusoidal stripe coding pattern, and compensates for smoke and dust obstruction through phase calculation.

[0071] The three-spectral imaging device generates a synchronous pulse signal based on FPGA, which simultaneously triggers the exposure of the three-modal devices; the acquisition time of each modality data is recorded by hardware timestamp to ensure timing alignment.

[0072] 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.

[0073] Specifically, frequency domain noise separation is performed on visible light images (weld surface texture images), spatial filtering is performed on structured light point clouds (three-dimensional topography images), and the thermal radiation characteristics of short-wave infrared images (molten pool heat distribution images) are enhanced to solve the problems of arc overexposure, smoke scattering and thermal signal attenuation, and output a high signal-to-noise ratio enhanced multispectral image set.

[0074] Visible light frequency domain processing requires performing a Fourier transform on the original visible light image to separate high-frequency arc noise from low-frequency weld texture, and retaining the low-frequency components to reconstruct a clear surface image.

[0075] The structured light point cloud optimization uses the moving least squares (MLS) method to smooth the point cloud and filter out outliers caused by splash particles; and reconstructs the shape data of the area obscured by smoke and dust based on the phase unwrapping algorithm.

[0076] Short-wave infrared feature enhancement requires non-uniformity correction (NUC) of the original thermal image to eliminate detector response differences; histogram equalization is used to improve the contrast between the molten pool area and the background.

[0077] S13. Based on 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.

[0078] Specifically, based on preset calibration parameters, visible light textures, short-wave infrared thermal images, and structured light 3D point clouds are mapped to a unified coordinate system. Subpixel-level spatial alignment is achieved through feature point matching and affine transformation, generating a registration dataset that binds "texture-temperature-coordinates", laying a spatial consistency foundation for 3D voxel fusion.

[0079] The intrinsic parameter matrix of each camera and the extrinsic parameter matrix relative to the world coordinate system are calculated using a checkerboard calibration board; the calibration parameters are optimized by bundle adjustment to reduce reprojection error; SIFT feature points are extracted from visible light and infrared images; mismatched points are eliminated by RANSAC algorithm; and the affine transformation matrix is ​​solved to map the pixel coordinates of 2D images to 3D point cloud spatial coordinates.

[0080] 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.

[0081] Specifically, the point cloud is first voxelized, and the AABB bounding box is calculated based on the spatial distribution of the structured light point cloud to set the voxel resolution. The point cloud is discretized into a voxel grid using a GPU-accelerated octree index, and the average spatial coordinates of each voxel are stored.

[0082] The 3D voxel model contains multi-channel attributes. For multi-channel attribute fusion, the following are included:

[0083] 1. Texture Channel: Interpolates and maps the RGB values ​​of the registered visible light image to the corresponding voxels;

[0084] 2. Temperature Channel: Converts the radiation intensity values ​​of shortwave infrared thermal images into temperature values ​​and writes them into voxels;

[0085] 3. Shape Channel: Records the three-dimensional coordinates and normal vector of the voxel center point.

[0086] 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.

[0087] In the description of this invention, a two-stage, multi-channel architecture for dynamic resource scheduling is constructed, geometric features and deep semantic features are extracted from the three-dimensional voxel model, and a cross-modal attention mechanism is combined to output the defect type classification result and defect confidence vector of the weld, including:

[0088] S21. Employ 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.

[0089] In the description of this invention, a multi-scale feature pyramid is used to output feature maps, constructing a two-stage decoupled framework for coarse detection and fine classification. The initial screening results of the coarse detection stage and the parallel processing channels corresponding to the fine classification stage are defined as follows:

[0090] 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.

[0091] Specifically, by designing a unified feature extraction backbone network, a multi-scale feature pyramid (FPN) is used to simultaneously output high-resolution geometric features and low-resolution semantic features, providing a shared feature base for coarse detection and fine classification, and avoiding redundant computation in two stages.

[0092] Hierarchical convolutional paths include high-resolution paths and low-resolution paths.

[0093] The high-resolution path uses a lightweight MobileNetV3 module stacked with 3 layers of deep separable convolutions to output a 1 / 8 resolution feature map (256×256 pixels), focusing on geometric details such as weld contour edges and misalignment.

[0094] A 4-layer ResNet-18 residual module is deployed in the low-resolution path, which is combined with dilated convolution (dilation=2) to expand the receptive field and output a 1 / 32 resolution feature map (64×64 pixels) to capture semantic information such as the melt pool temperature gradient field and texture mutation.

[0095] The feature maps of high / low resolution paths are input into the FPN, and multi-scale feature fusion is achieved through top-down and horizontal connections to generate three sets of outputs (1 / 8, 1 / 16, and 1 / 32 resolutions) to cover different scale detection needs.

[0096] S212. Based on the multi-scale feature pyramid feature map, spatial locations and risk labels of three types of preliminary screening results are generated through deformable convolution, realizing deformation-adaptive region proposal and risk classification. Among them, the preliminary screening results include regular welds, abnormal defects, and process anomalies.

[0097] Specifically, three sets of deformable convolutions (DCNv2) are embedded in the YOLOv8 inspection head to dynamically learn the weld deformation offset and improve robustness to irregular abrupt changes at the molten pool boundary.

[0098] The judgment logic and output format for the initial screening results, which include regular welds, abnormal defects, and process abnormalities, are explained below:

[0099] 1. For regular welds, the judgment logic is that the geometric parameters meet the standard (width error < threshold) and the contour smoothness ≥ threshold. The output format is a high confidence label (Confidence > 0.95).

[0100] 2. Abnormal defects are judged based on the spatial overlap between local texture mutation areas and temperature gradient abnormal areas, and the output is a defect candidate box (IoU>0.7).

[0101] 3. Process abnormality: The judgment 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. The output format is an abnormality mark + associated sensor ID.

[0102] 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.

[0103] Specifically, fine classification requires the establishment and binding of corresponding image processing channels. The channel binding strategy is explained below:

[0104] 1. Regular weld seam channel: Activate FPN1 / 8 high-resolution feature map routing, only receive geometric feature vectors of high-resolution feature maps, and force lightweight verification (MobileNetV3 classification head + B-spline contour fitting).

[0105] 2. Abnormal Defect Channel: Requires dual-modal input (visible light texture + infrared thermal image), forces the fusion of visible light texture (1 / 8 layer) and infrared thermal image (1 / 32 layer) features, and enables cross-modal attention mechanism;

[0106] 3. Process Anomaly Channel: Bind the timing data stream of the current sensor, input the coordinate sequence of the sudden temperature change point of the molten pool, and force the use of the temporal convolutional network (TCN) to analyze the root cause.

[0107] 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.

[0108] 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.

[0109] Specifically, based on the feature maps output by the multi-scale feature pyramid (including 1 / 8, 1 / 16, and 1 / 32 resolution feature maps), a deformable convolutional module (DCNv2) is deployed in the coarse detection stage to handle the geometric deformation problem of the weld. The specific implementation process is as follows:

[0110] 1. Deformation Offset Learning: For the multi-scale feature map output by FPN, the sampling point offset (Δ) is dynamically predicted using deformable convolutional kernels. x Δ y This allows the convolutional kernel to adapt to irregular deformations at the molten pool boundary (such as weld distortion caused by thermal deformation). The offset is generated by an additional convolutional layer, whose input is the geometric context information of the current feature map.

[0111] 2. Region Proposal Generation: Deformable convolution is used to slide and scan across feature maps at three scales, outputting the spatial locations of the three initial screening results, as detailed below:

[0112] 2.1 For regular welds, when the weld width error is less than the preset threshold (e.g., ±0.1mm) and the contour smoothness is greater than or equal to 0.9, output a high-confidence bounding box (Confidence>0.95).

[0113] 2.2 Detecting abnormal defects, including local texture abrupt changes (Laplacian response > threshold) and abnormal temperature gradient areas. The spatially overlapping regions are used to output defect candidate boxes (IoU>0.7).

[0114] 2.3. For process anomalies, analyze the region where the current fluctuation frequency is greater than the threshold (e.g., 10Hz) and the correlation coefficient with the temperature gradient is less than 0.3, and output the anomaly marker point and the associated sensor ID.

[0115] 3. Risk grading mechanism: Risk scores are calculated based on spatial location characteristics, as detailed below:

[0116] 3.1 Location confidence, based on the matching degree between deformable convolution sampling points and the actual weld edge (Hausdorff distance ≤ 0.05 mm).

[0117] 3.2 Risk Level: The risk level is determined by integrating the geometric deformation amplitude (curvature change rate) and thermodynamic parameters (temperature gradient entropy), and outputting three risk levels: low (L1), medium (L2), and high (L3).

[0118] 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.

[0119] Specifically, during the fine classification stage, the computing resource allocation is preset, for example, 70% for the abnormal defect channel, 20% for the process abnormal channel, and 10% for the regular weld channel. The GPU core utilization is controlled through CUDA streaming. The task queue status of each channel is monitored in real time. If a channel is idle (e.g., the regular weld queue is empty), its computing power is dynamically scheduled to other high-load channels (e.g., the abnormal defect channel), and resource conflicts are avoided through thread locks.

[0120] The parallel control mechanism creates an independent CUDA stream for each type of channel, schedules tasks through a priority queue (abnormal defect tasks > process abnormalities > regular welds), and designs an idle resource reallocator that automatically transfers the computing power quota of a channel to a higher priority channel when the task queue of a certain channel is empty.

[0121] 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.

[0122] Specifically, a coordinate system is reconstructed with the center of the weld pool as the origin, and the voxel model is translated to the center of the standard coordinate system through affine transformation. The principal axis direction of the weld is calculated based on principal component analysis (PCA), and the model is rotated to align the weld length direction with the Z-axis. The original voxels are resampled to a resolution of 0.1 mm³ using trilinear interpolation to ensure comparability of geometric scales for different welds, and histogram equalization is performed on the texture (RGB) and temperature (infrared) channels to enhance feature consistency.

[0123] 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.

[0124] Specifically, geometric features (contour curvature, misalignment) and deep semantic features (melt pool temperature gradient field, texture mutation mode) are extracted in parallel from the standardized voxel model, and feature interference is avoided through dual-path isolation.

[0125] For the geometric feature path, the contour curvature is first calculated. The weld edge point set is extracted in the XY section, and the second derivative is obtained by fitting the B-spline curve to output the maximum curvature value and position. Then, the misalignment is quantified by calculating the Hausdorff distance of the height difference between the base materials on both sides of the weld and outputting the absolute value of the misalignment.

[0126] For semantic feature paths, a temperature gradient field is first generated by scanning voxels in the molten pool region using a 3×3×3 Sobel operator, and the temperature change rate vector is output. Then, texture mutation detection is performed, and the 3D Laplacian operator is applied to the voxels of the visible light channel to enhance the edge response of defects such as pores and cracks.

[0127] Furthermore, based on a pre-defined channel protocol, visible light texture and infrared thermal gradient features are fused in abnormal defect regions. Channel interactive attention (CI-Attention) is used to enhance the cross-material generalization ability of small sample defects. In the implementation process, the infrared thermal image needs to be mapped to the visible light texture coordinate system through bilinear interpolation to achieve pixel-level spatial alignment. The visible light branch generates a query vector through 1×1 convolution, and the infrared branch generates a key-value vector. The cross-modal attention weight matrix is ​​calculated. The weighted features are added to the original visible light features, preserving the original texture details while strengthening the thermal gradient associated regions (such as the temperature drop area around the pores).

[0128] 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.

[0129] Specifically, based on the initial screening results output from the coarse inspection stage (including the spatial location and risk labels of three categories of areas: regular welds, abnormal defects, and process abnormalities), the following fine classification task allocation and processing flow is executed:

[0130] 1. Channel activation and feature routing: Activate the corresponding fine classification parallel processing channel based on the risk label.

[0131] 1.1 Regular weld seam region: Activate FPN1 / 8 high-resolution feature map routing, only receive geometric feature vectors (such as contour curvature, misalignment amount), perform B-spline contour fitting verification through lightweight MobileNetV3 classification head to confirm the compliance of geometric parameters.

[0132] 1.2 Abnormal Defect Regions: Visible light texture (FPN1 / 8 layers) and infrared thermal image (FPN1 / 32 layers) features are forcibly fused, enabling a cross-modal attention mechanism (CI-Attention). Specifically, the infrared thermal image is aligned to the visible light coordinate system via bilinear interpolation. The visible light branch generates a query vector, and the infrared branch generates a key-value vector. After calculating the attention weight matrix, the features are weighted and fused to strengthen the texture-thermal gradient correlation in defect regions (such as the temperature drop zone around pores).

[0133] 1.3 Process Abnormal Area: Bind the time-series data stream of the current sensor, input the coordinate sequence of the sudden temperature change point of the molten pool, and use a temporal convolutional network (TCN) to analyze the root cause of the correlation between current fluctuation and temperature gradient.

[0134] 2. Dynamic Resource Scheduling: Computational resources are allocated according to a preset priority (abnormal defects > process abnormalities > regular welds): the abnormal defect channel occupies 70% of the GPU computing power (FP16 precision), the process abnormal channel occupies 20% (INT8 quantization), and the regular weld channel occupies 10%. Computational power isolation is achieved through independent CUDA streams, and the task queue status is monitored in real time. If a channel is idle (e.g., the regular weld queue is empty), its computing power is dynamically scheduled to the high-load channel (e.g., the abnormal defect channel) through a thread lock mechanism, and the computing power quota is transferred using the idle resource reallocator.

[0135] 3. Feature standardization and classification output: Standardize the input features of each channel: Reconstruct the coordinate system with the center of the molten pool as the origin, align the main axis of the weld seam through PCA, and perform histogram equalization on the texture (RGB) and temperature (infrared) channels respectively to ensure feature consistency.

[0136] Specifically, for channel-specific tasks: the anomaly / defect channel extracts continuous edge features of minute defects through a residual feature enhancement module, and outputs defect type labels (porosity / crack / lack of fusion) by combining cross-modal attention. The process anomaly channel uses a temporal convolutional network (TCN) to analyze the sequence of abrupt temperature changes in the molten pool and outputs anomaly root cause labels (such as current overload or uneven cooling).

[0137] 4. Result Generation and Resource Release: Each channel outputs a defect type label and a multi-dimensional confidence vector. The vector elements represent the following aspects:

[0138] 4.1 Class Probability: The likelihood (0~1) that a defect belongs to the target class;

[0139] 4.2 Positioning reliability: Spatial position error score based on Hausdorff distance;

[0140] 4.3 Morphological integrity: Defect contour closure and thermal gradient continuity score.

[0141] 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.

[0142] Specifically, a feature data cache pool is established, memory resources are dynamically released according to the priority of classification results, and the channel load status is fed back to the resource scheduling engine to achieve computing power rebalancing.

[0143] For example, the following can be used as a reference for tiered caching strategies:

[0144] 1. For high-confidence results (confidence level > 95%), only the feature vector and classification label are saved, and the original voxel data is released immediately;

[0145] 2. Low confidence results (confidence level ≤ 80%) retain the original voxel data and intermediate features for iterative analysis of the S24 channel.

[0146] Real-time statistics of processing latency for each channel are collected. If the average latency of an abnormal or defective channel is greater than 20ms, the S214 resource scheduling engine is triggered to increase the GPU core quota.

[0147] 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. The quality score is output in different time periods through stage weighted calculation, and a confidence level mechanism is introduced to trigger the allocation of re-inspection resources.

[0148] In the description of this invention, a dynamic evaluation model for welding quality index is constructed based on the defect confidence vector, combined with real-time process parameters and molten pool state. The model outputs time-segmented quality scores through stage-weighted calculation and introduces a confidence grading mechanism to trigger re-inspection resource allocation, including:

[0149] S31. Based on the characteristics of the welding process, different dynamic time periods are divided, and the defect confidence vector, real-time process parameters and molten pool status are synchronized with time stamps and spatial coordinate mapping to achieve precise alignment.

[0150] Specifically, based on the welding process characteristics (arc initiation → stabilization → arc termination), the process can be divided into three dynamic time periods, achieving precise alignment of the defect confidence vector, process parameters, and molten pool state time axis, laying a time sequence foundation for time-segmented evaluation.

[0151] For example, the rules for dividing time periods can be referenced below:

[0152] 1. Arc initiation phase (0-10% of welding time): Current ramp-up period, monitoring the initial state of molten pool formation;

[0153] 2. Stable phase (10%-85% of welding time): The tolerance period for process parameter fluctuations, focusing on areas with high defect incidence;

[0154] 3. Arc decay phase (85%-100% of welding time): Current decay period, for detecting shrinkage defects (such as crater cracks).

[0155] 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.

[0156] In the description of this invention, based on the characteristics of the welding stage, the contribution weights of different input data are dynamically adjusted. An adaptively optimized comprehensive feature vector is generated by dynamically weighting the feature vectors of the three types of input data, including:

[0157] S321. Based on the molten pool state timing pattern and process parameter change rate, identify the current welding stage, output stage label and stage confidence level, realize accurate stage segmentation of the welding process, and provide stage basis for subsequent weight allocation.

[0158] Specifically, the dynamic characteristics of the welding process are analyzed in real time based on the time sequence pattern of the molten pool state (such as the rate of temperature change and the molten pool oscillation frequency) and the rate of change of process parameters (current slope and voltage fluctuation rate).

[0159] The molten pool temperature time-series curve is modeled using an LSTM network. Combined with the first derivative variation characteristics of process parameters (e.g., current change rate > 10A / ms is judged as arc initiation stage), the stage labels (arc initiation / stable / arc termination) and stage confidence (0-1 probability value) are output.

[0160] 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.

[0161] Specifically, for the three types of input data (molten pool state, process parameters, and defect confidence), quality sensitivity factors are set respectively: molten pool temperature gradient entropy (thermodynamic stability), current fluctuation variance (process stability), and defect density (quality risk).

[0162] The preset weight base value is called based on the stage label output by S321, and then the Gauss-Newton method is used to iteratively evaluate the impact of various data on the welding quality index. WQI The partial derivative of the partial derivative of the partial derivative is used to dynamically adjust the contribution weight. For example, during the arc initiation stage, the weight of the molten pool state data is increased from the initial 0.25 to 0.42 after 3 iterations because it contributes more to the prediction of cold cracks.

[0163] 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.

[0164] Specifically, based on the contribution weights generated iteratively, a gating mechanism is used to weight and fuse the three types of feature vectors: the melt pool state vector is multiplied by the Sigmoid gating coefficient (activated when the weight is >0.4), the process parameter vector is bound to the Tanh gating function, and the defect confidence vector is filtered for low-confidence features through ReLU gating.

[0165] The weighted vectors are concatenated into a comprehensive feature vector after channel normalization. Its anti-interference ability is reflected in the fact that when the arc changes abruptly, the gating mechanism reduces the weight of the molten pool state from 0.35 to 0.18, suppressing abnormal heat distribution noise.

[0166] 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.

[0167] Specifically, real-time comparison WQI When the predicted value deviates from the actual flaw detection result (such as the defect area deviation in X-ray inspection), a reverse correction is triggered if the prediction deviation is greater than 15%. This is achieved by calculating the partial derivative of the quality sensitivity factor with respect to the deviation using the gradient descent method, with a learning rate... η =0.01 Reverse update of the sensitivity factor in S322.

[0168] For example, when the porosity failure rate increases during the arc-ending stage, the system corrects the sensitivity coefficient of the defect density factor from 0.32 to 0.48, ultimately forming a "prediction-measurement-correction" closed loop.

[0169] 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.

[0170] In the description of this invention, a real-time dynamic evaluation model is constructed, which maps the comprehensive feature vector to a welding quality index for a standard interval, and statistically scores time periods, marking the location of key defect time periods, including:

[0171] S331. Adaptively standardize the input comprehensive feature vector, embed welding physical rule constraints, and fuse temporal and spatial features through an attention-enhanced gated recurrent unit model to dynamically map the normalized feature vector into a welding quality index for a standard interval.

[0172] 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-incidence window of defects through statistical process.

[0173] In the description of this invention, the formula for calculating the time period score is as follows:

[0174] .

[0175] 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.

[0176] It should be noted that, Q T This is a comprehensive evaluation result of the welding process at specific time points (such as the arc initiation and stabilization phases), calculated through weighted summation. Since all input terms in the formula have been 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.

[0177] 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.

[0178] 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 ).

[0179] 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.

[0180] 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.

[0181] 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.

[0182] 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.

[0183] 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:

[0184] 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.

[0185] 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).

[0186] 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.

[0187] 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.

[0188] Specifically, a comprehensive risk score is calculated for each sensitive region cluster. R This includes the following aspects:

[0189] 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;

[0190] 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 (weight). β =0.3;

[0191] 3. Area proportion factor ( A ratio ): The proportion of cluster area to the total weld area, weight γ =0.2.

[0192] 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 ( R The top 30% of clusters (in descending order of value) are given priority to occupy 70% of the computing power of the detection equipment, and the remaining resources are allocated to low- and medium-risk clusters in turn. Real-time load balancing is used to avoid detection bottlenecks.

[0193] S343. Adjust the grading threshold of the confidence label based on the re-inspection results.

[0194] Specifically, if a missed detection is found during the re-inspection of a high-confidence label area, the category probability threshold is lowered and the positioning reliability weight is increased; if the re-inspection of a low-confidence label area is correct, the morphological integrity score threshold is raised to reduce the false alarm rate.

[0195] 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.

[0196] Specifically, based on the defect type classification results and quality assessment indicators, key defect causal characteristics are extracted: if high porosity is detected, the correlation between abnormal molten pool temperature gradient and insufficient shielding gas flow rate is established; if crack confidence is high, the process parameters of edge misalignment and excessive cooling rate are combined. A reinforcement learning strategy (such as the Q-learning algorithm) is employed, with defect rate reduction and thermal deformation suppression as reward objectives, to search for the optimal solution in the process parameter space, generating optimized parameter combinations including current, voltage, and welding speed.

[0197] The optimized 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 thermo-elastic-plastic principle, time-varying parameters such as elastic modulus, yield strength, and coefficient of thermal expansion under temperature changes are calculated in real time, and the thermal deformation trend after parameter optimization is predicted.

[0198] For example, predicting the warpage of welded parts. y The reduction in 1 and the change in the area of ​​thermal stress concentration are studied. The thermal deformation process is dynamically demonstrated in a virtual welding environment and compared with real-time monitoring data from actual welding using DIC (Digital Image Correlation). If the predicted thermal deformation exceeds a safety threshold (e.g., residual stress > 70% of the material's yield strength), a reverse correction mechanism is triggered: the partial derivative of the parameter deviation with respect to thermal deformation is calculated using the gradient descent method, with a learning rate... η =0.01 Updated and optimized parameters.

[0199] The corrected optimization parameters are transmitted to the welding equipment control system via an industrial bus. Based on an adaptive control strategy, the equipment output is adjusted in real time: for example, when an abnormal molten pool temperature gradient is predicted, the current is dynamically reduced and the shielding gas flow rate is increased; when an excessive cooling rate is detected, the self-triggered reheating device is activated to reduce the cooling gradient. The system synchronously monitors the deviation between the actual welding parameters and the target values; if the deviation continues to exceed 5%, the digital twin simulation process is restarted.

[0200] Please see Figure 2It also provides a vision-based intelligent welding quality assessment system, including:

[0201] Synchronous acquisition module 1 is used to monitor the welding process in real time using multimodal imaging equipment, and synchronously acquire information on weld surface texture, molten pool heat distribution and three-dimensional morphology to construct a three-dimensional voxel model.

[0202] Multi-channel processing module 2 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.

[0203] The quality assessment module 3 is used to construct a dynamic assessment model of welding quality index based on the defect confidence vector and combined with real-time process parameters and molten pool status. It outputs time-segmented quality scores through stage weighted calculation and introduces a confidence level mechanism to trigger the allocation of re-inspection resources.

[0204] Among them, the synchronous acquisition module 1, the multi-channel processing module 2, and the quality assessment module 3 are connected in sequence.

[0205] In summary, by utilizing the above-mentioned technical solution of this invention, multimodal imaging equipment is used to simultaneously acquire information on weld surface texture, molten pool heat distribution, and three-dimensional morphology. A three-dimensional voxel model with texture-temperature-spatial coordinate binding is constructed to achieve multi-source heterogeneous data fusion analysis of welding quality. Combined with real-time process parameters and molten pool status, a dynamic evaluation model for welding quality index is constructed. Through stage-weighted calculation, time-segmented quality scores are output, significantly improving the comprehensiveness and real-time performance of the evaluation. Furthermore, a defect confidence grading mechanism and a dynamic allocation strategy for re-inspection resources are introduced to form a full-process monitoring and evaluation system for welding inspection, evaluation, and optimization. This drives dynamic adjustment of welding parameters and precise prevention and control of defect risks, effectively avoiding evaluation biases caused by neglecting the dynamics of welding in traditional methods.

[0206] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

Claims

1. A vision-based intelligent evaluation method for welding quality, characterized in that, The method comprises the following steps: S1, real-time monitoring the welding process by using a multi-modal imaging device, synchronously acquiring the surface texture of the weld, the heat distribution of the molten pool and the three-dimensional topography information, and constructing a three-dimensional voxel model; S2, building a two-stage multi-channel architecture for executing dynamic resource scheduling, extracting geometric features and deep semantic features from the three-dimensional voxel model, combining a cross-modal attention mechanism, and outputting the defect type classification result and the defect confidence vector of the weld; S3, based on the defect confidence vector, combining the real-time process parameters and the molten pool state, constructing a dynamic evaluation model of the welding quality index, and outputting the quality score of each period by stage weighting calculation; The two-stage multi-channel architecture for executing dynamic resource scheduling comprises: S21, outputting feature maps by using a multi-scale feature pyramid, constructing a two-stage decoupling framework of coarse detection and fine classification, and setting the preliminary screening result of the coarse detection stage and the corresponding parallel processing channel of the fine classification stage; S22, based on the feature maps output by the multi-scale feature pyramid, adapting to the weld deformation by using a deformable convolution, and outputting the preliminary screening result containing the spatial position and the risk level; S23, setting the input, processing and output constraints of the parallel processing channel of the fine classification stage, setting a dynamic resource allocation strategy based on the task priority, and realizing the computing power isolation and sharing of different parallel processing channels by thread parallel control; The two-stage decoupling framework of coarse detection and fine classification is constructed by using a multi-scale feature pyramid to output feature maps, and the parallel processing channel corresponding to the coarse detection stage and the fine classification stage is set, which comprises: S211, constructing a unified feature extraction backbone network, inputting feature maps of different paths into a multi-scale feature pyramid, realizing multi-scale feature fusion by top-down and horizontal connection, and generating three groups of outputs; S212, based on the multi-scale feature pyramid feature map, generating the spatial position and risk label of three types of preliminary screening results by using a deformable convolution, realizing deformation adaptive region proposal and risk grading; wherein, the preliminary screening result comprises a regular weld, an abnormal defect and a process anomaly; S213, according to the spatial position and risk label output by the coarse detection, cropping the local feature map from the multi-scale feature pyramid, activating the corresponding fine classification parallel processing channel and routing the related feature data; S214, using a stage-by-stage training mechanism to freeze the backbone network parameters, adding a gradient blocking layer at the output layer of the backbone network, and limiting the interference of the fine classification task backward propagation on the coarse detection feature extraction.

2. The method of claim 1, wherein, The method comprises the following steps: S11, deploying a three-spectrum imaging device combined by visible light, short-wave infrared and coded structured light, and synchronously capturing the surface texture of the weld, the heat distribution of the molten pool and the three-dimensional topography information; S12, performing frequency domain noise separation on the surface texture image of the weld, performing spatial filtering on the three-dimensional topography image, and enhancing the thermal radiation features of the molten pool heat distribution image; S13, map the weld surface texture, molten pool heat distribution and three-dimensional topography point cloud to a unified coordinate system based on preset calibration parameters, realize sub-pixel level spatial alignment through feature point matching and affine transformation, and generate texture-temperature-coordinate bound registration data set; S14, convert the point cloud into a discrete three-dimensional grid, fuse the texture color value, temperature gradient value and spatial coordinate value in each voxel, and construct a three-dimensional voxel model containing multi-channel attributes. 3.The visual-based welding quality intelligent evaluation method according to claim 1, characterized in that, The extraction of geometric features and deep semantic features in the three-dimensional voxel model, combined with the cross-modal attention mechanism, outputs the defect type classification result and the defect confidence vector of the weld include: S24, reconstruct the coordinate system with the weld pool center as the origin, perform spatial alignment and normalization on the three-dimensional voxel model, and perform histogram equalization on the texture and temperature channels to generate a standardized voxel model; S25, extract geometric features and deep semantic features from the standardized voxel model in parallel, and fuse visible light texture and infrared gradient features in the abnormal defect area, and realize residual feature enhancement of small sample defects through channel interaction attention; S26, according to the preliminary screening result in the coarse detection stage, assign to the corresponding parallel processing channel in the fine classification stage, execute the channel exclusive classification task, and output the defect type label and the defect confidence vector; S27, establish a feature data cache pool, dynamically release memory resources according to the classification result priority, and feed back the load state of each channel to trigger dynamic resource allocation adjustment. 4.The visual-based intelligent welding quality evaluation method according to claim 1, characterized in that, The dynamic evaluation model of the welding quality index is constructed based on the defect confidence vector, combined with real-time process parameters and molten pool state, and the time-period quality score is output through stage weighting calculation, including: S31, based on the welding process characteristics, different dynamic time 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; S32, based on the characteristics of the welding stage, the contribution weight of different input data is dynamically adjusted, and the adaptive optimized comprehensive feature vector is generated 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 in the standard interval, and count the time period score to mark the key defect period position.

5. The method of claim 4, wherein, The dynamic adjustment of the contribution weight of different input data based on the characteristics of the welding stage, the generation of the adaptive optimized comprehensive feature vector by dynamically weighting the feature vectors of the three types of input data include: S321, based on the molten pool state time sequence mode and the process parameter change rate, identify the current welding stage, output the stage label and the stage confidence; S322, 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; S323, based on the contribution weight output after iteration, gate weight fusion is performed on the feature vectors of the three types of input data to generate an anti-interference comprehensive feature vector; S324, fuse the prediction deviation of the welding quality index, and reversely correct the quality sensitive factor in the iteration process to realize closed-loop optimization of the contribution weight.

6. The method of claim 5, wherein the method further comprises: The real-time dynamic evaluation model is constructed to map the comprehensive feature vector to the welding quality index in a standard interval, and to calculate the time period score and mark the key defect time period position, including: S331, adaptive standardization processing is performed on the input comprehensive feature vector, physical rule constraints of welding are embedded, and time sequence features and spatial features are fused through attention enhanced gate recurrent unit model to dynamically map the normalized feature vector to the welding quality index in a standard interval; S332, based on the time sequence of the welding quality index, time period scores are dynamically calculated according to the welding stage, and defect high-risk windows are marked through statistical process; S333, 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, a multi-dimensional quality evaluation board is generated, the defect window position is located in real time, and a graded early warning signal is triggered.

7. The method of claim 6, wherein the method further comprises: The calculation formula of the time period score is: ； In the formula, Q T represents the current time period score; WQI t represents t weld quality score at the time instant; R c represents compliance rate score; S m represents status index; D d represents defect density factor; T represents weld period division; α 、 β 、 γ 、 δ all represent preset weight coefficients. 8.The visual-based intelligent welding quality evaluation method according to claim 4, characterized in that, After the real-time dynamic evaluation model is constructed to map the comprehensive feature vector to the welding quality index in a standard interval, and to calculate the time period score and mark the key defect time period position, the method further includes: S34, a plurality of confidence labels are divided according to the distribution characteristics of the defect confidence vector, the welding quality index is combined, the target sensitive area to be rechecked is identified, and rechecking resources are allocated; S35, process optimization parameters are generated based on the defect type and the quality evaluation result, and are input into the welding digital twin to pre-visualize the thermal deformation trend after parameter optimization, and drive the welding equipment to dynamically adjust parameters; The method according to the application further includes: S341, the defect confidence vector is 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; S342, according to the label density, the welding quality index mean value and the area proportion of the sensitive area cluster, the risk score is calculated by weighting, and when the risk score is greater than a set threshold, local rechecking is triggered; S343, the hierarchical judgment threshold of the confidence label is corrected according to the rechecking result.

9. A vision-based intelligent welding quality evaluation system for implementing the vision-based intelligent welding quality evaluation method of any one of claims 1-8, characterized in that, The method according to the application further includes: A synchronous acquisition module is configured to monitor a welding process in real time by using a multi-modal imaging device, to synchronously acquire weld surface texture, molten pool heat distribution and three-dimensional topography information, and to construct a three-dimensional voxel model; A multi-channel processing module is configured to build a two-stage multi-channel architecture for executing dynamic resource scheduling, to extract geometric features and deep semantic features in the three-dimensional voxel model, to combine a cross-modal attention mechanism, and to output a defect type classification result and a defect confidence vector of the weld; A quality evaluation module is configured to construct a dynamic evaluation model of the welding quality index based on the defect confidence vector, to combine real-time process parameters and molten pool states, and to output 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.

Images