Wind power steel pipe pile welding seam quality detection system based on machine vision

Abstract

The invention discloses a wind power steel pipe pile welding seam quality detection system based on machine vision, and the system comprises a welding seam image preprocessing module which generates a preprocessed welding seam image data set; the welding seam multi-layer semantic feature encoding module is used for extracting multi-layer semantic features and outputting a deep welding seam feature map; an improved FeatUp feature reconstruction module is used for generating a high-resolution welding seam feature map; the cavity attention fusion module is used for outputting a context guide welding seam feature map; the welding seam multi-type defect segmentation module is used for generating a welding seam defect detection result; the weld defect confidence evaluation module is used for obtaining a weld defect confidence evaluation result; and the welding seam quality classification module generates a welding seam quality grade classification result according to the welding seam defect detection result and the welding seam defect confidence degree evaluation result. According to the method, the distinction degree between defect features and background noise is remarkably improved, and the false alarm probability in weld defect detection is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of wind power steel pipe pile technology, and in particular to a machine vision-based weld quality inspection system for wind power steel pipe piles. Background Technology

[0002] With the increasing size of wind power equipment and the rapid development of offshore wind power projects, wind turbine steel pipe piles have become a key load-bearing structure in the foundation construction of wind farms. The quality of their welds directly affects the long-term safety and stable operation of wind turbines. Currently, the quality inspection of wind turbine steel pipe pile welds mainly relies on non-destructive testing methods such as industrial vision, X-ray imaging, and ultrasonic phased array to identify and locate various types of defects on the weld surface and inside.

[0003] In existing technologies, weld defect segmentation methods based on convolutional neural networks have been applied in industrial scenarios. However, they generally employ deep downsampling structures to enhance semantic representation capabilities, resulting in significant loss of spatial features of minute weld defects after multi-layer downsampling. Furthermore, traditional upsampling methods struggle to effectively recover the clear boundaries and fine-grained structures of weld defects when reconstructing high-resolution feature maps, leading to blurred edges and distorted contours in the segmentation results. This, in turn, affects the accuracy of defect size measurement and quality grade assessment.

[0004] On the other hand, the surface of the weld seam of wind power steel pipe piles has a complex industrial texture background, such as fish scale pattern, color difference in heat-affected zone, rust spots and welding spatter. The background interference signal is visually highly similar to the defect features, which makes convolutional neural networks prone to false detection and false detection. Existing attention mechanisms or dilated convolutions are often applied globally with fixed parameters, lacking adaptive design for weld structure geometry and width. They cannot flexibly adjust the receptive field and feature focusing ability according to different spatial locations and defect features, making it difficult to balance the detection of small defects and the suppression of complex backgrounds. Summary of the Invention

[0005] One objective of this invention is to propose a machine vision-based weld quality inspection system for wind power steel pipe piles. This invention significantly improves the distinguishability between defect features and background noise, and effectively reduces the false alarm probability in weld defect detection.

[0006] A machine vision-based weld quality inspection system for wind turbine steel pipe piles according to an embodiment of the present invention includes:

[0007] The weld image preprocessing module acquires the original image data of the weld area of ​​the wind power steel pipe pile to form a weld image data set, and performs preprocessing to generate a preprocessed weld image data set.

[0008] The weld multi-layer semantic feature encoding module inputs the preprocessed weld image data set into the convolutional neural network encoder, extracts multi-layer semantic features and outputs a deep weld feature map.

[0009] An improved FeatureUp feature reconstruction module is used to perform high-fidelity feature reconstruction on the deep weld feature map based on the improved FeatureUp algorithm, generating a high-resolution weld feature map.

[0010] The void attention fusion module performs adaptive void attention fusion processing on the high-resolution weld feature map and outputs a context-guided weld feature map.

[0011] The weld defect segmentation module takes the context-guided weld feature map as input to the decoder, generates a pixel-level weld defect segmentation mask, and performs small region removal, connected component analysis and edge refinement on the pixel-level weld defect segmentation mask to generate weld defect detection results.

[0012] The weld defect confidence assessment module calculates the weld defect confidence index based on the weld defect detection results, and obtains the weld defect confidence assessment results.

[0013] The weld quality classification module generates a weld quality level classification result based on the weld defect detection result and the weld defect confidence assessment result when the weld defect confidence assessment result is higher than the preset confidence threshold.

[0014] Optionally, the preprocessing includes image denoising, illumination equalization, reflection suppression, and texture enhancement.

[0015] Optionally, the weld multi-layer semantic feature encoding module includes:

[0016] In the convolutional neural network encoder structure, an encoding process consisting of multiple consecutive convolutional blocks is constructed, and the preprocessed weld seam image data set is sequentially subjected to layer-by-layer feature mapping operations, and the intermediate feature map set of the weld seam is output layer by layer.

[0017] During the calculation of the intermediate feature map of each layer of weld, the current input intermediate feature map of weld is the output intermediate feature map of weld of the previous layer;

[0018] The output of the last convolutional coding layer of the convolutional neural network encoder structure is defined as the intermediate feature map of the weld seam, which is then defined as the deep weld seam feature map.

[0019] Optionally, the improved FeatureUp feature reconstruction module includes:

[0020] A geometrical guidance field for welds is constructed based on preprocessed weld images, and the geometrical guidance field for welds and the feature map of deep welds are aligned to the same target space through spatial coordinate mapping.

[0021] For each target pixel coordinate in the target space used to generate a high-resolution weld feature map, an inverse mapping relationship from the target pixel coordinate to the deep weld feature map coordinate is established, and a neighborhood sampling set is determined with the deep weld feature map coordinate as the center. The number of neighborhood sampling points in the neighborhood sampling set is determined by the sensitivity of the target pixel coordinate to minute defects.

[0022] Based on the weld geometry guidance field, an anisotropic spatial metric matrix is ​​constructed at the coordinates of each target pixel, and different spatial bandwidths are assigned to the neighborhood sampling points along the weld direction and perpendicular to the weld direction.

[0023] Based on the improved FeatUp algorithm, at each target pixel coordinate of the high-resolution weld feature map, the normalized reconstruction weight of each sampling point in the neighborhood sampling set is calculated by combining the anisotropic spatial metric matrix, the deep weld feature map and the sensitivity to minute defects.

[0024] At each target pixel coordinate, the deep weld features of all sampling points in the neighborhood sampling set and the position encoding information of the target pixel coordinate are input into the learnable high-dimensional feature reconstruction mapping. The high-dimensional feature reconstruction mapping results of all sampling points are weighted and aggregated using normalized reconstruction weights to obtain the high-resolution weld feature vector at the target pixel coordinate.

[0025] The high-resolution weld feature map is generated by taking the high-resolution weld feature vectors of all target pixel coordinates in the target space according to the target space size.

[0026] Optionally, the hole attention fusion module includes:

[0027] Based on the weld geometry guidance field, the corresponding weld width estimate is calculated for each target pixel coordinate in the target space, and the weld width estimates of all target pixel coordinates are combined to form a weld width estimation map.

[0028] Based on the weld width estimation map, an expansion rate set is adaptively selected for the coordinates of each target pixel in the target space;

[0029] Based on the dilation rate set, a multi-scale dilated convolution operation is performed on the high-resolution weld feature map to obtain a multi-scale context feature map corresponding to the dilation rate.

[0030] In the target space, the feature vector of the high-resolution weld seam feature map at the corresponding coordinates is concatenated with the context feature vector of the multi-scale context feature map at the corresponding coordinates through the spatial attention mapping function, and then mapped. The spatial attention weights are obtained by using the Sigmoid function, and the spatial attention weights of all target pixel coordinates are combined to form the spatial attention weight vector.

[0031] Channel attention weights are generated based on high-resolution weld feature maps, and channel attention weight vectors are obtained.

[0032] In the target space, the weld width estimate at the target pixel coordinates is mapped together with the multi-scale context feature map at the expansion rate scale through the cross-scale attention mapping function to obtain the cross-scale attention response at the corresponding expansion rate scale. The cross-scale attention response at all expansion rate scales is normalized by an exponential function to obtain the cross-scale attention weights at all expansion rate scales at the target pixel coordinates.

[0033] Based on multi-scale context feature maps, spatial attention weight vectors, channel attention weight vectors, and cross-scale attention weights, feature weighted fusion is performed on high-resolution weld feature maps to obtain context-guided weld feature maps.

[0034] Optionally, the weld multi-type defect segmentation module includes:

[0035] In the decoder structure, the context-guided weld feature map is upsampled and fused layer by layer to generate the decoded weld feature map. The decoded weld feature map is then classified pixel by pixel using a classification mapping function to generate a pixel-level category logit tensor.

[0036] Perform Softmax normalization on the pixel-level class logit tensor at each target pixel coordinate to obtain the pixel-level class probability tensor;

[0037] Pixel-level weld defect segmentation masks are generated based on pixel-level category probability tensors, and defect category labels for each target pixel coordinate are generated by maximum probability decision.

[0038] Small region removal processing is performed on the pixel-level weld defect segmentation mask, and the remaining segmentation mask only includes the weld defect candidate regions that satisfy the minimum area constraint;

[0039] Perform connected component analysis on the preserved segmentation mask, and output the area, minimum bounding rectangle, and centroid coordinates of the defective connected component for each connected component.

[0040] After removing small areas, the weld defect segmentation mask is subjected to edge refinement processing to obtain an edge-refined weld defect segmentation mask.

[0041] Weld defect detection results are constructed by using defect category labeling, the area of ​​defect connected regions, the minimum bounding rectangle, centroid coordinates, and edge-refined weld defect segmentation mask.

[0042] Optionally, the weld defect confidence assessment module includes:

[0043] Context-guided weld feature maps are extracted based on weld defect segmentation masks, extracting all connected components for each weld defect category.

[0044] For each connected component in the context-guided weld feature map, the region confidence of the connected component is calculated.

[0045] For each connected component in the context-guided weld feature map, the boundary confidence of the connected component is calculated.

[0046] For each connected component in the context-guided weld feature map, the competition suppression confidence of the connected component is calculated.

[0047] The context-guided weld feature map performs a weighted summation of the region confidence, boundary confidence, and competition suppression confidence of the connected domains to obtain a comprehensive confidence index for the connected domains. All comprehensive confidence indices of the connected domains are then combined to form the weld defect confidence assessment result.

[0048] Optionally, the weld quality classification module includes:

[0049] Weld quality assessment standards were extracted from the historical database of weld quality inspection and the current quality specifications for the manufacturing and operation and maintenance of wind power steel pipe piles, and weld quality grading rules were constructed.

[0050] Based on the weld defect detection results and weld defect confidence assessment results, the number of effective defect connected regions, the area ratio of effective defect connected regions, and the distribution density of effective defect connected regions in the axial and circumferential directions of the weld are statistically analyzed for each weld defect category to form a set of weld defect assessment parameters.

[0051] Based on the set of weld defect assessment parameters and the weld quality grading rules, the weld quality status is determined, and the weld quality grade classification results are obtained.

[0052] Optionally, the weld quality grade classification results include:

[0053] Qualified weld: The number of effective connected domains of crack defects is zero, and the number of effective connected domains of porosity defects, slag inclusion defects and non-fusion defects does not exceed the first number threshold, and the area ratio of effective defect connected domains does not exceed the first area threshold, and the distribution density of effective defect connected domains in the axial direction and circumferential direction of the weld is lower than the first distribution density threshold.

[0054] Welds requiring re-inspection: There are no effective connected domains of through-crack defects, and the number of effective connected domains of at least one weld defect category exceeds the first quantity threshold but does not exceed the second quantity threshold, or the area ratio of at least one weld defect category exceeds the first area threshold but does not exceed the second area threshold, or the distribution density of effective defect connected domains in the axial or circumferential direction of the weld exceeds the first distribution density threshold but does not exceed the second distribution density threshold.

[0055] Unqualified weld: There is an effective connected domain of crack defects, or the number of effective connected domains of any weld defect category exceeds the second quantity threshold, or the area ratio of any weld defect category exceeds the second area threshold, or the distribution density of effective defect connected domains in the axial or circumferential direction of the weld exceeds the second distribution density threshold.

[0056] The beneficial effects of this invention are:

[0057] In the deep semantic feature reconstruction stage, this invention introduces geometric guidance information highly correlated with the weld structure of wind power steel pipe piles to specifically improve the FeatUp reconstruction process. During high-resolution feature reconstruction, it combines weld main direction information, weld structure consistency information, and sensitivity information to minor defects to adaptively adjust the neighborhood sampling density, sampling direction, and feature fusion weights. This enables the feature reconstruction process to preferentially propagate defect-related semantics along the weld direction while suppressing high-frequency interference introduced by fish-scale patterns and heat-affected zone textures in the direction perpendicular to the weld. The improved FeatUp algorithm can effectively restore the spatial structural continuity of low signal-to-noise ratio defects while maintaining deep semantic consistency, allowing the deep network to still have a stable response capability to minor defects in high-resolution space.

[0058] In the feature fusion and context modeling stages, this invention introduces an adaptive hole attention-guided segmentation mechanism based on weld width estimation. This mechanism decouples the receptive field of the hole convolution from a fixed range. Instead, it dynamically adjusts the expansion rate set according to the actual width of the weld at different spatial locations. A larger receptive field is introduced in wider weld areas to capture the overall structural context, while a smaller receptive field is used in narrower weld areas or local detail areas to avoid interference from irrelevant background information. This achieves a dynamic balance between context awareness and local fine segmentation capabilities, significantly improving the discriminability between defect features and background noise, and effectively reducing the false alarm probability in weld defect detection.

[0059] This invention combines a comprehensive confidence index with parameters such as the number of defects, the proportion of defect area, and the spatial distribution density of defects to form a weld quality grading and judgment rule. This enables weld inspection results to directly support engineering-level quality judgment and realizes a complete closed loop from pixel-level algorithm output to weld quality assessment report. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of a machine vision-based weld quality inspection system for wind power steel pipe piles proposed in this invention.

[0062] Figure 2 This is a schematic diagram of the structure of a machine vision-based wind power steel pipe pile weld quality inspection system that reconstructs high-resolution weld feature maps based on the FeatUp algorithm. Detailed Implementation

[0063] Example 1: Reference Figures 1-2 A machine vision-based weld quality inspection system for wind turbine steel pipe piles includes:

[0064] The weld image preprocessing module acquires the original image data of the weld area of ​​the wind power steel pipe pile to form a weld image data set, and performs preprocessing to generate a preprocessed weld image data set.

[0065] In this embodiment, the preprocessing includes image denoising, illumination equalization, reflection suppression, and texture enhancement.

[0066] The preprocessed weld seam image dataset is represented as a set of weld seam images that have undergone denoising, illumination equalization, reflection suppression, and texture enhancement. The set of weld seam images includes several weld seam images, each with a uniform number of pixels in height, width, and channels. The preprocessed weld seam image dataset is organized according to three dimensions: height, width, and number of channels, and represented as a three-dimensional tensor. The initial height of the weld seam image is the height dimension, the initial width of the weld seam image is the width dimension, and the number of channels of the weld seam image is the channel dimension.

[0067] The weld multi-layer semantic feature encoding module inputs the preprocessed weld image data set into the convolutional neural network encoder, extracts multi-layer semantic features and outputs a deep weld feature map.

[0068] In this embodiment, the weld multi-layer semantic feature encoding module includes:

[0069] In the convolutional neural network encoder structure, an encoding process consisting of multiple consecutive convolutional blocks is constructed, and the preprocessed weld seam image data set is sequentially subjected to layer-by-layer feature mapping operations, and the intermediate feature map set of the weld seam is output layer by layer.

[0070] Each convolutional block contains convolution operations, nonlinear activation function operations, and normalization operations in sequence. The intermediate feature map of the weld in each layer is defined by both spatial size and channel dimension. The spatial size decreases layer by layer with the encoder depth, while the channel dimension increases layer by layer with the encoder depth, which is used to enhance semantic abstraction capabilities.

[0071] During the calculation of the intermediate feature map of each layer of weld, the current input intermediate feature map of weld is the output intermediate feature map of weld of the previous layer;

[0072] The current output weld intermediate feature map is obtained by convolving the current convolution kernel weights with the input weld intermediate feature map, adding the bias term of the current convolution block, performing a normalization operation, and then processing it through a nonlinear activation function.

[0073] The output of the last convolutional coding layer of the convolutional neural network encoder structure is defined as the intermediate feature map of the weld seam, which is then defined as the deep weld seam feature map.

[0074] The deep weld feature map is obtained after processing all convolutional blocks within the convolutional neural network encoder structure. The spatial size is significantly reduced compared to the initial height and width, and the number of channels is significantly increased compared to the initial number of channels. The deep weld feature map centrally represents the multi-scale, complex semantic information and defect features in the weld image.

[0075] The FeatureUp feature reconstruction module is improved, and high-fidelity feature reconstruction of deep weld feature maps is performed based on the improved FeatureUp algorithm to generate high-resolution weld feature maps.

[0076] In this embodiment, the FeatureUp feature reconstruction module is improved, including:

[0077] A geometrical guidance field for welds is constructed based on preprocessed weld images, and the geometrical guidance field for welds and the feature map of deep welds are aligned to the same target space through spatial coordinate mapping.

[0078] In Example 1, the main direction information and structural consistency information of the weld are calculated within the pixel space of the preprocessed weld image. The weld geometric guidance field is constructed jointly by the main direction information and the structural consistency information. This field is then mapped to coordinates according to the spatial resolution ratio between the preprocessed weld image and the deep weld feature map, ensuring that the weld geometric guidance field is spatially aligned with the deep weld feature map. Figure 1 One-to-one correspondence aligns the weld geometry guidance field with the deep weld feature map to the same target space, which is used to constrain the feature propagation direction along the weld direction and the feature suppression direction perpendicular to the weld direction during high-resolution feature reconstruction.

[0079] The main direction information of the weld is obtained by performing local gradient direction statistics on the preprocessed weld image and performing direction aggregation within the weld area. The weld structure consistency information is obtained by measuring the direction consistency and grayscale change continuity of the weld fish scale pattern and heat-affected zone texture in the local neighborhood.

[0080] For each target pixel coordinate in the target space used to generate a high-resolution weld feature map, an inverse mapping relationship from the target pixel coordinate to the deep weld feature map coordinate is established, and a neighborhood sampling set is determined with the deep weld feature map coordinate as the center. The number of neighborhood sampling points in the neighborhood sampling set is determined by the sensitivity of the target pixel coordinate to minute defects.

[0081] In Example 1, the reverse mapping relationship maps the target pixel coordinates to the corresponding deep weld feature map coordinates according to the ratio of the target height of the high-resolution weld feature map to the feature height of the deep weld feature map, and the ratio of the target width of the high-resolution weld feature map to the feature width of the deep weld feature map. Centered on the deep weld feature map coordinates, a neighborhood sampling set is determined on the deep weld feature map. The neighborhood sampling set consists of multiple deep weld feature map coordinates within a preset sampling range surrounding the deep weld feature map coordinates, and is used to provide candidate feature support for high-resolution reconstruction of the target pixel coordinates.

[0082] The sensitivity to minute defects is obtained by applying a learnable defect sensitivity mapping to the feature vector at the coordinates of the deep weld feature map. The defect sensitivity includes the sensitivity to the type of minute defects in the weld of the wind power steel pipe pile and the sensitivity to the type of large-scale defects in the weld of the wind power steel pipe pile. The number of neighborhood sampling points is equal to the sum of the number of basic neighborhood sampling points and the weighted term of the minute defect sensitivity. For areas where the defect sensitivity is higher than the threshold, denser sampling is used to improve the microstructure reconstruction density.

[0083] Based on the weld geometry guidance field, an anisotropic spatial metric matrix is ​​constructed at the coordinates of each target pixel, and different spatial bandwidths are assigned to the neighborhood sampling points along the weld direction and perpendicular to the weld direction.

[0084] In Example 1, the process of constructing the anisotropic space metric matrix includes:

[0085] At the target pixel coordinates, the unit vector of the main direction of the weld is obtained based on the main direction field of the weld. The unit vector of the main direction of the weld is used as the first coordinate axis direction, and the direction perpendicular to the unit vector of the main direction of the weld is used as the second coordinate axis direction to establish a local orientation coordinate system at the target pixel coordinates. The weld structure consistency strength value at the target pixel coordinates is obtained based on the weld structure consistency map, and the weld structure consistency strength value is mapped to a spatial bandwidth parameter along the main direction of the weld and a spatial bandwidth parameter perpendicular to the main direction of the weld. The larger the weld structure consistency strength value, the larger the value of the spatial bandwidth parameter along the main direction of the weld, and the smaller the value of the spatial bandwidth parameter perpendicular to the main direction of the weld. The local orientation coordinate system and the corresponding spatial bandwidth parameter are combined to generate an anisotropic spatial metric matrix for measuring the spatial distribution relationship of the neighborhood sampling points. This allows the anisotropic spatial metric matrix to allow longer feature propagation distances along the main direction of the weld and restricts the feature propagation range perpendicular to the main direction of the weld, so as to match the spatial distribution characteristics of crack defects extending along the weld and locally concentrated porosity defects in the weld of wind power steel pipe piles.

[0086] ;

[0087] Where p represents the target pixel coordinates in the target space of the high-resolution weld feature map, and the target pixel coordinates correspond to a specific spatial location in the weld area of ​​the wind power steel pipe pile. This represents the anisotropic spatial metric matrix constructed at the target pixel coordinates. This represents the unit vector of the principal direction of the weld at the target pixel coordinates, used to indicate the principal extension direction of the weld at that location in the wind turbine steel pipe pile. This represents a two-dimensional rotation matrix constructed from the unit vectors of the principal direction of the weld. Represents the transpose of a two-dimensional rotation matrix. This represents the spatial bandwidth parameter along the main direction of the weld at the target pixel coordinates, used to control the propagation range of features along the main direction of the weld. This parameter represents the spatial bandwidth at the target pixel coordinates, perpendicular to the main direction of the weld seam. It is used to limit the propagation range of features in the direction perpendicular to the weld seam. and These represent the reciprocals of the squares of the spatial bandwidth parameters along the main weld direction and perpendicular to the main weld direction, respectively. The middle diagonal matrix represents the directional weight matrix with different spatial metric intensities applied in the main weld direction and perpendicular to the main weld direction. The directional weight matrix and the rotation matrix work together to enable the anisotropic spatial metric matrix to allow feature fusion over a longer distance in the main weld direction, while restricting the participation of irrelevant texture features in the perpendicular to the main weld direction.

[0088] Based on the improved FeatUp algorithm, at each target pixel coordinate of the high-resolution weld feature map, the normalized reconstruction weight of each sampling point in the neighborhood sampling set is calculated by combining the anisotropic spatial metric matrix, the deep weld feature map and the sensitivity to minute defects.

[0089] ;

[0090] Where q represents the inverse mapping coordinates corresponding to the target pixel coordinates in the deep weld feature map, which is the mapping point from high-resolution space to low-resolution space. This represents the coordinates of the i-th neighboring sampling point selected in the deep weld feature map, centered on the inverse mapping coordinates. This represents the feature vector of the deep weld feature map at the inverse mapping coordinates. This represents the feature vector of the deep weld feature map at the coordinates of the neighboring sampling points. The bandwidth parameter representing the similarity of deep features. Represents the L2 norm, This represents the sensitivity enhancement coefficient for minute defects. This represents a sensitivity scalar for minute defect types at the coordinates of a neighborhood sampling point in the deep weld feature map. This represents the normalized reconstruction weight of the i-th neighborhood sampling point for the target pixel coordinates of the high-resolution weld feature map. This represents the number of sampling points in the neighborhood sampling set.

[0091] The formula, through a normalized reconstruction weight setting method, can effectively avoid interference from weld scale patterns, heat-affected zone textures, and random noise on defect features during high-resolution feature reconstruction. This ensures that crack-like defects continuously distributed along the weld direction maintain good structural coherence after feature reconstruction, while preventing the diffusion of irrelevant textures perpendicular to the weld direction. The weight enhancement mechanism for micro-defect areas significantly improves the response intensity of low signal-to-noise ratio defects such as micro-cracks and micro-porosity, reducing the risk of fine defects being smoothed or submerged during deep feature reconstruction. This improves the detection rate, boundary positioning accuracy, and reliability of multi-type defect detection in wind power steel pipe pile welds.

[0092] At each target pixel coordinate, the deep weld features of all sampling points in the neighborhood sampling set and the position encoding information of the target pixel coordinate are input into the learnable high-dimensional feature reconstruction mapping. The high-dimensional feature reconstruction mapping results of all sampling points are weighted and aggregated using normalized reconstruction weights to obtain the high-resolution weld feature vector at the target pixel coordinate.

[0093] In Example 1, a high-dimensional feature reconstruction map is constructed at each target pixel coordinate. The high-dimensional feature reconstruction map is composed of multiple cascaded feature transformation units, which are used to jointly map the deep weld features corresponding to each sampling point in the neighborhood sampling set with the position encoding information of the target pixel coordinates. The deep weld features serve as semantic input to represent the multi-scale semantic information related to defects in the weld region. The position encoding information of the target pixel coordinates serves as spatial input to represent the relative spatial position of the target pixel in the weld structure in the high-resolution weld feature map. During the high-dimensional feature reconstruction process, the high-dimensional feature reconstruction map adaptively adjusts its feature transformation weights for different target pixel coordinates, so that crack defects continuously distributed along the main direction of the weld maintain directional continuity in the reconstructed features, and porosity defects with local clusters maintain spatial focus in the reconstructed features. After obtaining the high-dimensional feature reconstruction results corresponding to each neighborhood sampling point, the high-dimensional feature reconstruction results of all sampling points are weighted and fused according to the normalized reconstruction weights to generate a high-resolution weld feature vector at the target pixel coordinates.

[0094] The high-resolution weld feature map is generated by taking the high-resolution weld feature vectors of all target pixel coordinates in the target space according to the target space size.

[0095] In Example 1, based on the two-dimensional spatial position of the target pixel coordinates in the target space, the corresponding high-resolution weld feature vector is written into the feature storage location corresponding to the target pixel coordinates. The above writing operation is repeated for all target pixel coordinates in the target space, so that the high-resolution weld feature vectors corresponding to all target pixel coordinates are arranged in a regular manner in the target space according to the spatial arrangement order of the target height and the target width. By stacking the high-resolution weld feature vectors of all target pixel coordinates in a consistent spatial position and splicing them in the channel dimension, a high-resolution weld feature map continuously defined within the target height and target width spatial range is formed.

[0096] The spatial dimensions of a high-resolution weld feature map are equal to the target height and target width of the high-resolution weld feature map. The number of feature channels in a high-resolution weld feature map is equal to the feature dimension of a single high-resolution weld feature vector. The spatial dimensions of a high-resolution weld feature map are defined by the target height and target width. The number of feature channels is the target number of channels. The coordinates of each target pixel in the weld geometry guidance field include the weld geometry guidance information at the coordinates.

[0097] The void attention fusion module performs adaptive void attention fusion processing on the high-resolution weld feature map and outputs a context-guided weld feature map.

[0098] In this embodiment, the hollow attention fusion module includes:

[0099] Based on the weld geometry guidance field, the corresponding weld width estimate is calculated for each target pixel coordinate in the target space, and the weld width estimates of all target pixel coordinates are combined to form a weld width estimation map.

[0100] Extract the unit vector of the main direction of the weld and the weld structure consistency strength value at the target pixel coordinates. Search for positions where the weld structure consistency strength value is lower than a preset threshold on both sides of the main direction of the weld. With the target pixel coordinates as the center and the searched positions on both sides as the endpoints, calculate the Euclidean distance between the endpoints as the weld width estimate at the target pixel coordinates. Combine and arrange all the weld width estimates at the target pixel coordinates in the target space to form a weld width estimation map, which is used to represent the actual width variation of the wind power steel pipe pile weld at each spatial location.

[0101] Based on the weld width estimation map, an expansion rate set is adaptively selected for the coordinates of each target pixel in the target space;

[0102] The expansion rate set includes several expansion rates. Each expansion rate is the product of the estimated weld width and the expansion rate width ratio coefficient. After rounding down, it is limited to the minimum and maximum expansion rates. All expansion rates form the expansion rate set corresponding to the target pixel coordinates. The expansion rate width ratio coefficient is set manually.

[0103] Based on the dilation rate set, a multi-scale dilated convolution operation is performed on the high-resolution weld feature map to obtain a multi-scale context feature map corresponding to the dilation rate.

[0104] In Example 1, using the expansion rate as the sampling interval, several hole sampling positions are determined around the target pixel coordinates in the high-resolution weld feature map according to a preset hole convolution sampling offset mode. At each hole sampling position, the corresponding feature vector is read from the high-resolution weld feature map, and the feature vector is weighted element-wise with the convolution kernel weight coefficient corresponding to the current expansion rate. The weighted results of all hole sampling positions under the same expansion rate are accumulated to obtain the context feature vector corresponding to the expansion rate. The above process is repeated for all target pixel coordinates in the target space, so that each target pixel coordinate obtains a context feature vector under the corresponding expansion rate. The context feature vectors corresponding to all target pixel coordinates are arranged in the spatial order of the target space to form a multi-scale context feature map corresponding to the expansion rate.

[0105] In the target space, the feature vector of the high-resolution weld seam feature map at the corresponding coordinates is concatenated with the context feature vector of the multi-scale context feature map at the corresponding coordinates through the spatial attention mapping function, and then mapped. The spatial attention weights are obtained by using the Sigmoid function, and the spatial attention weights of all target pixel coordinates are combined to form the spatial attention weight vector.

[0106] The spatial attention weight value is between zero and one, representing the spatial attention intensity of the weld defect area at the target pixel coordinates.

[0107] Channel attention weights are generated based on high-resolution weld feature maps, and channel attention weight vectors are obtained.

[0108] In Example 1, in the high-resolution weld feature map, for each feature channel, spatial aggregation operation is performed on the feature values ​​of the feature channel at all target pixel coordinates in the target space to obtain the channel description value corresponding to the feature channel. The spatial aggregation operation is obtained by statistically aggregating all feature values ​​of the feature channel within the target height and target width range, which is used to represent the overall response intensity of the feature channel to the defect semantics in the entire wind power steel pipe pile weld area. The channel description values ​​corresponding to all feature channels are combined into a channel description vector according to the feature channel order, and the channel description vector is input into the channel attention mapping function for nonlinear mapping processing to generate a channel response value corresponding to each feature channel. Normalization activation processing is performed on the channel response value to obtain the channel attention weight corresponding to each feature channel. The channel attention weights of all feature channels are combined according to the feature channel order to form a channel attention weight vector.

[0109] In the target space, the weld width estimate at the target pixel coordinates is mapped together with the multi-scale context feature map at the expansion rate scale through the cross-scale attention mapping function to obtain the cross-scale attention response at the corresponding expansion rate scale. The cross-scale attention response at all expansion rate scales is normalized by an exponential function to obtain the cross-scale attention weights at all expansion rate scales at the target pixel coordinates.

[0110] Each cross-scale attention weight value is between zero and one, and the sum of the cross-scale attention weights across all expansion scales is one.

[0111] ;

[0112] in, This represents the target pixel coordinates in the target space of the high-resolution weld feature map. Place, No. Cross-scale attention weights corresponding to each dilated convolutional dilation rate scale. The index number represents the dilation rate scale and is used to distinguish the context scale corresponding to different dilation rates in dilated convolutions. Indicates the target pixel coordinates The total number of selected dilatation scales. This represents a cross-scale attention mapping function, used to jointly map the estimated weld width at the target pixel coordinates with the multi-scale context feature map at the corresponding expansion rate scale, generating a scale response value that reflects the degree of matching between that scale and the current weld structure. This represents the target pixel coordinates in the target space of a high-resolution weld feature map. Estimated weld width at the location, Indicates the target pixel coordinates At this point, the first method is adopted. Each hole convolution dilation rate Multi-scale contextual feature maps obtained by performing dilated convolution on high-resolution weld feature maps. Indicates the target pixel coordinates At this point, the first method is adopted. Each hole convolution dilation rate Multi-scale contextual feature maps obtained by performing dilated convolution on high-resolution weld feature maps. This represents an exponential function.

[0113] Based on multi-scale context feature maps, spatial attention weight vectors, channel attention weight vectors, and cross-scale attention weights, feature weighted fusion is performed on high-resolution weld feature maps to obtain context-guided weld feature maps.

[0114] ;

[0115] in, Represents the spatial attention weight vector. This represents the channel attention weight vector.

[0116] The weld defect segmentation module takes the context-guided weld feature map as input to the decoder, generates a pixel-level weld defect segmentation mask, and performs small region removal, connected component analysis and edge refinement on the pixel-level weld defect segmentation mask to generate weld defect detection results.

[0117] In this embodiment, the weld multi-type defect segmentation module includes:

[0118] In the decoder structure, the context-guided weld feature map is upsampled and fused layer by layer to generate the decoded weld feature map. The decoded weld feature map is then classified pixel by pixel using a classification mapping function to generate a pixel-level category logit tensor.

[0119] In the decoder structure, upsampling operations are performed on the current context-guided weld feature map sequentially according to the preset decoding level order. The upsampled context-guided weld feature map is then fused with the decoded features at the same spatial scale to restore the spatial resolution of the feature map layer by layer, generating a decoded weld feature map. When the spatial size of the decoded weld feature map is restored to be consistent with the target space, the decoded weld feature map is input into the classification mapping function. Pixel-wise category mapping processing is performed on the feature vector corresponding to each target pixel coordinate in the decoded weld feature map to generate a pixel-level category logit tensor. At each target pixel coordinate in the target space, the pixel-level category logit tensor contains a category response value that corresponds one-to-one with the total number of weld defect categories, which is used to represent the classification response intensity of the target pixel coordinate on each weld defect category.

[0120] Perform Softmax normalization on the pixel-level class logit tensor at each target pixel coordinate to obtain the pixel-level class probability tensor;

[0121] At each target pixel coordinate, the sum of the probability values ​​of all categories in the pixel-level category probability tensor is one. Each probability value of the pixel-level category probability tensor represents the pixel-level probability that the target pixel coordinate belongs to the weld defect category. The pixel-level category probability tensor is completely consistent with the target height and target width of the high-resolution weld feature map in terms of spatial size. The number of feature channels is equal to the total number of weld defect categories. The total number of weld defect categories includes background category, crack defect category, porosity defect category, slag inclusion defect category and non-fusion defect category.

[0122] Pixel-level weld defect segmentation masks are generated based on pixel-level category probability tensors, and defect category labels for each target pixel coordinate are generated by maximum probability decision.

[0123] In Example 1, within the target space, the pixel-level probability values ​​of each weld defect category corresponding to the target pixel coordinates in the pixel-level category probability tensor are read. The pixel-level probability values ​​of each weld defect category corresponding to the target pixel coordinates are compared one by one, and the weld defect category with the highest probability value is selected as the defect category label of the target pixel coordinate. The defect category label of the target pixel coordinate is written into the position corresponding to the target pixel coordinate in the form of a category index to obtain the pixel-level weld defect segmentation mask.

[0124] Small region removal processing is performed on the pixel-level weld defect segmentation mask, and the remaining segmentation mask only includes the weld defect candidate regions that satisfy the minimum area constraint;

[0125] The small region removal process specifically involves calculating all connected components under each weld defect category, counting the number of pixels in each connected component, and setting the corresponding connected component as the background category in the segmentation mask if the number of pixels in the connected component is less than the minimum area threshold.

[0126] Perform connected component analysis on the preserved segmentation mask, and output the area, minimum bounding rectangle, and centroid coordinates of the defective connected component for each connected component.

[0127] The connected component analysis process involves extracting all connected components for each weld defect category, and outputting the area, minimum bounding rectangle, and centroid coordinates of each connected component. The area of ​​the connected component is the number of pixels within it, the minimum bounding rectangle is the smallest rectangle that can cover the connected component, and the centroid coordinates are the average coordinates of the spatial positions of all pixels in the connected component.

[0128] After removing small areas, the weld defect segmentation mask is subjected to edge refinement processing to obtain an edge-refined weld defect segmentation mask.

[0129] During the edge refinement process, for all connected domains of each weld defect category, the boundaries of the connected domains are refined and their shapes are optimized to obtain weld defect segmentation results with clear boundaries.

[0130] Weld defect detection results are constructed by using defect category labeling, the area of ​​defect connected regions, the minimum bounding rectangle, centroid coordinates, and edge-refined weld defect segmentation mask.

[0131] The weld defect confidence assessment module calculates the weld defect confidence index based on the weld defect detection results, and obtains the weld defect confidence assessment results.

[0132] In this embodiment, the weld defect confidence assessment module includes:

[0133] Context-guided weld feature maps are extracted based on weld defect segmentation masks, extracting all connected components for each weld defect category.

[0134] The set of connected components for each weld defect category is defined as the set of connected components for the corresponding category. Each connected component represents a spatially connected defect region under the same category. Each connected component includes the set of all target pixel coordinates belonging to that connected component.

[0135] For each connected component in the context-guided weld feature map, the region confidence of the connected component is calculated.

[0136] The region confidence score is obtained by statistically aggregating the pixel-level probability values ​​of all target pixel coordinates in the weld defect category probability tensor corresponding to the target category.

[0137] For each connected component in the context-guided weld feature map, the boundary confidence of the connected component is calculated.

[0138] Boundary confidence is obtained by statistically aggregating the pixel-level probability values ​​of the boundary pixels of the connected region corresponding to the target category in the weld defect category probability tensor.

[0139] For each connected component in the context-guided weld feature map, the competition suppression confidence of the connected component is calculated.

[0140] Competition suppression confidence indicates the degree of competition between target pixel coordinates and non-target defect categories within a connected component. Specifically, for all target pixel coordinates within the connected component, the maximum pixel-level probability value is selected in all non-target defect categories. Then, the sum of all maximum probability values ​​is divided by the number of target pixel coordinates within the connected component to obtain the average maximum probability value of the connected component across all non-target defect categories. Subtracting this average maximum probability value from 1 yields the competition suppression confidence of the connected component. Competition suppression confidence measures the class discrimination ability of the connected component in the target defect category. The higher the competition suppression confidence, the stronger the class discrimination ability of the connected component in the target defect category.

[0141] The context-guided weld feature map performs a weighted summation of the region confidence, boundary confidence, and competition suppression confidence of the connected domains to obtain a comprehensive confidence index for the connected domains. All comprehensive confidence indices of the connected domains are then combined to form the weld defect confidence assessment result.

[0142] The sum of the weighting coefficients of the regional confidence, the boundary confidence, and the competition inhibition confidence is equal to one. The comprehensive confidence index of each connected component is used to quantitatively represent the detection reliability of the connected component on the target defect category. The comprehensive confidence index of all connected components together constitutes the confidence assessment result of the weld defects of wind power steel pipe piles.

[0143] The weld quality classification module generates a weld quality level classification result based on the weld defect detection result and the weld defect confidence assessment result when the weld defect confidence assessment result is higher than the preset confidence threshold.

[0144] In this embodiment, the weld quality classification module includes:

[0145] Weld quality assessment standards were extracted from the historical database of weld quality inspection and the current quality specifications for the manufacturing and operation and maintenance of wind power steel pipe piles, and weld quality grading rules were constructed.

[0146] Based on the weld defect detection results and weld defect confidence assessment results, the number of effective defect connected regions, the area ratio of effective defect connected regions, and the distribution density of effective defect connected regions in the axial and circumferential directions of the weld are statistically analyzed for each weld defect category to form a set of weld defect assessment parameters.

[0147] In Example 1, the number of effective defect connected components is determined by iterating through all defect connected components under each weld defect category, selecting connected components with a comprehensive confidence index higher than a preset confidence threshold as effective defect connected components, and counting the number of effective defect connected components as the number of effective defect connected components for the weld defect category.

[0148] The effective defect connected region area ratio is calculated by summing the pixel areas of all effective defect connected regions under each weld defect category, and then calculating the ratio of the sum of pixel areas to the total number of pixels of all target pixel coordinates within the weld detection area.

[0149] The distribution density of the weld along the axial and circumferential directions is based on the minimum bounding rectangle and centroid coordinates of each effective defect connected domain in the weld defect detection results. The projection distribution of the effective defect connected domains in the weld length and width directions is statistically analyzed. The weld detection area is divided into several equal-length statistical intervals in the length and width directions, respectively. The number of effective defect connected domain centroids falling into the equal-length statistical intervals in each statistical interval is counted, and the number and the corresponding interval length are normalized to obtain the axial and circumferential distribution densities of the weld, which represent the spatial concentration of weld defects in the overall weld structure.

[0150] Based on the set of weld defect assessment parameters and the weld quality grading rules, the weld quality status is determined, and the weld quality grade classification results are obtained.

[0151] In this embodiment, the weld quality grade classification results include:

[0152] Qualified weld: The number of effective connected domains of crack defects is zero, and the number of effective connected domains of porosity defects, slag inclusion defects and non-fusion defects does not exceed the first number threshold, and the area ratio of effective defect connected domains does not exceed the first area threshold, and the distribution density of effective defect connected domains in the axial direction and circumferential direction of the weld is lower than the first distribution density threshold.

[0153] Welds requiring re-inspection: There are no effective connected domains of through-crack defects, and the number of effective connected domains of at least one weld defect category exceeds the first quantity threshold but does not exceed the second quantity threshold, or the area ratio of at least one weld defect category exceeds the first area threshold but does not exceed the second area threshold, or the distribution density of effective defect connected domains in the axial or circumferential direction of the weld exceeds the first distribution density threshold but does not exceed the second distribution density threshold.

[0154] Unqualified weld: There is an effective connected domain of crack defects, or the number of effective connected domains of any weld defect category exceeds the second quantity threshold, or the area ratio of any weld defect category exceeds the second area threshold, or the distribution density of effective defect connected domains in the axial or circumferential direction of the weld exceeds the second distribution density threshold.

[0155] In Example 1, during the detection and quality assessment of multiple types of defects in the welds of wind power steel pipe piles, there are significant differences in the quantity, spatial distribution characteristics, and impact on structural safety of different types of weld defects. In order to balance the engineering applicability of the detection results with the generalization ability of the algorithm, the first quantity threshold, the second quantity threshold, the first distribution density threshold, and the second distribution density threshold are all set by combining historical data statistics and engineering specifications.

[0156] The first and second quantity thresholds are used to constrain the number of effective defect connected domains within the same weld area. The values ​​are determined based on the statistical analysis of samples of welds judged as qualified, welds requiring re-inspection, and unqualified by manual or traditional non-destructive testing methods in the historical inspection data of wind power steel pipe pile manufacturing and operation and maintenance. The distribution range of the number of effective defect connected domains corresponding to different defect categories in each type of weld sample is calculated. The upper limit of the number of defect connected domains that frequently appear in qualified weld samples is set as the first quantity threshold, and the lower limit of the number of defect connected domains that significantly increase in unqualified weld samples is set as the second quantity threshold. Thus, the first and second quantity thresholds can reflect the critical characteristics of the transition of wind power steel pipe pile welds from an acceptable state to an unacceptable state in engineering practice.

[0157] The first and second distribution density thresholds are used to constrain the spatial concentration of effective defect connected domains in the weld axial and circumferential directions. The values ​​are determined by dividing the weld length or circumferential unfolded length into equal-length statistical intervals in the weld axial and circumferential directions, statistically analyzing the distribution of the centroid coordinates of effective defect connected domains in historical test samples within each statistical interval, and calculating the statistical distribution of the number of centroids of defect connected domains per unit length. The first distribution density threshold is set as the upper limit of the number of centroids per unit length when the centroid distribution of defect connected domains in qualified weld samples is relatively discrete and uniform. The second distribution density threshold is set as the lower limit of the number of centroids per unit length when the centroids of defect connected domains in unqualified weld samples are significantly concentrated and exhibit a clustered distribution, reflecting the engineering characteristics of defects accumulating in local areas of the weld and significantly impacting structural safety.

[0158] When the confidence assessment result of weld defect is lower than the preset confidence threshold and is marked as a false detection, the iterative sample mining process is triggered. Iterative samples are collected and difficult samples are added to the weld image dataset. The iterative samples are used to incrementally train the convolutional neural network encoder, the improved FeatUp algorithm and the adaptive dilated attention fusion processing. The process is then returned to extract multi-layer semantic features and continue execution.

[0159] Example 2: On a wind power foundation production line, an inspection system was deployed in the automated weld inspection process. In a single batch inspection task, the system received a data packet containing 15 original weld images, each with a resolution of 2048×2048, covering three longitudinal weld seams and two circumferential weld seams in the steel pipe piles, with a total inspection length of approximately 13.2 meters. The inspected objects included various common defects found in real-world manufacturing scenarios, such as cracks, porosity, slag inclusions, and lack of fusion.

[0160] The system sequentially performs Gaussian filtering for noise reduction on each original image, normalizes the grayscale range to [0, 1], and uses local adaptive histogram equalization to improve the contrast of the weld area. Taking the third image as an example, the noise variance decreased from the original 0.019 to 0.006, and the signal-to-noise ratio improved by 31%. After the main region of the original image is identified, the mean grayscale gradient of the edges is improved to 0.42.

[0161] The preprocessed image is processed layer by layer by the encoder. Taking the third image as an example, after 5 levels of convolution, the output deep weld feature map size is 64×64×512. At coordinates (29,17), the maximum response value of channel #112 is 0.78, which is initially identified as the feature focal point of the porosity region; while the crack feature response value corresponding to channel #309 at (45,60) is 0.64.

[0162] The system automatically analyzes the preprocessed image and establishes the main direction field of the weld. The average main direction of the third image is 91.5°, the average structural consistency score is 0.82, and the maximum consistency interval width is 31 pixels. Through spatial alignment, the main direction corresponds one-to-one with the feature map coordinates. For row 26 of the third image, the system back-maps the weld main direction range in that row to the deep feature map, setting the number of neighborhood sampling points to 16 (8 basic points + 8 points weighted by the sensitivity of small defects). For the sensitive area of ​​small pores (26,35), the sampling is encrypted to 24 points. Taking the target pixel (26,35) as an example, the spatial bandwidth parameter... =9.2, =3.1, the anisotropic space metric matrix is ​​automatically generated, after the reconstruction weight (26,35) is normalized, the maximum value corresponds to the sampling point (9,13) of the deep feature map, the value is 0.132, after high-dimensional feature reconstruction, the feature vector length of this point is 256, and the consistency index of defect direction distribution is 0.89.

[0163] After all target pixels are written, a high-resolution weld feature map is generated. The feature map has 18% better coherence in the crack propagation direction and 23% lower noise response mean square error.

[0164] The system automatically calculates the weld width at all points within the target space. The mean width estimate is 18.4 mm, and the standard deviation is 2.7 mm. Taking pixel (41,78) as an example, the estimated width is 20.1 mm. The dilation rate set {2,3,5} is adaptively selected, and the dilated convolution branch at this point outputs three sets of multi-scale contextual feature vectors. The spatial attention weight is 0.94, and the channel attention weight reaches 0.88 at the pore channel (#48). The cross-scale attention weights are 0.61 (r=2), 0.27 (r=3), and 0.12 (r=5), respectively. After fusion, the context-guided feature response at point (41,78) is 1.17.

[0165] The system upsamples the context-guided weld feature map layer by layer, fusing channel features after each upsampling to restore it to 2048×2048. The classification mapping function outputs logit[porosity=2.86, crack=1.19, inclusion=0.52, lack of fusion=0.27, background=-1.92] at (41,78). After Softmax normalization, the probability of porosity at this point is 0.67, which is the largest category.

[0166] After probabilistic decision-making for all pixels, a segmentation mask is generated. Five pore connected regions (maximum area 22 pixels), two crack connected regions (maximum length 47 pixels), one slag inclusion connected region (area 48 pixels), and zero unfused connected regions were detected in the image. All connected regions were filtered out using minimum area filtering (threshold 8 pixels), and the edge width was reduced by an average of 1.7 pixels after boundary thinning.

[0167] For the first connected component of the stomatal category, the confidence level is 0.88 for the region, 0.83 for the boundary, and 0.90 for the competition inhibition. The weighting coefficients are 0.5, 0.3, and 0.2, respectively, and the final comprehensive confidence level is 0.868. The confidence levels of all connected components are higher than the system-set threshold of 0.80.

[0168] The system reads the number, area ratio, and spatial density of defect connected regions. Taking the third image as an example, the number of effective connected regions of pores is 5, the area ratio is 0.0094 (< the first area threshold of 0.012), and the pore density is all below the first density threshold of 0.07 / 100 pixels. The system determines that the weld area is qualified. In the 12th image of Example 2, there is one crack connected region with a length of 56 pixels. The system directly outputs a non-qualified judgment.

[0169] To verify the superiority of the method of the present invention, the team constructed a training set containing 600 images of weld defects in wind power steel pipe piles and 200 independent test sets based on real industrial scenarios, and compared them with traditional U-Net, DeepLabV3+, and fixed parameter dilated convolution segmentation methods. The data is shown in Table 1 below.

[0170] Table 1. Comparison of Segmentation Accuracy

[0171] method mIoU(%) Recall(%) Precision (%) F1 score (%) Traditional U-Net 71.8 74.2 78.1 76.1 DeepLabV3+ 77.5 80.6 82.3 81.4 Fixed-hole convolution 80.3 82.7 84.0 83.3 Method of the present invention 87.4 91.2 89.7 90.4

[0172] In areas with micropores and fine cracks, the method of this invention achieves a false negative rate of 2.3%, significantly better than the comparative methods (U-Net false negative rate 12.4%, DeepLabV3+ false negative rate 8.8%, and fixed-difference convolution false negative rate 6.0%). On an Intel Xeon E5 server with an NVIDIA RTX3080, the average time for full-process detection of a single 2048×2048 pixel image is 1.28 seconds, which is significantly faster than traditional methods (U-Net 2.10 seconds, DeepLabV3+ 1.85 seconds, and fixed-difference convolution 1.52 seconds), demonstrating higher engineering usability.

[0173] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine vision-based weld quality inspection system for wind power steel pipe piles, characterized in that, include: The weld image preprocessing module acquires the original image data of the weld area of ​​the wind power steel pipe pile to form a weld image data set, and performs preprocessing to generate a preprocessed weld image data set. The weld multi-layer semantic feature encoding module inputs the preprocessed weld image data set into the convolutional neural network encoder, extracts multi-layer semantic features and outputs a deep weld feature map. An improved FeatureUp feature reconstruction module is used to perform high-fidelity feature reconstruction on the deep weld feature map based on the improved FeatureUp algorithm, generating a high-resolution weld feature map. The void attention fusion module performs adaptive void attention fusion processing on the high-resolution weld feature map and outputs a context-guided weld feature map. The weld defect segmentation module takes the context-guided weld feature map as input to the decoder, generates a pixel-level weld defect segmentation mask, and performs small region removal, connected component analysis and edge refinement on the pixel-level weld defect segmentation mask to generate weld defect detection results. The weld defect confidence assessment module calculates the weld defect confidence index based on the weld defect detection results, and obtains the weld defect confidence assessment results. The weld quality classification module generates a weld quality level classification result based on the weld defect detection result and the weld defect confidence assessment result when the weld defect confidence assessment result is higher than the preset confidence threshold.

2. The machine vision-based weld quality inspection system for wind power steel pipe piles according to claim 1, characterized in that, The preprocessing includes image denoising, illumination equalization, reflection suppression, and texture enhancement.

3. The machine vision-based weld quality inspection system for wind power steel pipe piles according to claim 1, characterized in that, The weld multi-layer semantic feature encoding module includes: In the convolutional neural network encoder structure, an encoding process consisting of multiple consecutive convolutional blocks is constructed, and the preprocessed weld seam image data set is sequentially subjected to layer-by-layer feature mapping operations, and the intermediate feature map set of the weld seam is output layer by layer. During the calculation of the intermediate feature map of each layer of weld, the current input intermediate feature map of weld is the output intermediate feature map of weld of the previous layer; The output of the last convolutional coding layer of the convolutional neural network encoder structure is defined as the intermediate feature map of the weld seam, which is then defined as the deep weld seam feature map.

4. The machine vision-based weld quality inspection system for wind power steel pipe piles according to claim 1, characterized in that, The improved FeatureUp feature reconstruction module includes: A geometrical guidance field for welds is constructed based on preprocessed weld images, and the geometrical guidance field for welds and the feature map of deep welds are aligned to the same target space through spatial coordinate mapping. For each target pixel coordinate in the target space used to generate a high-resolution weld feature map, an inverse mapping relationship from the target pixel coordinate to the deep weld feature map coordinate is established, and a neighborhood sampling set is determined with the deep weld feature map coordinate as the center. The number of neighborhood sampling points in the neighborhood sampling set is determined by the sensitivity of the target pixel coordinate to minute defects. Based on the weld geometry guidance field, an anisotropic spatial metric matrix is ​​constructed at the coordinates of each target pixel, and different spatial bandwidths are assigned to the neighborhood sampling points along the weld direction and perpendicular to the weld direction. Based on the improved FeatUp algorithm, at each target pixel coordinate of the high-resolution weld feature map, the normalized reconstruction weight of each sampling point in the neighborhood sampling set is calculated by combining the anisotropic spatial metric matrix, the deep weld feature map and the sensitivity to minute defects. At each target pixel coordinate, the deep weld features of all sampling points in the neighborhood sampling set and the position encoding information of the target pixel coordinate are input into the learnable high-dimensional feature reconstruction mapping. The high-dimensional feature reconstruction mapping results of all sampling points are weighted and aggregated using normalized reconstruction weights to obtain the high-resolution weld feature vector at the target pixel coordinate. The high-resolution weld feature map is generated by taking the high-resolution weld feature vectors of all target pixel coordinates in the target space according to the target space size.

5. The machine vision-based weld quality inspection system for wind power steel pipe piles according to claim 1, characterized in that, The hollow attention fusion module includes: Based on the weld geometry guidance field, the corresponding weld width estimate is calculated for each target pixel coordinate in the target space, and the weld width estimates of all target pixel coordinates are combined to form a weld width estimation map. Based on the weld width estimation map, an expansion rate set is adaptively selected for the coordinates of each target pixel in the target space; Based on the dilation rate set, a multi-scale dilated convolution operation is performed on the high-resolution weld feature map to obtain a multi-scale context feature map corresponding to the dilation rate. In the target space, the feature vector of the high-resolution weld seam feature map at the corresponding coordinates is concatenated with the context feature vector of the multi-scale context feature map at the corresponding coordinates through the spatial attention mapping function, and then mapped. The spatial attention weights are obtained by using the Sigmoid function, and the spatial attention weights of all target pixel coordinates are combined to form the spatial attention weight vector. Channel attention weights are generated based on high-resolution weld feature maps, and channel attention weight vectors are obtained. In the target space, the weld width estimate at the target pixel coordinates is mapped together with the multi-scale context feature map at the expansion rate scale through the cross-scale attention mapping function to obtain the cross-scale attention response at the corresponding expansion rate scale. The cross-scale attention response at all expansion rate scales is normalized by an exponential function to obtain the cross-scale attention weights at all expansion rate scales at the target pixel coordinates. Based on multi-scale context feature maps, spatial attention weight vectors, channel attention weight vectors, and cross-scale attention weights, feature weighted fusion is performed on high-resolution weld feature maps to obtain context-guided weld feature maps.

6. The machine vision-based weld quality inspection system for wind power steel pipe piles according to claim 1, characterized in that, The weld multi-type defect segmentation module includes: In the decoder structure, the context-guided weld feature map is upsampled and fused layer by layer to generate the decoded weld feature map. The decoded weld feature map is then classified pixel by pixel using a classification mapping function to generate a pixel-level category logit tensor. Perform Softmax normalization on the pixel-level class logit tensor at each target pixel coordinate to obtain the pixel-level class probability tensor; Pixel-level weld defect segmentation masks are generated based on pixel-level category probability tensors, and defect category labels for each target pixel coordinate are generated by maximum probability decision. Small region removal processing is performed on the pixel-level weld defect segmentation mask, and the remaining segmentation mask only includes the weld defect candidate regions that satisfy the minimum area constraint; Perform connected component analysis on the preserved segmentation mask, and output the area, minimum bounding rectangle, and centroid coordinates of the defective connected component for each connected component. After removing small areas, the weld defect segmentation mask is subjected to edge refinement processing to obtain an edge-refined weld defect segmentation mask. Weld defect detection results are constructed by using defect category labeling, the area of ​​defect connected regions, the minimum bounding rectangle, centroid coordinates, and edge-refined weld defect segmentation mask.

7. The machine vision-based weld quality inspection system for wind power steel pipe piles according to claim 1, characterized in that, The weld defect confidence assessment module includes: Context-guided weld feature maps are extracted based on weld defect segmentation masks, extracting all connected components for each weld defect category. For each connected component in the context-guided weld feature map, the region confidence of the connected component is calculated. For each connected component in the context-guided weld feature map, the boundary confidence of the connected component is calculated. For each connected component in the context-guided weld feature map, the competition suppression confidence of the connected component is calculated. The context-guided weld feature map performs a weighted summation of the region confidence, boundary confidence, and competition suppression confidence of the connected domains to obtain a comprehensive confidence index for the connected domains. All comprehensive confidence indices of the connected domains are then combined to form the weld defect confidence assessment result.

8. The machine vision-based weld quality inspection system for wind power steel pipe piles according to claim 1, characterized in that, The weld quality classification module includes: Weld quality assessment standards were extracted from the historical database of weld quality inspection and the current quality specifications for the manufacturing and operation and maintenance of wind power steel pipe piles, and weld quality grading rules were constructed. Based on the weld defect detection results and weld defect confidence assessment results, the number of effective defect connected regions, the area ratio of effective defect connected regions, and the distribution density of effective defect connected regions in the axial and circumferential directions of the weld are statistically analyzed for each weld defect category to form a set of weld defect assessment parameters. Based on the set of weld defect assessment parameters and the weld quality grading rules, the weld quality status is determined, and the weld quality grade classification results are obtained.

9. The machine vision-based weld quality inspection system for wind power steel pipe piles according to claim 8, characterized in that, The weld quality grade classification results include: Qualified weld: The number of effective connected domains of crack defects is zero, and the number of effective connected domains of porosity defects, slag inclusion defects and non-fusion defects does not exceed the first number threshold, and the area ratio of effective defect connected domains does not exceed the first area threshold, and the distribution density of effective defect connected domains in the axial direction and circumferential direction of the weld is lower than the first distribution density threshold. Welds requiring re-inspection: There are no effective connected domains of through-crack defects, and the number of effective connected domains of at least one weld defect category exceeds the first quantity threshold but does not exceed the second quantity threshold, or the area ratio of at least one weld defect category exceeds the first area threshold but does not exceed the second area threshold, or the distribution density of effective defect connected domains in the axial or circumferential direction of the weld exceeds the first distribution density threshold but does not exceed the second distribution density threshold. Unqualified weld: There is an effective connected domain of crack defects, or the number of effective connected domains of any weld defect category exceeds the second quantity threshold, or the area ratio of any weld defect category exceeds the second area threshold, or the distribution density of effective defect connected domains in the axial or circumferential direction of the weld exceeds the second distribution density threshold.

Images