Pitting extreme depth distribution prediction method based on image of un-rust galvanized steel wire

Abstract

The application discloses a pitting extreme depth distribution prediction method based on a rusted galvanized steel wire image, takes a steel wire group as a prediction object, collects a rusted surface image of a training steel wire group in a training stage, generates a local image block through pretreatment, extracts corrosion characteristic features and carries out group-level fusion to form a training sample feature vector; three-dimensional measurement is carried out on the same steel wire group after rust removal, a pitting depth field is constructed, block extreme values are extracted and extreme value distribution is fitted, the obtained extreme pitting depth distribution parameters are taken as target outputs, and a pitting extreme depth distribution regression prediction model is trained; in a prediction stage, the same treatment is carried out on an image of a steel wire group to be measured and the trained model is input, and the pitting extreme depth distribution parameters of the steel wire group to be measured are output. The application establishes a quantitative mapping between the rusted image features and the three-dimensional pitting extreme value distribution parameters after rust removal, does not need rust removal and three-dimensional measurement in the prediction stage, and improves the field applicability of corrosion detection of parallel steel wire system components such as in-service bridge cables, hangers and main cables.

Description

Technical Field

[0001] This invention relates to the fields of corrosion detection of steel wires in bridge cables, assessment of pitting damage in metals, computer vision, and intelligent information processing, and specifically to a method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wires. Background Technology

[0002] Parallel galvanized high-strength steel wire is widely used as the main load-bearing component in cable-stayed structures such as bridge cables, suspenders, and main cables. During long-term service, this type of steel wire is susceptible to degradation of the galvanized layer, accumulation of rust products, and localized pitting corrosion due to factors such as humid heat cycling, chloride corrosion, rainwater infiltration, sheath aging, and water accumulation in the anchorage area. For high-strength steel wire, deep pitting corrosion leads to significant stress concentration, which is a major cause of fatigue crack initiation, reduced load-bearing capacity, and increased risk of wire breakage. Therefore, the pitting depth distribution parameter, obtained from block extreme value statistics and extreme value distribution fitting of the pitting depth field, is an important indicator for assessing steel wire corrosion risk, strength degradation analysis, and wire breakage risk.

[0003] Existing quantitative assessment methods for pitting damage on steel wires mainly include post-rust removal three-dimensional scanning, three-dimensional contour measurement, and microscopic morphology measurement, but they still have the following shortcomings:

[0004] (1) Methods such as three-dimensional scanning after rust removal usually require sampling, rust removal or pickling, and measurement under laboratory conditions. The detection process is complex and somewhat destructive, making it difficult to directly use for rapid on-site assessment of multiple galvanized steel wires in the window area of ​​in-service bridge components.

[0005] (2) Existing image-based corrosion identification methods are mostly used for corrosion area identification, corrosion level classification, rust area statistics or defect location detection. They mainly reflect the apparent corrosion characteristics in the unremoved state and are difficult to directly output the extreme depth distribution parameters of the steel wire group-level pitting corrosion obtained by extreme value statistics from the three-dimensional pitting corrosion data after rust removal.

[0006] (3) The image of the unrusted surface reflects the coverage of rust products, while the three-dimensional data after rust removal reflects the morphology of the base material loss. It is usually difficult to stably correspond between a single local image block and a single pit. If a point-by-point or local correspondence is forcibly established, it is easily affected by registration error and label uncertainty.

[0007] (4) Existing methods usually lack a modeling process that uses the steel wire group as a statistical unit to correspond the corrosion characterization features of multiple unremoved local image blocks in the same steel wire group with the extreme value distribution parameters obtained from the three-dimensional pitting corrosion data of the steel wire group after rust removal.

[0008] Therefore, there is a need for a method that can predict the extreme depth distribution parameters of pitting corrosion of steel wire groups based on the surface image of unremoved galvanized steel wires without requiring rust removal, pickling, or three-dimensional morphology measurement of the steel wire group under test during the prediction stage. Summary of the Invention

[0009] To address the problems of existing methods for assessing pitting depth in steel wires, which rely on post-rust removal 3D measurements and are difficult to use for rapid on-site assessment of in-service bridge components, and the difficulty in directly correlating images of unremoved surfaces with extreme statistical parameters of the true pitting depth after rust removal, this invention proposes a method for predicting the distribution of extreme pitting depth based on images of unremoved galvanized steel wires. The specific technical solution is as follows:

[0010] A method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire, characterized in that the method includes:

[0011] S1: Establish a training steel wire group set including multiple steel wire groups, the multiple training steel wire groups covering different corrosion stages, each training steel wire group including multiple galvanized steel wires, and obtain images of the unrusted surface of multiple galvanized steel wires in each steel wire group;

[0012] S2: The image of the unrusted surface is first preprocessed, and then a sliding window is used to divide it into blocks to generate multiple local image blocks belonging to the same wire group;

[0013] S3: Extract the corrosion characterization features of each local image patch to form a set of local image patch features belonging to each wire group;

[0014] S4: Select multiple subsets of local image block feature vectors that are at least partially different from the local image block feature set of the same wire group, and perform group-level fusion respectively to construct multiple sample feature vectors of the same wire group;

[0015] S5: After removing corrosion products, obtain three-dimensional surface data for multiple galvanized steel wires corresponding to the same steel wire group. Based on the three-dimensional surface data, construct a pitting depth field, extract the extreme depth of block pitting corrosion and perform extreme value distribution fitting to obtain the pitting extreme depth distribution parameters corresponding to the steel wire group.

[0016] S6: Combine the multiple sample feature vectors obtained from S4 corresponding to the same wire group with the pitting extreme depth distribution parameters obtained from S5 to form an input-output training pair, and use the input-output training pair formed by multiple training wire groups to train the pitting extreme depth distribution regression prediction model.

[0017] S7: Obtain images of the unrusted surfaces of multiple galvanized steel wires in the test group, and perform the same preprocessing, local image patch generation, corrosion characterization feature extraction, local image patch feature selection and group-level fusion as in S2~S4 to obtain the sample feature vector of the test group.

[0018] S8: Input the sample feature vector of the steel wire group to be tested into the trained pitting extreme depth distribution regression prediction model, and output the pitting extreme depth distribution parameters of the steel wire group to be tested.

[0019] Furthermore, the images of the unrusted surface include single images, multiple images, or multi-view images covering the visible surface of the galvanized steel wire;

[0020] In step S2, when the image of the unrusted surface contains multiple galvanized steel wires, the region of interest containing the galvanized steel wires is first extracted from the image, and then the galvanized steel wires in the region of interest are segmented into regions or instances to obtain a single steel wire image or a steel wire region image. Based on the single steel wire image or the steel wire region image, multiple local image blocks belonging to the same steel wire group are generated.

[0021] Furthermore, the steel wire sets used for training include galvanized steel wire sets from actual bridge service, galvanized steel wire sets subjected to artificial accelerated corrosion, or a combination of both;

[0022] The training wire set contains multiple wire sets corresponding to different corrosion stages or different corrosion states, which are used to establish the mapping relationship between the corrosion characterization features of unremoved surface images and the pitting depth distribution parameters under different corrosion development degrees.

[0023] Furthermore, the preprocessing in S2 includes sequentially performing wire region segmentation, boundary extraction, centerline fitting, and geometric correction;

[0024] The geometric correction is based on coordinate remapping of the wire centerline and its normal direction to correct the bent or tilted wire image into a strip image extending along the wire axis.

[0025] Furthermore, the corrosion characterization features in S3 include color features, texture features, edge features, and deep learning features; wherein, the color features include RGB features and HSV features, the texture features include local binary pattern features and gray-level co-occurrence matrix features, the edge features include edge statistical features based on gradient operators, and the deep learning features are obtained through a pre-trained neural network or a trained feature extraction network.

[0026] Furthermore, in step S4, the selection method for the subset of local image block feature vectors includes any one of random selection, layered selection according to wire number, or layered selection according to image position;

[0027] For the same group of steel wires, multiple sample feature vectors belonging to the same group are formed by selecting feature subsets of different local image blocks and fusing them at the group level.

[0028] Furthermore, the group-level fusion includes any one of average pooling, max pooling, mean statistics, discrete statistics, Top-k local feature aggregation, weighted pooling, attention-weighted aggregation, or multi-instance learning aggregation;

[0029] The extreme value distribution fitting in S5 adopts either Gumbel distribution fitting or generalized extreme value distribution fitting.

[0030] The pitting extreme depth distribution parameters include location parameters, scale parameters, and quantile parameters. The quantile parameters include the pitting depth quantile values ​​corresponding to a preset quantile level determined by the extreme value statistical distribution.

[0031] Furthermore, the construction of the pitting depth field in S5 specifically includes:

[0032] Three-dimensional surface data of galvanized steel wire after removing corrosion products are obtained by structured light three-dimensional scanning, blue light three-dimensional scanning, laser three-dimensional scanning, three-dimensional contour measurement or microscopic morphology measurement.

[0033] The three-dimensional surface data is processed to determine the radial positional relationship of the measurement point relative to the reference cylindrical surface;

[0034] The zero-depth reference surface is determined based on the reference cylindrical surface obtained by fitting the uncorroded area, the nominal radius of the steel wire, the measured uncorroded diameter, or the preset reference cylindrical surface.

[0035] The radial difference between the measurement point and the zero-depth reference surface is taken as the pitting depth to obtain the pitting depth field.

[0036] Furthermore, the extraction of the pitting extreme depth distribution parameter in S5 specifically includes:

[0037] Along the length of the galvanized steel wire, the pitting depth field is divided into multiple segments according to the preset segment division rules. The maximum pitting depth is extracted in each segment and taken as the pitting extreme depth of that segment.

[0038] The block pitting corrosion extreme depth of each section of multiple galvanized steel wires in the same steel wire group constitutes a block extreme value sequence;

[0039] The segmentation rules include any one of the following: segmentation based on fixed axial length, segmentation based on fixed point cloud quantity, or segmentation based on effective detection length and other proportions.

[0040] Furthermore, when the actual effective statistical length of the wire group to be tested is inconsistent with the effective statistical length of the wire group used for training, the predicted pitting extreme depth distribution parameter is corrected for length effect.

[0041] The length effect correction includes keeping the scale class parameter characterizing the dispersion of the distribution unchanged, and correcting the parameter characterizing the distribution location and / or quantile parameter according to the ratio of the effective statistical length of the wire group used for training to the actual effective statistical length of the wire group to be tested.

[0042] The beneficial effects of this invention are as follows:

[0043] (1) This invention takes steel wire group as the basic prediction object and establishes a mapping relationship between the corrosion characterization features of the unrusted surface image and the extreme statistical parameters of the three-dimensional measured pitting depth of the same steel wire group after rust removal. This enables the unrusted surface image to be used to predict the extreme distribution parameters of the true pitting depth of the steel wire group. This eliminates the need for rust removal, pickling, or three-dimensional morphology measurement of the steel wire group under test during the prediction stage. This improves the on-site applicability of corrosion detection for parallel steel wire system components such as cables, hangers, and main cables of in-service bridges.

[0044] (2) In this invention, the same steel wire group is used as the corresponding unit of the target output parameter. The feature vector of the training sample formed by the steel wire group is respectively connected with the same pitting corrosion extreme value depth distribution parameter to form multiple input-output training pairs, which improves the flexibility of training sample construction and enables the model to learn the correspondence between different local corrosion combinations and the extreme value distribution parameter of the steel wire group.

[0045] (3) This invention converts multiple local corrosion observations within the same steel wire group into fixed-dimensional sample feature vectors through group-level fusion, so that the model input and the steel wire group-level pitting corrosion extreme depth distribution parameters are consistent at the level, avoiding the errors and label uncertainties caused by point-by-point spatial registration of a single local image block with a single corrosion pit.

[0046] (4) The present invention outputs the pitting corrosion extreme depth distribution parameters, rather than just the corrosion type, rust area or defect location. Therefore, it can provide a more direct quantitative basis for steel wire strength degradation analysis, fatigue risk analysis, wire breakage risk assessment and bridge operation and maintenance decision-making. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the overall process of predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire, according to an embodiment of the present invention.

[0048] Figure 2 This is a flowchart of the image preprocessing process for rust-free galvanized steel wire in an embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram illustrating the generation of local image blocks in an embodiment of the present invention.

[0050] Figure 4 This is a flowchart of the extraction and group-level fusion of local image block erosion characterization features in an embodiment of the present invention.

[0051] Figure 5 This is a flowchart of the three-dimensional surface data processing and pitting depth field construction after rust removal in an embodiment of the present invention.

[0052] Figure 6 This is a flowchart of the process for extracting the extreme depth of pitting corrosion and fitting the extreme value distribution in an embodiment of the present invention. Detailed Implementation

[0053] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0054] On the one hand, such as Figure 1 As shown, this invention provides a method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire, including a training phase and a prediction phase. This method uses a group of multiple galvanized steel wires as the basic prediction object. It should be noted that this invention predicts the extreme depth distribution parameters of pitting corrosion corresponding to the entire group of steel wires, not the pitting depth of a single local image patch, nor the maximum pitting depth of a single galvanized steel wire. Steps one to six correspond to the model training process, and steps seven to eight correspond to the prediction process of the group of steel wires to be tested.

[0055] Step 1: Establish a training wire set including multiple wire groups, each of which includes multiple galvanized wires; obtain images of the unrusted surfaces of the multiple galvanized wires in each wire group.

[0056] During the training phase, the steel wire group can include multiple galvanized steel wires from the same salt spray corrosion time, the same corrosion test stage, the same service inspection area, or the same window inspection area. The multiple steel wire groups included in the steel wire group set preferably correspond to different corrosion stages or different corrosion states. This allows the prediction model to learn the correspondence between the corrosion characterization features of the unrusted surface image under different corrosion development degrees and the distribution parameters of the pitting depth extreme value of the steel wire group.

[0057] The preferred set of steel wire assemblies for training includes galvanized steel wire assemblies used in actual bridge operations, galvanized steel wire assemblies subjected to artificially accelerated corrosion, or a combination of both.

[0058] In this embodiment of the invention, hot-dip galvanized high-strength steel wire was used as the training sample. The nominal diameter of the wire was 7 mm, the zinc coating mass was approximately 343 g / m², and the corresponding zinc coating thickness was approximately 48 μm. Galvanized steel wire samples at different corrosion stages were obtained using a salt spray test. The salt spray test temperature was 35 ± 2 °C, and the corrosive medium was a 5% sodium chloride solution. The salt spray corrosion times included 25 days, 35 days, 45 days, 55 days, 65 days, 80 days, 95 days, and 110 days. Each corrosion stage included 10 galvanized steel wires, thus forming 8 groups of steel wires for training, totaling 80 galvanized steel wires. The above-mentioned wire dimensions, zinc coating parameters, corrosive medium, corrosion time, and number of wires per group are only one specific embodiment, and the invention is not limited thereto.

[0059] For the galvanized steel wires in each wire group, surface images were acquired without removing corrosion products. Images of the untreated surface include single images, multiple images, or multi-view images covering the visible surface of the galvanized steel wire. The image acquisition range, image scale, lighting conditions, color calibration methods, reflection suppression methods, or preprocessing methods were kept consistent between the training and prediction phases to minimize the impact of differences in shooting conditions on corrosion characterization features.

[0060] Step 2: First, preprocess the image of the unrusted surface, then divide it into blocks using a sliding window to generate multiple local image blocks belonging to the same wire group.

[0061] During image acquisition or preprocessing, one or more of the following processes can be performed: illumination control, white balance correction, color calibration, or reflection suppression, in order to reduce the impact of differences in shooting conditions on corrosion characterization features.

[0062] If the original image contains multiple steel wires, the region of interest containing the steel wires is first extracted from the original image. Then, the steel wires within the region of interest are segmented into regions or instances to obtain a single steel wire image or a steel wire region image. Based on the single steel wire image or the steel wire region image, multiple local image patches belonging to the same training steel wire group are generated.

[0063] like Figure 2 As shown, image preprocessing includes sequentially performing wire region segmentation, boundary extraction, centerline fitting, and geometric correction.

[0064] The segmentation of the wire region can be achieved using manual annotation, threshold segmentation, edge detection, semantic segmentation models, or instance segmentation models. Geometric correction is based on coordinate remapping of the wire centerline and its normal direction to correct the curved or tilted wire image into a strip-shaped image extending along the wire axis.

[0065] In this embodiment of the invention, a binary mask M(i,j) is first constructed based on the wire region, wherein when pixel (i,j) belongs to the wire region, M(i,j)=1; when pixel (i,j) does not belong to the wire region, M(i,j)=0.

[0066] For the i-th row of the image, calculate the horizontal centroid c of the foreground pixels in the wire region based on the binary mask. i :

[0067]

[0068] Discrete center point (c) obtained from each row i i) Fit the centerline of the steel wire. Preferably, a cubic polynomial P(v) is used to fit the discrete center points.

[0069] For a pixel (x, y) in the corrected image, its vertical coordinate v in the original image is represented as:

[0070]

[0071] Among them, v min The coordinates are the starting row coordinates of the wire region in the original image.

[0072] Furthermore, the normal lateral component of the centerline used for lateral remapping at position v is represented as:

[0073]

[0074] Among them, P ' (v) is the first derivative of the centerline fitting function P(v) at v.

[0075] Correspondingly, the pixel (x,y) in the corrected image can be mapped to the coordinates (u,v) in the original image, and its horizontal coordinate u is expressed as:

[0076]

[0077] Where w is the width of the corrected image.

[0078] Through the above coordinate remapping, the image of a bent or tilted wire can be corrected into a strip image extending along the axial direction of the wire.

[0079] After preprocessing is completed, such as Figure 3 As shown, along the axial direction of the galvanized steel wire, the geometrically corrected steel wire image is divided into multiple local image blocks by sliding window segmentation.

[0080] Preferably, the training and prediction phases use the same block size. In this embodiment of the invention, the local image block is a 224×224 pixel image block, and overlapping areas can be set between adjacent sliding windows, for example, an overlap rate of 30%. The pixel size and overlap rate of the local image blocks can be adjusted according to the camera resolution, wire diameter, shooting distance, and model input requirements. Furthermore, the number of local image blocks can be expanded using enhancement methods such as rotation and mirroring, without changing the basic characteristics of the erosion morphology.

[0081] Step 3: Extract the corrosion characterization features of each local image block to form a set of local image block features belonging to each wire group.

[0082] For the k-th wire group, all its local image patches are combined into a local image patch set:

[0083]

[0084] Among them, P k I represents the set of local image patches for the k-th wire group. k,j N represents the j-th local image patch in the k-th wire group. k This represents the total number of local image blocks corresponding to this group of steel wires.

[0085] For the set of local image patches P k Corrosion characterization features are extracted from each local image patch in the image, forming a set of local image patch features belonging to the k-th wire group:

[0086]

[0087] Among them, f k,j This represents the corrosion characterization feature of the k-th local image block in the wire group.

[0088] Corrosion characterization features are represented in the form of numerical feature vectors. Corrosion characterization features of multiple categories can be concatenated to form a multi-source corrosion characterization feature vector for a local image patch.

[0089] For the k-th local image patch used in the wire group, its multi-source corrosion characterization feature vector can be expressed as:

[0090]

[0091] Among them, f k,j RGB and f k,j HSV Color characteristics are used to characterize the color changes of rust products and the apparent degradation state of the zinc plating layer; f k,j LBP and f k,j GLCMTexture features are used to characterize the roughness, mottledness, and local texture differences of corrosion products; f k,j Sobel Edge features are used to characterize corrosion boundaries, local abrupt changes, and pitting corrosion-related morphological changes; f k,j Deep These are optional deep learning features. Deep learning features can be learned features obtained through pre-trained neural networks or trained feature extraction networks. Specifically, for example... Figure 4 As shown.

[0092] In this embodiment of the invention, color features include RGB features and HSV features. Normalized histograms for the R, G, and B channels of the RGB image are extracted, with each channel divided into 10 intervals, forming a 30-dimensional RGB feature. Normalized histograms for the H, S, and V channels of the HSV image are extracted, with each channel divided into 10 intervals, forming a 30-dimensional HSV feature. Texture features include Local Binary Pattern (LBP) features and Gray-Level Co-occurrence Matrix (GLCM) features. LBP features can form a 10-dimensional LBP feature, and GLCM features can form a 48-dimensional GLCM feature. Edge features are calculated using the Sobel operator to compute gradient magnitude maps, and edge statistical features, including edge density, gradient mean, gradient standard deviation, maximum gradient magnitude, and gradient intensity histogram, are extracted to form a 14-dimensional Sobel edge feature. Deep learning features can be obtained through pre-training the EfficientNet-B0 network, and 1280-dimensional deep learning features are extracted after removing the final classification layer.

[0093] Therefore, in this embodiment of the invention, the spliced ​​local image patch multi-source erosion representation feature vector has 1412 dimensions. If deep learning features are not used, the local image patch multi-source erosion representation feature vector is formed by splicing RGB features, HSV features, LBP features, GLCM features, and Sobel edge features. The above feature categories, feature dimensions, and network structures are only preferred embodiments. This invention does not limit the specific dimensions and extraction models of various features. For example, RGB features and HSV features in color features can be used simultaneously or selectively; binary pattern features and gray-level co-occurrence matrix features in texture features can be used simultaneously or selectively.

[0094] It should be noted that this step yields a set of corrosion characterization features for multiple local image blocks within the same wire group. The corrosion characterization features of a single local image block are not directly used as a supervisory sample for the depth of a single pit.

[0095] Step 4: Select multiple subsets of local image block feature vectors that are at least partially different from the local image block feature set of the same wire group, and perform group-level fusion on each subset to construct multiple sample feature vectors belonging to the same wire group.

[0096] From the local image patch feature set F of the k-th wire group k Select n local image patch features to form the t-th local image patch feature subset:

[0097]

[0098] Among them, S k,t f represents the subset of local image patch features selected from the k-th training wire group at the t-th time. k,tq This represents the q-th local image patch feature in the feature subset, where n is the number of local image patch features participating in group-level fusion.

[0099] The selection of local image patch features can be done randomly, or by selecting in layers according to wire number or by selecting in layers according to image position. In this embodiment of the invention, each subset of local image patch features contains 5 local image patch features, i.e., n=5.

[0100] For a local image patch feature subset S k,t Group-level fusion is performed to obtain the sample feature vector:

[0101]

[0102] Among them, X k,t Let A(⋅) be the feature vector of the training sample formed by the t-th selection of the k-th training wire group, and let A(⋅) be the group-level fusion function.

[0103] The same group of steel wires can be used to select and fuse multiple different subsets of local image features at the group level to form multiple sample feature vectors:

[0104]

[0105] Among them, T k This represents the number of sample feature vectors formed by the k-th wire group.

[0106] Group-level fusion functions include one or more of the following: average pooling, max pooling, mean statistics, discrete statistics, Top-k local feature aggregation, weighted pooling, attention-weighted aggregation, or multi-instance learning aggregation. Group-level fusion employs a fusion method independent of the order of local image patches, fusing features from multiple local image patches selected within the same wire group into a fixed-dimensional sample feature vector.

[0107] In this embodiment of the invention, average pooling is used as the group-level fusion method, that is, the average of multiple feature vectors in the same feature subset of the same local image block is calculated dimension by dimension:

[0108]

[0109] Where n is the feature subset S of the local image patch k,t The number of local image patch features.

[0110] When the feature vector representing the multi-source erosion of each local image patch is 1412-dimensional, the feature vector of the training samples obtained after average pooling is still 1412-dimensional.

[0111] In another implementation, max pooling, Top-k local feature aggregation, or attention-weighted aggregation can be used to enhance the contribution of severely eroded local regions to the feature vectors of training samples.

[0112] The purpose of group-level fusion is to transform multiple localized corrosion observations within the same steel wire group into fixed-dimensional sample feature vectors. These sample feature vectors are used to characterize the overall corrosion level and localized severe corrosion state resulting from multiple localized corrosion observations within the same steel wire group.

[0113] Since the subsequent target output parameters are the wire group-level pitting extreme depth distribution parameters, rather than the parameters of a single local image patch, group-level fusion can keep the model input and target output parameters consistent at the hierarchical level.

[0114] Step 5: After removing corrosion products, obtain three-dimensional surface data for multiple galvanized steel wires corresponding to the same steel wire group. Construct a pitting depth field based on the three-dimensional surface data, extract the extreme depth of block pitting, and fit the extreme value distribution to obtain the pitting extreme depth distribution parameters corresponding to the steel wire group.

[0115] When removing corrosion products from multiple galvanized steel wires corresponding to the same steel wire group, pickling, mechanical cleaning, or other methods that do not significantly change the pitting morphology of the substrate can be used.

[0116] like Figure 5 As shown, the construction of the pitting depth field specifically includes:

[0117] Three-dimensional surface data of galvanized steel wire after removing corrosion products are obtained by structured light three-dimensional scanning, blue light three-dimensional scanning, laser three-dimensional scanning, three-dimensional contour measurement or microscopic morphology measurement.

[0118] Three-dimensional data processing is performed on the three-dimensional surface data to determine the radial positional relationship of the measurement point relative to the reference cylindrical surface;

[0119] The zero-depth reference surface is determined based on the reference cylindrical surface obtained by fitting the uncorroded area, the nominal radius of the steel wire, the measured uncorroded diameter, or the preset reference cylindrical surface.

[0120] The radial difference between the measurement point and the zero-depth reference surface is taken as the pitting depth to obtain the pitting depth field.

[0121] In this embodiment of the invention, after coordinate processing, the axial coordinate z is taken as the direction of the steel wire axis, and the xy cross-sectional coordinate is established with the cross-sectional direction perpendicular to the steel wire axis. The radial distance of the measurement point is expressed as:

[0122]

[0123] Pitting depth d i Represented as:

[0124]

[0125] Where R0 is the reference radius corresponding to the zero-depth reference cylinder, r i Let d be the radial distance of the i-th measurement point. i Let be the pitting depth at the i-th measurement point.

[0126] When the measurement point is below the zero-depth reference cylindrical surface, d i This indicates the radial material loss depth at this location relative to the zero-depth reference cylindrical surface.

[0127] like Figure 6 As shown, along the length of the galvanized steel wire, the pitting depth field is divided into multiple segments according to the preset segment division rules. The maximum pitting depth is extracted in each segment and used as the block pitting extreme depth of that segment.

[0128] The preset segment division rules can be any one of the following: division by fixed axial length, division by fixed point cloud quantity, or division by effective detection length.

[0129] In this embodiment of the invention, a statistical segment is divided every 1 mm along the axial direction of the steel wire, and the maximum pitting depth in each 1 mm segment is selected as the block pitting extreme depth of that segment; when the remaining length at the end is less than a complete segment, the end region can be discarded or merged into an adjacent segment.

[0130] For the k-th wire group used for training, the block extreme value sequence is composed of the block pitting corrosion extreme values ​​of each segment of multiple galvanized wires within the group:

[0131]

[0132] Among them, B k Let b represent the sequence of extreme pitting depths for the k-th wire group. k,m M represents the extreme pitting depth of the m-th block in the wire group. k This indicates the number of extreme pitting depths corresponding to the steel wire group.

[0133] The extreme value distribution of the block extreme value sequence is fitted to obtain the pitting depth distribution parameters corresponding to the kth wire group used for training, which are used as the target output parameters for the wire group used for training.

[0134] Preferably, the block extreme value sequence is fitted using a Gumbel distribution, and its distribution function is:

[0135]

[0136] Where μ is the position parameter and σ is the scale parameter.

[0137] Furthermore, the pitting depth quantile value Q corresponding to the preset quantile level p. p for:

[0138]

[0139] In this embodiment of the invention, p is set to 0.99, resulting in a 99% cleavage depth Q. 99 Therefore, μ can be... k σ k and Q 99,k The target output parameter for the k-th training wire group is denoted as:

[0140]

[0141] It should be noted that the target output parameter Y k These are parameters at the wire group level, rather than parameters for a single local image patch, a single image, or a single galvanized wire. The same wire group used for training can form multiple training sample feature vectors X. k,t Multiple training sample feature vectors are respectively associated with the same wire group-level target output parameter Y. k correspond.

[0142] Step 6: Combine the feature vectors of multiple samples obtained in Step 4 with the pitting depth distribution parameters obtained in Step 5 to form an input-output training pair, and use the input-output training pair formed by multiple training wire groups to train the pitting depth distribution regression prediction model.

[0143] This invention establishes a statistical correspondence at the wire group level, rather than a point-by-point spatial correspondence between local image patches and three-dimensional pits after corrosion products are removed. The training sample feature vector is obtained by group-level fusion of features from multiple local image patches within the same training wire group, and the target output parameter is obtained by fitting the extreme value distribution of the pitting corrosion extreme depths of multiple galvanized wires within the same training wire group. The two form an input-output training pair with the same training wire group as the corresponding unit.

[0144] Using the same wire group as the corresponding unit for the target output parameter, the multiple sample feature vectors formed from this wire group in step four are respectively combined with the same point erosion extreme value depth distribution parameter obtained from this wire group in step five to form multiple input-output training pairs. For the k-th wire group used for training, its input-output training pair can be expressed as:

[0145]

[0146] Among them, X k,t Y represents the sample feature vector formed by the t-th selection and group-level fusion of the k-th wire group. k T represents the depth distribution parameter of the same pitting corrosion extreme value corresponding to the k-th wire group. k This represents the number of sample feature vectors formed by the k-th wire group.

[0147] The input-output training of multiple steel wire groups is used to train a regression prediction model for the extreme depth distribution of pitting corrosion, so as to establish the mapping relationship between the combination of local corrosion features of galvanized steel wire surface images in the unremoved state and the parameters of the extreme depth distribution of pitting corrosion after the removal of corrosion products from the same steel wire group.

[0148] During model training and performance evaluation, the training set, validation set, and test set are preferably divided according to the training wire group, or at least ensure that all training sample feature vectors derived from the same training wire group do not appear in the training set, validation set, and test set at the same time, so as to avoid label leakage of the same wire group-level target output parameters between different datasets.

[0149] Pitting extreme depth distribution regression prediction models include gradient boosting tree regression model, random forest regression model, support vector regression model, neural network regression model, multi-objective regression model, or a combination thereof.

[0150] In this embodiment of the invention, the XGBoost multi-objective regression model is used as the prediction model. The input to the prediction model is the feature vector X of the training samples. k,t The output is:

[0151]

[0152] Where, μ pred σ pred and Q p,pred Let Q represent the predicted pitting depth quantiles corresponding to the location parameter, scale parameter, and preset quantile level p, respectively; when p = 0.99, Q p,pred The 99% cleavage depth Q 99,pred .

[0153] Furthermore, the pitting depth distribution parameters output by the model can be subject to reasonable constraints or post-processing to ensure that the scale parameters are positive and the quantile parameters are non-negative. In one specific implementation, the scale parameters are logarithmically transformed before model training and prediction.

[0154] Furthermore, cross-validation, random search, grid search, or other hyperparameter optimization methods can be used to optimize the parameters of the prediction model. In one specific implementation, 5-fold cross-validation and random sampling iteration can be used to optimize the number of trees, learning rate, maximum tree depth, subsample sampling ratio, column sampling ratio, and regularization coefficient of the XGBoost model with the minimum mean squared error as the optimization objective.

[0155] The above-described model and its parameter optimization methods are merely preferred implementation methods. This invention does not limit the specific type of prediction model or the specific hyperparameters.

[0156] Step 7: Obtain images of the unrusted surfaces of multiple galvanized steel wires in the test group, and perform the same preprocessing, local image block generation, corrosion characterization feature extraction, local image block feature selection and group-level fusion as in S2~S4 to obtain the sample feature vector of the test group.

[0157] During the prediction phase, images of the unrusted surface of the cable, boom, main cable, or tie rod window detection area to be tested are acquired. The unrusted surface images contain multiple galvanized steel wires to be tested. Preferably, the image acquisition range, image scale, lighting conditions, or preprocessing method during the prediction phase are consistent with those during the training phase, or are consistent with those during the training phase after white balance correction, color calibration, or reflection suppression.

[0158] The region of interest (ROI) containing galvanized steel wire is extracted from the image of the unrusted surface. Then, the galvanized steel wire within the ROI is segmented into regions or instances to obtain an image of the galvanized steel wire to be tested. Multiple galvanized steel wires within the same windowed detection area are identified as a group of wires to be tested.

[0159] The images of galvanized steel wires in the test group undergo the same or corresponding preprocessing as the training phase, including local image block generation, corrosion characterization feature extraction, local image block feature selection, and group-level fusion, to obtain the sample feature vector of the test group.

[0160] Specifically, the image of the galvanized steel wire to be tested is subjected to boundary extraction, centerline fitting, and geometric correction to generate multiple local image blocks along the wire axis; color features, texture features, edge features, and optional deep learning features are extracted from each local image block; multiple local image block features are selected from the local image block feature set of the steel wire group to be tested to form a local image block feature subset; and then the local image block feature subset is fused at the group level to obtain the sample feature vector of the steel wire group to be tested.

[0161] For the same group of steel wires to be tested, a subset of local image features can be selected once or multiple times, and one or more sample feature vectors can be formed respectively. When multiple sample feature vectors are formed, the multiple sample feature vectors can be input into the trained prediction model, and the multiple prediction results can be averaged, weighted averaged, statistically analyzed by median or other statistical methods to obtain the pitting extreme depth distribution parameters of the steel wire group to be tested.

[0162] Step 8: Input the sample feature vector of the steel wire group to be tested into the trained pitting corrosion extreme depth distribution regression prediction model, and output the pitting corrosion extreme depth distribution parameters of the steel wire group to be tested.

[0163] The extreme depth distribution parameters of pitting corrosion of the steel wire group under test include location parameters, scale parameters, and the pitting depth quantile value Q corresponding to the preset quantile level p. p .

[0164] In this embodiment of the invention, p=0.99, and the output includes position parameters, scale parameters, and the 99% quantile erosion depth Q. 99 .

[0165] The prediction stage does not require rust removal, pickling, or three-dimensional morphology measurement of the galvanized steel wires in the test wire group.

[0166] Furthermore, when the actual effective statistical length of the wire group under test is inconsistent with the effective statistical length of the wire group used for training, the pitting depth distribution parameters output by the model can be corrected based on the length effect of the extreme value distribution. In the embodiment of the invention using Gumbel distribution fitting, the scale parameter remains unchanged, and the position parameter and the pitting depth quantile value corresponding to the preset quantile level p are corrected according to the ratio of the effective statistical length of the wire group used for training to the actual effective statistical length of the wire group under test, the expression of which is:

[0167]

[0168]

[0169]

[0170] Among them, L train L represents the effective statistical length of the wire during the training phase. actual σ represents the actual effective statistical length corresponding to the group of steel wires to be tested. pred μ pred and Q p,pred These represent the scale parameter, location parameter, and pitting depth quantile corresponding to the preset quantile level p, respectively, from the prediction model output. final μ final and Q p,finalThese represent the scale parameter, location parameter, and pitting depth quantile after length effect correction. When p = 0.99, Q... p,pred The 99% cleavage depth Q 99,pred Q p,final The 99% cleavage depth Q is the length-effect corrected value. 99,final .

[0171] The length effect correction mentioned above is an optional step, applicable to scenarios where it is necessary to compare the pitting depth distribution parameters between different effective statistical lengths.

[0172] Furthermore, the predicted pitting depth distribution parameters can be compared with a preset threshold to obtain the corrosion risk level of the steel wire group to be tested or its corresponding window detection area.

[0173] The extreme depth distribution parameter of pitting corrosion can also be used for steel wire strength degradation analysis, fatigue risk analysis, wire breakage risk assessment, or bridge operation and maintenance decision-making.

[0174] This invention is illustrated using examples of 7mm hot-dip galvanized high-strength steel wire, 8 salt spray corrosion stages, 10 steel wires per stage, 224×224 pixel axial local image blocks, n=5 local image block feature subsets, 1412-dimensional multi-source corrosion characterization features, average pooling group-level fusion, three-dimensional scanning pitting depth field, Gumbel distribution fitting, and XGBoost multi-objective regression model. However, this invention is not limited to the specific combination described above. Those skilled in the art can, according to actual detection conditions, replace artificial accelerated corrosion samples with actual bridge service samples, replace average pooling with max pooling, Top-k local feature aggregation, weighted pooling, attention-weighted aggregation, or multi-instance learning aggregation, replace Gumbel distribution with generalized extreme value distribution, and replace the XGBoost regression model with other regression prediction models. All of these modifications do not affect the core technical concept of this invention: predicting the extreme depth distribution parameters of pitting corrosion of steel wire groups from unremoved galvanized steel wire surface images. As long as the steel wire group is still used as the corresponding unit, the training sample feature vector formed by the group-level fusion of features of multiple unremoved local image blocks, and the input-output training pair are formed with the pitting depth distribution parameters of the same steel wire group, all belong to the core technical concept of this invention.

[0175] To verify the effectiveness of the method of this invention, the trained regression prediction model for pitting corrosion extreme depth distribution was used to predict the pitting corrosion extreme depth distribution parameters of the steel wire group under test. The prediction results were then compared with the measured fitting parameters obtained based on the three-dimensional surface data after rust removal, the block pitting corrosion extreme depth sequence, and the Gumbel distribution. The comparison results may include the location parameter μ, the scale parameter σ, and the high-quantile pitting corrosion depth Q. p The results, including the absolute or relative errors, are shown in Table 1.

[0176] Table 1. Comparison of predicted and measured fitting parameters for extreme distribution of pitting corrosion on steel wire in a specific embodiment.

[0177]

[0178] As shown in Table 1, the prediction results based on the surface image of the rust-free steel wire generally agree well with the measured fitting results, indicating that the method can accurately predict the extreme depth distribution of pitting corrosion on the steel wire. Specifically, the relative error of the scale parameter σ is only 3.56%, indicating that the model accurately predicts the dispersion of pitting corrosion depth. The 99% quantile pitting corrosion depth Q, which is of greater concern in engineering evaluation, is also a significant factor. 99 The relative error was 5.80%, and the predicted value of 206.47 μm was slightly higher than the measured fitted value of 195.15 μm, indicating that the method has a certain degree of conservatism in the assessment of extreme pitting depth, which is beneficial for safety assessment. The relative error of the position parameter μ was 15.31%, indicating that the model has a good characterization ability for the overall level of the distribution of extreme pitting depth.

[0179] Overall, Table 1 verifies the feasibility and engineering application potential of this method in predicting the extreme depth distribution of pitting corrosion from images of untreated surfaces, and can provide parameter basis for subsequent evaluation of steel wire corrosion damage and load-bearing performance assessment.

[0180] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for predicting the pitting extreme depth distribution based on the image of a non-depatinated galvanized steel wire, characterized in that, The method includes: S1: Establish a training steel wire group set including multiple steel wire groups, the multiple training steel wire groups covering different corrosion stages, each training steel wire group including multiple galvanized steel wires, and obtain images of the unrusted surface of multiple galvanized steel wires in each steel wire group; S2: The image of the unrusted surface is first preprocessed, and then a sliding window is used to divide it into blocks to generate multiple local image blocks belonging to the same wire group; S3: Extract the corrosion characterization features of each local image patch to form a set of local image patch features belonging to each wire group; S4: Select multiple subsets of local image block feature vectors that are at least partially different from the local image block feature set of the same wire group, and perform group-level fusion respectively to construct multiple sample feature vectors of the same wire group; S5: After removing corrosion products, obtain three-dimensional surface data for multiple galvanized steel wires corresponding to the same steel wire group. Based on the three-dimensional surface data, construct a pitting depth field, extract the extreme depth of block pitting corrosion and perform extreme value distribution fitting to obtain the pitting extreme depth distribution parameters corresponding to the steel wire group. S6: Combine the multiple sample feature vectors obtained from S4 corresponding to the same wire group with the pitting extreme depth distribution parameters obtained from S5 to form an input-output training pair, and use the input-output training pair formed by multiple training wire groups to train the pitting extreme depth distribution regression prediction model. S7: Obtain images of the unrusted surfaces of multiple galvanized steel wires in the test group, and perform the same preprocessing, local image patch generation, corrosion characterization feature extraction, local image patch feature selection and group-level fusion as in S2~S4 to obtain the sample feature vector of the test group. S8: Input the sample feature vector of the steel wire group to be tested into the trained pitting extreme depth distribution regression prediction model, and output the pitting extreme depth distribution parameters of the steel wire group to be tested.

2. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 1, characterized in that, The images of the unrusted surface include single images, multiple images, or multi-view images covering the visible surface of the galvanized steel wire; In step S2, when the image of the unrusted surface contains multiple galvanized steel wires, the region of interest containing the galvanized steel wires is first extracted from the image, and then the galvanized steel wires in the region of interest are segmented into regions or instances to obtain a single steel wire image or a steel wire region image. Based on the single steel wire image or the steel wire region image, multiple local image blocks belonging to the same steel wire group are generated.

3. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 1, characterized in that, The steel wire sets used for training include galvanized steel wire sets from actual bridge service, galvanized steel wire sets from artificially accelerated corrosion, or a combination of both. The training wire set contains multiple wire sets corresponding to different corrosion stages or different corrosion states, which are used to establish the mapping relationship between the corrosion characterization features of unremoved surface images and the pitting depth distribution parameters under different corrosion development degrees.

4. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 1, characterized in that, The preprocessing in S2 includes sequentially performing wire region segmentation, boundary extraction, centerline fitting, and geometric correction. The geometric correction is based on coordinate remapping of the wire centerline and its normal direction to correct the bent or tilted wire image into a strip image extending along the wire axis.

5. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 1, characterized in that, The corrosion characterization features in S3 include color features, texture features, edge features, and deep learning features; wherein, the color features include RGB features and HSV features, the texture features include local binary pattern features and gray-level co-occurrence matrix features, the edge features include edge statistical features based on gradient operators, and the deep learning features are obtained through a pre-trained neural network or a trained feature extraction network.

6. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 1, characterized in that, In step S4, the selection method for the subset of local image block feature vectors includes any one of random selection, layered selection according to wire number, or layered selection according to image position. For the same group of steel wires, multiple sample feature vectors belonging to the same group are formed by selecting feature subsets of different local image blocks and fusing them at the group level.

7. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 6, characterized in that, The group-level fusion includes average pooling, max pooling, mean statistics, dispersion statistics, and Top-... k Any one of the following: local feature aggregation, weighted pooling, attention-weighted aggregation, or multi-instance learning aggregation; The extreme value distribution fitting in S5 adopts either Gumbel distribution fitting or generalized extreme value distribution fitting. The pitting extreme depth distribution parameters include location parameters, scale parameters, and quantile parameters. The quantile parameters include the pitting depth quantile values ​​corresponding to a preset quantile level determined by the extreme value statistical distribution.

8. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 1, characterized in that, The construction of the pitting depth field in S5 specifically includes: Three-dimensional surface data of galvanized steel wire after removing corrosion products are obtained by structured light three-dimensional scanning, blue light three-dimensional scanning, laser three-dimensional scanning, three-dimensional contour measurement or microscopic morphology measurement. The three-dimensional surface data is processed to determine the radial positional relationship of the measurement point relative to the reference cylindrical surface; The zero-depth reference surface is determined based on the reference cylindrical surface obtained by fitting the uncorroded area, the nominal radius of the steel wire, the measured uncorroded diameter, or the preset reference cylindrical surface. The radial difference between the measurement point and the zero-depth reference surface is taken as the pitting depth to obtain the pitting depth field.

9. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 8, characterized in that, The extraction of the pitting extreme depth distribution parameters in S5 specifically includes: Along the length of the galvanized steel wire, the pitting depth field is divided into multiple segments according to the preset segment division rules. The maximum pitting depth is extracted in each segment and taken as the pitting extreme depth of that segment. The block pitting corrosion extreme depth of each section of multiple galvanized steel wires in the same steel wire group constitutes a block extreme value sequence; The segmentation rules include any one of the following: segmentation based on fixed axial length, segmentation based on fixed point cloud quantity, or segmentation based on effective detection length and other proportions.

10. The method for predicting the extreme depth distribution of pitting corrosion based on images of unremoved galvanized steel wire according to claim 1, characterized in that, When the actual effective statistical length of the wire group to be tested is inconsistent with the effective statistical length of the wire group used for training, the predicted pitting extreme depth distribution parameter is corrected for length effect. The length effect correction includes keeping the scale class parameter characterizing the dispersion of the distribution unchanged, and correcting the parameter characterizing the distribution location and / or quantile parameter according to the ratio of the effective statistical length of the wire group used for training to the actual effective statistical length of the wire group to be tested.

Images