Method for identifying steel bar cracks based on deep learning

Abstract

The application relates to the technical field of feature extraction and retrieval, and discloses a method for identifying steel mesh cracks based on deep learning, which collects digital images and physical side lengths through a visual sensor and a laser range finder, extracts crack skeleton points and grid intersection node coordinates through a deep learning network, then extracts anti-noise structure invariants based on the relationship between spatial topological distance and stress distribution, and generates a database retrieval primary key for each crack skeleton point in combination with global coordinates, finally, the matching similarity score of the retrieval primary key and a historical retrieval key value is calculated, a dynamic retrieval tolerance threshold is generated according to the structure invariants, and database matching retrieval is completed through constraint judgment. The technical scheme effectively overcomes feature extraction noise and data tilt, and realizes high-precision and rapid retrieval matching of homologous cracks under multiple scales.

Description

Technical Field

[0001] This invention relates to the field of feature extraction and retrieval technology, specifically a method for identifying cracks in steel mesh based on deep learning. Background Technology

[0002] During the service life of building engineering and bridge structures, surface cracks in steel mesh are an important indicator for assessing the service condition of the structure. Currently, visual sensors are often deployed to collect digital images of cracks, and then deep learning networks are used to process the images. This existing technology can quickly extract the pixel coordinates and morphological features of cracks. However, when performing homologous damage retrieval in a massive historical crack evolution database, existing methods face significant technical bottlenecks.

[0003] First, existing image recognition only outputs isolated two-dimensional coordinates, which are completely detached from the physical support structure of the underlying steel mesh. Simple coordinate comparison cannot reflect the spatial topological relationship between cracks and mesh nodes. Furthermore, pure geometric distance ignores the main stress transmission characteristics of the steel structure, which leads to cracks with drastically different stress states being misclassified by the system. Second, crack skeleton points are extremely densely distributed in local areas. If features are directly extracted and stored, it is easy to cause index distribution skew in the database retrieval tree. A large number of similar features lead to hash collisions, which severely slows down the retrieval and matching speed. In addition, the lighting in actual detection scenarios is complex, and the local edges extracted by deep learning models have random jitter noise. This high-frequency oscillation of features will cause the query key value of the same damage to change drastically at different times, resulting in a break in historical tracing. Finally, the physical scale of steel mesh varies greatly in different projects. Existing systems mostly rely on fixed similarity thresholds for judgment. This rigid judgment standard is prone to causing missed detections at large scales and false alarms at small scales when processing heterogeneous data.

[0004] Therefore, there is an urgent need for a deep learning-based method to identify cracks in steel mesh, in order to solve the problems of topological missingness, index skew, noise interference, and scale inconsistency that exist in database retrieval using pure image coordinates. Summary of the Invention

[0005] This invention provides a method for identifying cracks in steel mesh based on deep learning, which helps to solve the problems mentioned in the background art.

[0006] This invention provides the following technical solution: a method for identifying cracks in steel mesh based on deep learning, comprising:

[0007] Initial operation steps:

[0008] Visual sensors and laser rangefinders are deployed at the inspection site. The visual sensors acquire digital images of the steel mesh, which are then processed by a pre-set deep learning network to obtain the set of coordinates of crack skeleton points and the set of coordinates of steel mesh intersection nodes.

[0009] The standard physical side length of the reinforcing mesh is obtained by measuring with a laser rangefinder.

[0010] Connect to the historical crack database to obtain the retrieval key values ​​of historical crack features;

[0011] Invariant extraction and primary key generation steps:

[0012] Based on the spatial topological distance and stress distribution relationship between the crack skeleton point and the mesh intersection node, structural invariants that suppress local extraction noise are extracted, and combined with the standard physical side length and global coordinates, a database retrieval primary key is generated for each crack skeleton point.

[0013] Database comparison and retrieval steps:

[0014] Calculate the matching similarity score between the retrieval primary key and the retrieval key value of the historical crack feature, and adaptively generate a dynamic retrieval tolerance threshold based on the structural invariant. Constrain the matching similarity score and the dynamic retrieval tolerance threshold. If the threshold requirement is met, it is determined to be a homologous damage feature, and the database matching retrieval is completed.

[0015] Optionally, in the invariant extraction and primary key generation step, the calculation of the crack node topological distance matrix is ​​carried out in the following steps:

[0016] Using the coordinates of each crack skeleton point and the coordinates of each steel mesh intersection node output in the initial operation steps, calculate the difference between the two on the horizontal axis and the difference on the vertical axis respectively.

[0017] Square the difference on the horizontal axis and the difference on the vertical axis respectively;

[0018] The results of squaring the two values ​​are summed, and the sum is squared to obtain the topological distance between each crack skeleton point and each rebar mesh intersection node.

[0019] Optionally, after obtaining the topological distance matrix, the lateral offset rate is introduced to calculate the directional stress correlation factor. The specific steps are as follows:

[0020] Calculate the absolute value of the difference in the x-coordinate between a specific crack skeleton point and a specific intersection node of the rebar mesh;

[0021] Divide the absolute value by the standard physical side length of the steel mesh obtained in the initial operation step to obtain the offset ratio;

[0022] Add the offset ratio to the numerical value to obtain the adjustment coefficient;

[0023] The calculated topological distance is multiplied by the adjustment coefficient to obtain the directional stress correlation factor.

[0024] Optionally, after obtaining the directional stress correlation factor, the structural index dispersion coefficient is calculated, and the specific steps are as follows:

[0025] For a specific crack skeleton point and a specific rebar mesh intersection node, the topological distance between the two is divided by the standard physical side length of the rebar mesh, and the negative of the ratio is taken as the exponent. The exponentiation is performed with the base of the natural logarithm as the base to obtain the distance attenuation term.

[0026] Multiply the directional stress correlation factor between the two by the distance attenuation term to obtain the weighted correlation value;

[0027] For a specific crack skeleton point, the weighted correlation values ​​calculated by the point and all nodes in the set of coordinates of the intersection nodes of the steel mesh are summed to obtain the structural index discrete coefficient of the specific crack skeleton point.

[0028] Optionally, after obtaining the discrete coefficients of the structure index, the structure invariants are calculated, and the specific steps are as follows:

[0029] The structural index discrete coefficients of all skeleton points in the set of crack skeleton point coordinates are summed to obtain the global discrete coefficient sum.

[0030] Multiply the structural index discrete coefficient of a specific crack skeleton point by the total number of skeleton points to obtain the magnified feature value of that specific point.

[0031] Dividing the amplified feature value by the sum of the global discrete coefficients yields the structural invariant of the specific crack skeleton point.

[0032] Optionally, after calculating the structural invariants, the database retrieval primary key is generated, and the specific steps are as follows:

[0033] Calculate the square of the x-coordinate and the square of the y-coordinate of a specific crack skeleton point, and then sum the two.

[0034] The square root of the summation result is used to obtain the absolute positional distance;

[0035] Divide the absolute position distance by the standard physical side length of the steel mesh to obtain the normalized position parameters;

[0036] The structural invariant of the specific crack skeleton point is multiplied by the normalized location parameter to obtain the generated database retrieval primary key.

[0037] Optionally, in the database comparison and retrieval step, the matching similarity score is calculated, specifically through the following steps:

[0038] Calculate the difference between the generated database retrieval primary key and the obtained historical feature retrieval key value, and take the absolute value of the difference;

[0039] Using the negative number of the extracted structural invariant as the exponent, and performing exponential operation with the base of the natural logarithm as the base, the characteristic decay factor is obtained;

[0040] Multiplying the absolute value by the feature decay factor yields the matching similarity score residual.

[0041] Optionally, to generate the dynamic retrieval tolerance threshold, after obtaining the structural invariants, the intermediate statistics of the global feature distribution are calculated. The specific steps are as follows:

[0042] From the calculated topological distance matrix, extract the maximum topological distance with the largest value;

[0043] Summ the structural invariants at all crack skeleton points;

[0044] Divide the summation result by the total number of crack skeleton points to obtain the global average structural invariant.

[0045] Optionally, the calculation of the dynamic retrieval tolerance threshold is performed, and the specific steps are as follows:

[0046] Divide the standard physical side length by the extracted maximum topological distance to obtain the scale mapping ratio.

[0047] The calculated global average structural invariant is multiplied by the scale mapping ratio to obtain the generated dynamic retrieval tolerance threshold.

[0048] The present invention has the following beneficial effects:

[0049] 1. This technical solution primarily targets the specific application environment of identifying and tracing the historical evolution of cracks in reinforced steel mesh under stress during structural health monitoring in buildings and engineering projects. The solution transforms isolated crack coordinates extracted by visual sensors into structural invariants and retrieval keys containing stress characteristics, combined with the physical boundaries and topological distances of the reinforced steel mesh. It then adaptively generates dynamic tolerance thresholds based on the mesh size, ultimately performing similarity comparisons in a historical database. This mechanism was designed because simple two-dimensional image coordinates, detached from the underlying stress environment, would face severe bottlenecks when directly used for retrieval in massive databases: dense crack coordinates easily trigger database indexing. Distribution skew and retrieval congestion; unavoidable local noise in the pre-extraction of deep learning can cause drastic fluctuations in the retrieval features of the same crack in different detection cycles; moreover, the physical scale of various steel meshes varies greatly in reality, and fixed comparison standards cannot be universally applied; under this specific environment, this solution can reconstruct the fragile image recognition results into stable and noise-resistant geometric topological scalars, avoiding the risk of retrieval paralysis caused by jitter in the underlying feature extraction and data density; at the same time, it endows the retrieval system with the dynamic evaluation capability of adaptive physical scale, thereby achieving efficient and high-precision cross-cycle source tracing and matching of historical damage data under complex detection environments and variable mesh specifications;

[0050] 2. By using the method of summing the squares of the differences between the horizontal and vertical coordinates and taking the square root, the spatial topological distance matrix between each crack skeleton point and each rebar mesh intersection node output by the deep learning network is accurately calculated. This successfully eliminates the lack of structural semantics in the coordinates of a single crack image, forcibly binding the originally isolated crack pixels to the geometric topological network of the surrounding solid rebar skeleton. This makes the subsequent feature representation no longer limited to the surface morphology of the crack itself, but can profoundly reflect the basic spatial distribution logic of the damage location relative to the overall stress nodes, thereby greatly improving the feature extraction dimension and effectively overcoming the feature matching disorder problem that traditional image comparison technology is prone to when facing complex mesh backgrounds due to the lack of spatial reference.

[0051] 3. By multiplying the pure spatial topological distance obtained from the previous calculation with an adjustment coefficient that includes the lateral offset rate, a directional stress correlation factor that can comprehensively reflect the stress propagation trend is constructed, realizing the deep integration of basic geometric features and structural stress features. This factor not only inherits the distance of the crack point from the core intersection node, but also amplifies the specific differences in deviation from the principal stress transmission direction. This allows the final output correlation factor to accurately distinguish between longitudinal cracks parallel to the direction of the reinforcing bars and shear cracks perpendicular to the direction of stress. This effectively solves the problem of misjudging structural damage with similar appearance but completely different causes due to the lack of stress direction constraints in historical database retrieval, and significantly improves the physical rigor of damage identification.

[0052] 4. By multiplying the directional stress correlation factor with the distance attenuation term to obtain the weighted correlation value, and by summing the weighted correlation values ​​across all intersection nodes for a specific crack skeleton point, the structural index discrete coefficient that can characterize the spatial independence of nodes is successfully calculated. This process forcibly projects the dense local crack coordinates into the attenuation influence domain of the global nodes for dispersal and recombination, effectively suppressing the feature homogenization phenomenon caused by the extremely close spatial positions of adjacent crack points, eliminating the distribution skew and hash collisions caused by the simultaneous influx of massive highly similar features into the historical database, and ensuring the smooth database comparison and stable system operation in high-concurrency retrieval scenarios.

[0053] 5. By summarizing all discrete coefficients in the set of crack skeleton point coordinates to obtain the global sum, and multiplying the discrete coefficient of a single skeleton point by the total number to obtain the amplified feature value, and dividing the global sum, a highly robust structural invariant was successfully extracted. This operation can effectively smooth the random jitter noise that inevitably occurs when deep learning models extract edge pixels in environments with drastic changes in lighting or cluttered backgrounds. Even if there is a local pixel shift in the same crack coordinates extracted in different batches, its relative proportion feature in the global network remains absolutely constant. This completely overcomes the technical barrier of high-frequency oscillation of query key values ​​and frequent breaks in historical matching tracing caused by local extraction noise.

[0054] 6. By performing square summation and square root operations on the horizontal and vertical coordinates of the crack skeleton points respectively to obtain the absolute position distance, and dividing it by the standard physical side length of the steel mesh, a normalized position parameter that completely eliminates size dependence is obtained. Finally, the structural invariant is multiplied by this normalized parameter to generate a unique database retrieval primary key. This successfully completes the deep dimensionality reduction and cross-boundary transformation from multidimensional geometric coordinates to one-dimensional standard scalar key values. This retrieval primary key highly condenses the stress invariant characteristics at the microscale and the absolute spatial addressing information at the macroscale, enabling the massive and complex structural damage data to be seamlessly integrated into the existing general database index architecture. This greatly reduces the memory consumption and computing power consumption when comparing massive heterogeneous data, and achieves fast and accurate positioning.

[0055] 7. By calculating the absolute difference between the currently generated retrieval primary key and the inherent feature key value in the historical crack database, and constructing a feature decay factor with the negative number of the structural invariant as the exponent, the two are multiplied to obtain the final matching similarity score residual. This breaks the limitation of the traditional algorithm that relies solely on a single distance residual for rigid comparison. When evaluating the similarity between two cracks, the system can adaptively adjust the tolerance according to the singularity of the current crack's own structural features. When encountering extremely obvious and unusual local damage, the decay factor will actively tighten the similarity tolerance boundary, thereby effectively preventing conventional heterogeneous damage from being misjudged as homogeneous evolutionary features under complex environmental interference, and improving the database retrieval identification ability.

[0056] 8. By extracting the global average structural invariant and multiplying it with the scale mapping ratio formed by the standard physical side length and the maximum topological distance, a dynamic retrieval tolerance threshold directly controlled by the current engineering scale was successfully generated adaptively. This threshold calculation mechanism reverses the problem of fixed thresholds being incompatible with cross-scale grids. Whether the detection object is a micro precast slab with fine aperture or a heavy bridge frame with a huge span, the system can automatically adjust the judgment scale according to the maximum topological tensile range. In large grids, the matching boundary is automatically relaxed to avoid serious missed detections, and in small grids, the scale is automatically tightened to eliminate massive false alarms. This provides a dynamic matching constraint boundary for big data crack evolution analysis across engineering types. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the basic process of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Example 1, refer to Figure 1 Methods for identifying cracks in steel mesh based on deep learning include:

[0060] Initial operation steps:

[0061] Visual sensors and laser rangefinders are deployed at the inspection site. The visual sensors acquire digital images of the rebar mesh, which are then processed by a pre-set deep learning network to obtain a set of coordinates for crack skeleton points and rebar mesh intersection nodes. Crack skeleton points refer to the set of pixels on the central axis of cracks in the rebar mesh. In this technical solution, a deep learning model first extracts the macroscopic connected region pixels of the crack. Then, a morphological thinning algorithm (e.g., the Zhang-Suen skeleton extraction algorithm) is used to remove the width features of the crack, retaining only the central connected line with a width of a single pixel. Each pixel coordinate on this connected line is defined as a crack skeleton point. Rebar mesh intersection nodes refer to the structural intersection center points formed by the orthogonal or cross-tying of longitudinal reinforcing bars and transverse distributed reinforcing bars in industrial building rebar mesh. In this technical solution, the intersection positions of the rebar skeleton in the image are first identified, and the geometric center pixel coordinates of the image of the intersection physical region are calculated. The set of these center pixel coordinates is defined as a rebar mesh node.

[0062] The standard physical side length of the reinforcing mesh is obtained by measuring with a laser rangefinder.

[0063] Connect to the historical crack database to obtain the retrieval key values ​​of historical crack features;

[0064] Invariant extraction and primary key generation steps:

[0065] Based on the spatial topological distance and stress distribution relationship between the crack skeleton point and the mesh intersection node, structural invariants that suppress local extraction noise are extracted, and combined with the standard physical side length and global coordinates, a database retrieval primary key is generated for each crack skeleton point.

[0066] Database comparison and retrieval steps:

[0067] Calculate the matching similarity score between the retrieval primary key and the retrieval key value of the historical crack feature, and adaptively generate a dynamic retrieval tolerance threshold based on the structural invariant. Constrain the matching similarity score and the dynamic retrieval tolerance threshold. If the threshold requirement is met, it is determined to be a homologous damage feature, and the database matching retrieval is completed.

[0068] In the invariant extraction and primary key generation steps, the topological distance matrix of the crack nodes is calculated, and the specific steps are as follows:

[0069] Using the coordinates of each crack skeleton point and the coordinates of each steel mesh intersection node output in the initial operation steps, calculate the difference between the two on the horizontal axis and the difference on the vertical axis respectively.

[0070] Square the difference on the horizontal axis and the difference on the vertical axis respectively;

[0071] The results of squaring the two values ​​are summed, and the sum is squared to obtain the topological distance between each crack skeleton point and each rebar mesh intersection node.

[0072] After obtaining the topological distance matrix, the lateral offset rate is introduced to calculate the directional stress correlation factor. The specific steps are as follows:

[0073] Calculate the absolute value of the difference in the x-coordinate between a specific crack skeleton point and a specific intersection node of the rebar mesh;

[0074] Divide the absolute value by the standard physical side length of the steel mesh obtained in the initial operation step to obtain the offset ratio;

[0075] Add the offset ratio to the numerical value to obtain the adjustment coefficient;

[0076] The calculated topological distance is multiplied by the adjustment coefficient to obtain the directional stress correlation factor.

[0077] After obtaining the directional stress correlation factor, the structural index dispersion coefficient is calculated. The specific steps are as follows:

[0078] For a specific crack skeleton point and a specific rebar mesh intersection node, the topological distance between the two is divided by the standard physical side length of the rebar mesh, and the negative of the ratio is taken as the exponent. The exponentiation is performed with the base of the natural logarithm as the base to obtain the distance attenuation term.

[0079] Multiply the directional stress correlation factor between the two by the distance attenuation term to obtain the weighted correlation value;

[0080] For a specific crack skeleton point, the weighted correlation values ​​calculated by the point and all nodes in the set of coordinates of the intersection nodes of the steel mesh are summed to obtain the structural index discrete coefficient of the specific crack skeleton point.

[0081] After obtaining the discrete coefficients of the structure index, the structure invariants are calculated. The specific steps are as follows:

[0082] The structural index discrete coefficients of all skeleton points in the set of crack skeleton point coordinates are summed to obtain the global discrete coefficient sum.

[0083] Multiply the structural index discrete coefficient of a specific crack skeleton point by the total number of skeleton points to obtain the magnified feature value of that specific point.

[0084] Dividing the amplified feature value by the sum of the global discrete coefficients yields the structural invariant of the specific crack skeleton point.

[0085] After calculating the structural invariants, the database retrieval primary key is generated. The specific steps are as follows:

[0086] Calculate the square of the x-coordinate and the square of the y-coordinate of a specific crack skeleton point, and then sum the two.

[0087] The square root of the summation result is used to obtain the absolute positional distance;

[0088] Divide the absolute position distance by the standard physical side length of the steel mesh to obtain the normalized position parameters;

[0089] The structural invariant of the specific crack skeleton point is multiplied by the normalized location parameter to obtain the generated database retrieval primary key.

[0090] In the database comparison and retrieval step, the matching similarity score is calculated, and the specific steps are as follows:

[0091] Calculate the difference between the generated database retrieval primary key and the obtained historical feature retrieval key value, and take the absolute value of the difference;

[0092] Using the negative number of the extracted structural invariant as the exponent, and performing exponential operation with the base of the natural logarithm as the base, the characteristic decay factor is obtained;

[0093] Multiplying the absolute value by the feature decay factor yields the matching similarity score residual.

[0094] To generate the dynamic retrieval tolerance threshold, after obtaining the structural invariants, the intermediate statistics of the global feature distribution are calculated. The specific steps are as follows:

[0095] From the calculated topological distance matrix, extract the maximum topological distance with the largest value;

[0096] Summ the structural invariants at all crack skeleton points;

[0097] Divide the summation result by the total number of crack skeleton points to obtain the global average structural invariant.

[0098] The specific steps to complete the calculation of the dynamic retrieval tolerance threshold are as follows:

[0099] Divide the standard physical side length by the extracted maximum topological distance to obtain the scale mapping ratio.

[0100] The calculated global average structural invariant is multiplied by the scale mapping ratio to obtain the generated dynamic retrieval tolerance threshold. This technical solution is mainly aimed at the specific application environment of identifying and tracing the historical evolution of cracks in reinforced steel mesh under stress in building and engineering structural health monitoring. The solution transforms the isolated crack coordinates extracted by the visual sensor into structural invariants and retrieval keys containing stress features by combining the physical boundary and topological distance of the reinforced steel mesh. A dynamic tolerance threshold is adaptively generated based on the mesh size, and finally, similarity comparison is performed in the historical database. This mechanism is designed because simple two-dimensional image coordinates, detached from the underlying stress environment, face serious bottlenecks when directly used for massive database retrieval: dense crack coordinates easily cause database indexing issues. The problem stems from several factors: First, the retrieval process suffers from inconsistencies in the mesh layout and retrieval congestion. Second, the unavoidable local noise in the pre-processing of deep learning can cause significant fluctuations in the retrieval features of the same crack across different detection cycles. Third, the vastly different physical scales of various steel meshes in reality make fixed comparison standards unsuitable. Under these specific conditions, this solution can reconstruct fragile image recognition results into stable, noise-resistant geometric topological scalars, mitigating the risk of retrieval paralysis caused by jitter in low-level feature extraction and data density. Simultaneously, it endows the retrieval system with the ability to dynamically evaluate physical scales, thereby achieving efficient and high-precision cross-cycle source matching of historical damage data in complex detection environments and with varying mesh specifications.

[0101] Example 2, the method for identifying cracks in steel mesh based on deep learning, further includes:

[0102] In the invariant extraction and primary key generation steps, to establish the basic spatial topological relationship between crack points and rebar mesh nodes, the crack node topological distance matrix is ​​calculated, where the topological distance is calculated as follows:

[0103] In the formula: Indicates the first The crack skeleton point and the first Topological distance between individual steel mesh nodes; The output of the initial operation step is respectively the first The horizontal and vertical coordinates of the crack skeleton points; The output of the initial operation step is respectively the first The horizontal and vertical coordinates of each rebar mesh node are calculated. By summing the squares of the differences between the horizontal and vertical coordinates and taking the square root, the spatial topological distance matrix between each crack skeleton point and each rebar mesh intersection node output by the deep learning network is accurately calculated. This successfully eliminates the lack of structural semantics in the coordinates of a single crack image, forcibly binding the originally isolated crack pixels to the geometric topological network of the surrounding solid rebar skeleton. This makes the subsequent feature representation no longer limited to the surface morphology of the crack itself, but can profoundly reflect the basic spatial distribution logic of the damage location relative to the overall stress nodes, thereby greatly improving the feature extraction dimension and effectively overcoming the feature matching error problem that traditional image comparison technology is prone to when facing complex mesh backgrounds due to the lack of spatial reference.

[0104] After obtaining the topological distance matrix, to correct the inadequacy of pure geometric distance in reflecting stress transmission, a lateral offset rate is introduced to calculate the directional stress correlation factor, specifically as follows:

[0105] In the formula: Indicates the first The crack point and the first A directional correlation factor between individual grid nodes, used to characterize the stress propagation trend; The calculated topological distance; The first The first crack skeleton point and the first The x-coordinate of each grid node; The standard physical side length of the steel mesh measured in the initial operation steps is used. By multiplying the pure spatial topological distance obtained from the previous calculation with an adjustment coefficient that includes the lateral offset rate, a directional stress correlation factor that can comprehensively reflect the stress propagation trend is constructed. This achieves a deep integration of the basic geometric features and the structural stress features. This factor not only inherits the degree of distance between the crack point and the core intersection node, but also amplifies the specific differences in deviation from the principal stress transmission direction. This allows the final output correlation factor to accurately distinguish between longitudinal cracks parallel to the steel bar direction and shear cracks perpendicular to the stress direction. This effectively solves the problem of misjudging structural damage with similar appearance but completely different causes due to the lack of stress direction constraints in historical database retrieval, and significantly improves the physical rigor of damage identification.

[0106] After obtaining the directional stress correlation factor, in order to suppress the skewness of the distribution of search features in the database index tree, the structural index dispersion coefficient is calculated, specifically as follows: ;

[0107] In the formula: Indicates the first The structural index of the crack skeleton points is the discrete coefficient; The total number of nodes in the set of coordinates of the intersecting nodes of the steel mesh obtained in the initial operation steps; The calculated directional stress correlation factor; The calculated topological distance; The standard physical side length of the reinforcing mesh measured in the initial operation step; The base of the natural logarithm is used. By multiplying the directional stress correlation factor with the distance attenuation term to obtain the weighted correlation value, and then summing the weighted correlation values ​​for all intersection nodes for a specific crack skeleton point, the structural index discrete coefficient that can characterize the spatial independence of the nodes is successfully calculated. This process forcibly projects the dense local crack coordinates into the attenuation influence domain of the global nodes for dispersal and recombination, effectively suppressing the feature homogenization phenomenon caused by the extremely close spatial positions of adjacent crack points, eliminating the distribution skew and hash collisions caused by the simultaneous influx of massive highly similar features into the historical database, and ensuring the smooth database comparison and stable system operation in high-concurrency retrieval scenarios.

[0108] After obtaining the discrete coefficients of the structure index, in order to eliminate the interference of local noise during the preliminary extraction on the retrieval, the structure invariant is calculated, specifically as follows:

[0109] In the formula: Indicates the first The structural invariants of each crack skeleton point; For the calculated first The structural index discrete coefficients of each point; For the first The structural index discrete coefficients of each point For traversal index; The total number of skeleton points in the set of crack skeleton point coordinates obtained in the initial operation step. By summing all the discrete coefficients in the set of crack skeleton point coordinates to obtain the global sum, and multiplying the discrete coefficient of a single skeleton point by the total number to obtain the amplified feature value, the structure invariant with strong anti-interference ability is successfully extracted. This operation can effectively smooth the random jitter noise that inevitably occurs when the deep learning model extracts edge pixels in environments with drastic changes in lighting or cluttered backgrounds. Even if there is a local pixel shift in the same crack coordinates extracted in different batches, its relative proportion feature in the global network remains absolutely constant, thus completely overcoming the technical barrier of high-frequency oscillation of query key values ​​and frequent breakage of historical matching tracing caused by local extraction noise.

[0110] After calculating the structural invariants, they are converted into standard scalar key values ​​that can be directly used for database indexing, generating the database retrieval primary key. The specific calculation is as follows:

[0111] In the formula: Represented as the first The primary key for database retrieval is generated from the crack skeleton points; These are the structural invariants obtained from the calculation; The first The horizontal and vertical coordinates of the crack skeleton points; The standard physical side length is given. The absolute positional distance is obtained by summing the squares and taking the square root of the horizontal and vertical coordinates of the crack skeleton points, respectively. This distance is then divided by the standard physical side length of the reinforcing mesh to obtain a normalized positional parameter that completely eliminates size dependence. Finally, the structural invariant is multiplied by this normalized parameter to generate a unique database retrieval key. This successfully completes a deep dimensionality reduction and cross-boundary transformation from multidimensional geometric coordinates to one-dimensional standard scalar key values. This retrieval key highly condenses the stress invariant characteristics at the microscale and the absolute spatial addressing information at the macroscale, enabling the seamless integration of massive and complex structural damage data into existing general database index architectures. This significantly reduces memory usage and computational consumption when comparing massive amounts of heterogeneous data, achieving rapid and accurate positioning.

[0112] In the database comparison and retrieval step, the matching similarity score is calculated as follows:

[0113] In the formula: Indicates the first The characteristics of the crack are compared with those in the database. The residual of the matching similarity score for each historical crack feature; This is the calculated primary key for database retrieval; The first one obtained in the initial operation step Historical feature retrieval key value; For extracted structural invariants; The base of the natural logarithm is used. By calculating the absolute difference between the currently generated retrieval primary key and the inherent feature key value in the historical crack database, and constructing a feature decay factor with the negative number of the structural invariant as the exponent, the two are multiplied to obtain the final matching similarity score residual. This breaks the limitation of traditional algorithms that rely solely on a single distance residual for rigid comparison. When evaluating the similarity between two cracks, the system can adaptively adjust the tolerance based on the singularity of the current crack's own structural features. When encountering extremely obvious and unusual local damage, the decay factor will actively tighten the similarity tolerance boundary, thereby effectively preventing conventional heterogeneous damage from being misjudged as homogeneous evolutionary features under complex environmental interference, and improving the identification ability of database retrieval.

[0114] To generate the dynamic retrieval tolerance threshold, after obtaining the structural invariants, the intermediate statistics of the global feature distribution are calculated, specifically as follows:

[0115] Extract the maximum topological distance from the topological distance matrix. ;

[0116] Calculate the global average structural invariants of the current crack: ;

[0117] In the formula: To obtain the global average structure invariants; The total number of points in the crack skeleton; These are the structural invariants obtained from the calculation.

[0118] After obtaining the intermediate statistics of the global feature distribution, the dynamic retrieval tolerance threshold is calculated, specifically as follows:

[0119] In the formula: This indicates the dynamic retrieval tolerance threshold; The calculated global average structure invariants; The standard physical side length; The maximum topological distance is extracted. By extracting the global average structural invariant and multiplying it by the scale mapping ratio formed by the standard physical side length and the maximum topological distance, a dynamic retrieval tolerance threshold directly controlled by the current engineering scale is successfully generated adaptively. This threshold calculation mechanism reverses the problem of fixed thresholds being incompatible with cross-scale grids. Whether the detection object is a micro precast slab with fine aperture or a heavy bridge frame with a huge span, the system can automatically adjust the judgment scale according to the maximum topological tensile range. In large grids, it automatically relaxes the matching boundary to avoid serious missed detections, and in small grids, it automatically tightens the review scale to eliminate massive false alarms. Thus, it provides a dynamic matching constraint boundary for big data crack evolution analysis across engineering types.

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0121] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying cracks in reinforcing mesh based on deep learning, characterized in that, include: Initial operation steps: Visual sensors and laser rangefinders are deployed at the inspection site. The visual sensors acquire digital images of the steel mesh, which are then processed by a pre-set deep learning network to obtain the set of coordinates of crack skeleton points and the set of coordinates of steel mesh intersection nodes. The standard physical side length of the reinforcing mesh is obtained by measuring with a laser rangefinder. Connect to the historical crack database to obtain the retrieval key values ​​of historical crack features; Invariant extraction and primary key generation steps: Based on the spatial topological distance and stress distribution relationship between the crack skeleton point and the mesh intersection node, structural invariants that suppress local extraction noise are extracted, and combined with the standard physical side length and global coordinates, a database retrieval primary key is generated for each crack skeleton point. Database comparison and retrieval steps: Calculate the matching similarity score between the retrieval primary key and the retrieval key value of the historical crack feature, and adaptively generate a dynamic retrieval tolerance threshold based on the structural invariant. Constrain the matching similarity score and the dynamic retrieval tolerance threshold. If the threshold requirement is met, it is determined to be a homologous damage feature, and the database matching retrieval is completed.

2. The method for identifying cracks in reinforcing mesh based on deep learning according to claim 1, characterized in that, In the invariant extraction and primary key generation steps, the topological distance matrix of the crack nodes is calculated, and the specific steps are as follows: Using the coordinates of each crack skeleton point and the coordinates of each steel mesh intersection node output in the initial operation steps, calculate the difference between the two on the horizontal axis and the difference on the vertical axis respectively. Square the difference on the horizontal axis and the difference on the vertical axis respectively; The results of squaring the two values ​​are summed, and the sum is squared to obtain the topological distance between each crack skeleton point and each rebar mesh intersection node.

3. The method for identifying cracks in reinforcing mesh based on deep learning according to claim 2, characterized in that, After obtaining the topological distance matrix, the lateral offset rate is introduced to calculate the directional stress correlation factor. The specific steps are as follows: Calculate the absolute value of the difference in the x-coordinate between a specific crack skeleton point and a specific intersection node of the rebar mesh; Divide the absolute value by the standard physical side length of the steel mesh obtained in the initial operation step to obtain the offset ratio; Add the offset ratio to the numerical value to obtain the adjustment coefficient; The calculated topological distance is multiplied by the adjustment coefficient to obtain the directional stress correlation factor.

4. The method for identifying cracks in reinforcing mesh based on deep learning according to claim 3, characterized in that, After obtaining the directional stress correlation factor, the structural index dispersion coefficient is calculated. The specific steps are as follows: For a specific crack skeleton point and a specific rebar mesh intersection node, the topological distance between the two is divided by the standard physical side length of the rebar mesh, and the negative of the ratio is taken as the exponent. The exponentiation is performed with the base of the natural logarithm as the base to obtain the distance attenuation term. Multiply the directional stress correlation factor between the two by the distance attenuation term to obtain the weighted correlation value; For a specific crack skeleton point, the weighted correlation values ​​calculated by the point and all nodes in the set of coordinates of the intersection nodes of the steel mesh are summed to obtain the structural index discrete coefficient of the specific crack skeleton point.

5. The method for identifying cracks in reinforcing mesh based on deep learning according to claim 4, characterized in that, After obtaining the discrete coefficients of the structure index, the structure invariants are calculated. The specific steps are as follows: The structural index discrete coefficients of all skeleton points in the set of crack skeleton point coordinates are summed to obtain the global discrete coefficient sum. Multiply the structural index discrete coefficient of a specific crack skeleton point by the total number of skeleton points to obtain the magnified feature value of that specific point. Dividing the amplified feature value by the sum of the global discrete coefficients yields the structural invariant of the specific crack skeleton point.

6. The method for identifying cracks in reinforcing mesh based on deep learning according to claim 5, characterized in that, After calculating the structural invariants, the database retrieval primary key is generated. The specific steps are as follows: Calculate the square of the x-coordinate and the square of the y-coordinate of a specific crack skeleton point, and then sum the two. The square root of the summation result is used to obtain the absolute positional distance; Divide the absolute position distance by the standard physical side length of the steel mesh to obtain the normalized position parameters; The structural invariant of the specific crack skeleton point is multiplied by the normalized location parameter to obtain the generated database retrieval primary key.

7. The method for identifying cracks in reinforcing mesh based on deep learning according to claim 6, characterized in that, In the database comparison and retrieval step, the matching similarity score is calculated, and the specific steps are as follows: Calculate the difference between the generated database retrieval primary key and the obtained historical feature retrieval key value, and take the absolute value of the difference; Using the negative number of the extracted structural invariant as the exponent, and performing exponential operation with the base of the natural logarithm as the base, the characteristic decay factor is obtained; Multiplying the absolute value by the feature decay factor yields the matching similarity score residual.

8. The method for identifying cracks in reinforcing mesh based on deep learning according to claim 5, characterized in that, To generate the dynamic retrieval tolerance threshold, after obtaining the structural invariants, the intermediate statistics of the global feature distribution are calculated. The specific steps are as follows: From the calculated topological distance matrix, extract the maximum topological distance with the largest value; Summ the structural invariants at all crack skeleton points; Divide the summation result by the total number of crack skeleton points to obtain the global average structural invariant.

9. The method for identifying cracks in reinforcing mesh based on deep learning according to claim 8, characterized in that, The specific steps to complete the calculation of the dynamic retrieval tolerance threshold are as follows: Divide the standard physical side length by the extracted maximum topological distance to obtain the scale mapping ratio. The calculated global average structural invariant is multiplied by the scale mapping ratio to obtain the generated dynamic retrieval tolerance threshold.

Images