Concrete bridge damage detection and residual life prediction method based on machine vision

Abstract

The invention provides a concrete bridge damage detection and residual life prediction method based on machine vision. The method comprises the following steps: acquiring a concrete bridge surface image and a corresponding structure coordinate thereof; the structure coordinates are explicitly embedded into the concrete bridge surface image, and a coordinate surface image is obtained; the coordinate surface image is input into an image recognition model, image types and corresponding structure coordinates are recognized, the image types comprise an extrinsic damage type image and a mark type image, and the mark type image is a mark image generated by manual testing; determining each first life influence parameter according to the external display damage image and the corresponding structure coordinate; determining each second life influence parameter according to each mark type image and the corresponding structure coordinate; and estimating the remaining life of the concrete bridge according to each first life influence parameter and each second life influence parameter. By implementing the method, the comprehensiveness of concrete bridge damage detection and the accuracy of residual life prediction can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of concrete bridge life prediction, and in particular relates to a method for concrete bridge damage detection and remaining life prediction based on machine vision. Background Art

[0002] Remaining life assessment of concrete bridges is crucial for ensuring their safe operation. Traditional bridge inspection and life assessment methods suffer from numerous shortcomings. Traditional manual inspection methods suffer from subjectivity, numerous blind spots, and data discretization, making them inadequate for the precise inspection of long and large concrete bridges. While existing computer vision-based inspection technologies have made some progress in damage identification, life prediction methods often rely on a single, overt image of damage, such as concrete cracking, steel corrosion, cracks, spalling, and exposed rebar, due to the limitations of computer vision. These images only reflect the extent of the visible damage. Using only a single type of image information to predict remaining life can lead to significant deviations in remaining life assessment. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a method for concrete bridge damage detection and remaining life prediction based on machine vision to meet the demand for improving the accuracy of the remaining life of concrete bridges.

[0004] In order to achieve the above object, the present invention provides the following technical solutions:

[0005] According to a first aspect, the present invention provides a method for damage detection and remaining life prediction of concrete bridges based on machine vision, comprising: obtaining a surface image of a concrete bridge and its corresponding structural coordinates; explicitly embedding the structural coordinates into the surface image of the concrete bridge to obtain a coordinated surface image; inputting the coordinated surface image into an image recognition model to identify the image type and the corresponding structural coordinates, the image type including explicit damage images and marked images, wherein the marked images are marked images generated by artificial testing; determining each first life influencing parameter based on each explicit damage image and the corresponding structural coordinates; determining each second life influencing parameter based on each marked image and the corresponding structural coordinates; and estimating the remaining life of the concrete bridge based on each first life influencing parameter and each second life influencing parameter.

[0006] According to a second aspect, the present invention provides a device for detecting damage and predicting remaining life of a concrete bridge based on machine vision, comprising: an acquisition module for acquiring a surface image of a concrete bridge and its corresponding structural coordinates; a coordinate module for embedding the structural coordinates into the surface image of the concrete bridge to obtain a coordinated surface image; an identification module for inputting the coordinated surface image into an image recognition model to identify the image type and the corresponding structural coordinates, wherein the image type includes an explicit damage image and a marked image, wherein the marked image is a marked image generated by artificial testing; a first life influencing parameter determination module for determining each first life influencing parameter based on each explicit damage image and the corresponding structural coordinates; a second life influencing parameter determination module for determining each second life influencing parameter based on each marked image and the corresponding structural coordinates; and a remaining life determination module for estimating the remaining life of the concrete bridge based on each first life influencing parameter and each second life influencing parameter.

[0007] According to a third aspect, an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the steps of the machine vision-based method for concrete bridge damage detection and remaining life prediction described in the first aspect or any embodiment of the first aspect.

[0008] According to a fourth aspect, an embodiment of the present invention provides a computer storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method for machine vision-based concrete bridge damage detection and remaining life prediction as described in the first aspect or any embodiment of the first aspect.

[0009] An embodiment of the present invention provides a method for concrete bridge damage detection and remaining life prediction based on machine vision. Compared with existing computer vision detection technologies that rely solely on external damage images, this method introduces labeled images. By analyzing the labeled images, second life influencing parameters (such as concrete carbonization depth and steel corrosion rate) are obtained. Combined with the first life influencing parameters determined by the external damage images (such as cracks, spalling and other geometric features), a complete life assessment system is constructed from the two dimensions of apparent damage and material degradation mechanism. This multi-source data fusion method can not only capture visible surface damage, but also deeply analyze the degradation process inside the material, making up for the defect of traditional methods that only focus on external features, and improving the comprehensiveness of concrete bridge damage detection and the accuracy of remaining life prediction.

[0010] Other advantages, objectives and features of the present invention will be described in the following description and will be apparent to those skilled in the art to some extent, or those skilled in the art can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0012] Figure 1 This is a specific example flow chart of a method for concrete bridge damage detection and remaining life prediction based on machine vision in the present invention;

[0013] Figure 2 A flowchart of a specific example of comparing a digital model of a concrete bridge with a standard model of a concrete bridge to determine multiple structural differences in the present invention;

[0014] Figure 3 This is a specific example flow chart of estimating the remaining life of a concrete bridge according to various first life-influencing parameters, various second life-influencing parameters, and a plurality of structural difference points in the present invention;

[0015] Figure 4 This is a principle block diagram of a specific example of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; internal connections between two components; wireless connections or wired connections. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0018] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0019] The embodiment of the present invention provides a method for concrete bridge damage detection and remaining life prediction based on machine vision, such as Figure 1 As shown, including:

[0020] S101, obtaining a surface image of a concrete bridge and its corresponding structural coordinates;

[0021] S102, embedding the structural coordinates into the concrete bridge surface image to obtain a coordinate surface image;

[0022] S103, inputting the coordinate surface image into an image recognition model to identify the image type and the corresponding structural coordinates. The image types include visible damage images and marked images. The marked images are marked images generated by manual testing.

[0023] S104, determining each first lifespan influencing parameter based on each external damage image and the corresponding structural coordinates;

[0024] S105, determining each second lifespan influencing parameter based on each marked image and the corresponding structure coordinates;

[0025] S106: Estimate the remaining life of the concrete bridge according to the first life influencing parameters and the second life influencing parameters.

[0026] For example, images of concrete bridge surfaces can be acquired using drones equipped with RTK modules. A drone equipped with a high-precision RTK-GPS module, a high-definition camera, and an IMU is selected. Using drone mapping software, the route is planned, and the bridge is tilted at multiple angles. Each time the shutter is triggered, the 3D coordinates (X, Y, Z) and attitude data (heading, pitch, and roll) at the moment of capture are simultaneously recorded. After acquisition, the image sequence is processed using Pix4Dmapper software. Feature point matching and bundle adjustment are used to generate a 3D point cloud of the bridge surface. A correspondence between image pixels and structural coordinates is established. Coordinate calibration is performed using ground control points to obtain the concrete surface image and the corresponding structural coordinates.

[0027] Next, the obtained concrete surface image and the corresponding structural coordinates are fused, that is, the structural coordinate information is integrated into the concrete surface image to construct a coordinate surface image. Specifically, the mapping table of image pixels and structural coordinates can be read, and the three-dimensional coordinates (X, Y, Z) corresponding to each target pixel point are superimposed on the non-critical area of ​​the image in the form of white semi-transparent text. The target pixel point can be a pixel determined according to a preset rule. For example, the image is divided into a grid of the target size, then the target pixel point is the pixel point of the four corners corresponding to the divided grid line. This embodiment does not limit the method of determining the target pixel, and those skilled in the art can determine it as needed.

[0028] The coordinateized surface image is segmented to produce multiple images. These segmented images are then fed into an image recognition model, which outputs the image types and coordinates within the segmented images, including visible damage images and labeled images. Visible damage images represent visible damage to the bridge, such as cracks and spalling, while labeled images represent artificially generated images, such as images showing the color change of a hole at a specific location when exposed to phenolphthalein. This image recognition model can be based on the YOLO model, with the output layer modified to include visible damage and labeled images. An optical character recognition (OCR) branch is added to the YOLO model to implement coordinate recognition of the coordinateized surface images. The YOLO model is constructed using a dataset of 5,000 bridge surface images, of which 3,000 contain visible damage, such as cracks and spalling, 1,000 contain various artificially labeled patterns, and 1,000 contain clean background images. Data augmentation operations, including random flipping, brightness adjustment, Gaussian blurring, and random cropping, are performed on the dataset to expand it to 20,000 samples. The training was performed on an NVIDIA RTX4090 GPU with a batch size of 16 for 300 epochs. The cosine annealing decay strategy was used, and the validation set loss was used as the early stopping basis.

[0029] For images of visible damage, geometric features of the damage, such as crack length, width, and spalling area, are extracted, along with spatial distribution features. The coordinates are then mapped to bridge structural locations (e.g., midspan, support, web, etc.), generating a multidimensional feature vector containing the damage type, geometric parameters, and location as the first life influencing parameter. For labeled images, relevant parameters and corresponding structural coordinates are extracted from the labeled images based on the test purpose, serving as the second life influencing parameter. Based on the first and second life influencing parameters, the remaining life of a concrete bridge can be estimated by combining the first and second life influencing parameters into a single data set. A prediction model is constructed using an XGBoost or LSTM neural network. The remaining life label is calculated by subtracting the bridge's design life from its service life, combined with expert assessments, and corrected. A sample is then constructed using data of the same type as the first and second life influencing parameters. This sample and corresponding remaining life label are then fed into the prediction model. Training is performed using the root mean square error (RMSE) as the error metric to obtain a trained prediction model. The combined data set of the first and second life influencing parameters is then fed into the prediction model to obtain the predicted remaining life.

[0030] An embodiment of the present invention provides a method for concrete bridge damage detection and remaining life prediction based on machine vision. Compared with existing computer vision detection technologies that rely solely on external damage images, this method introduces labeled images. By analyzing the labeled images, second life influencing parameters (such as concrete carbonization depth and steel corrosion rate) are obtained. Combined with the first life influencing parameters determined by the external damage images (such as cracks, spalling and other geometric features), a complete life assessment system is constructed from the two dimensions of apparent damage and material degradation mechanism. This multi-source data fusion method can not only capture visible surface damage, but also deeply analyze the degradation process inside the material, making up for the defect of traditional methods that only focus on external features, and improving the comprehensiveness of concrete bridge damage detection and the accuracy of remaining life prediction.

[0031] As an optional implementation, determining each first lifespan influencing parameter based on each external damage image and the corresponding structural coordinates includes:

[0032] Extracting geometric feature parameters of external damage images;

[0033] Determining whether there is an identified random speckle image at the structural coordinates of the explicit damage image, where the random speckle image is one of the marked images;

[0034] When there is a random speckle image, the mechanical parameters are analyzed based on the random speckle image;

[0035] The mechanical parameters, geometric characteristic parameters and corresponding structural coordinates of each external damage image at the structural coordinates are input into a pre-established first damage attention model for importance sorting to determine the key external damage image;

[0036] When there is no random speckle image, the geometric characteristic parameters and corresponding structural coordinates of each visible damage image are input into the pre-established second damage attention model for importance sorting to determine the key visible damage image;

[0037] The mechanical parameters, geometric characteristic parameters and damage location at the corresponding coordinates of the key external damage image are taken as the first life influencing parameters.

[0038] For example, for crack damage, an edge detection algorithm is used to extract the crack outline and calculate the crack length, width, area, and strike angle. For spalling damage, threshold segmentation is used to separate the spalling area and calculate parameters such as spalling area, perimeter, and depth. For exposed rebar damage, morphological operations are used to refine the rebar outline and calculate parameters such as exposed rebar length and corrosion area percentage. These calculated geometric feature parameters are stored in a data structure and associated with the structural coordinates (X, Y, Z) of the corresponding visible damage image.

[0039] Then, it is determined whether a recognized random speckle image exists at the structural coordinates of the explicit damage image. A random speckle image is a marker image with a randomly distributed speckle pattern that is artificially sprayed or printed on the surface of a concrete bridge. It is mainly used for non-contact measurement of structural deformation and mechanical parameters. Specifically, the random speckle image in the marker image is screened out from the output of the image recognition model, and its structural coordinates are extracted. For each structural coordinate of the explicit damage image, its Euclidean distance to the structural coordinates of all random speckle images is calculated. A distance threshold is set, such as 0.5 meters. If the distance between the coordinates of the random speckle image and the coordinates of the explicit damage image is less than the threshold, then the random speckle image is determined to exist at the structural coordinates of the explicit damage image; otherwise, it is determined not to exist.

[0040] When there is a random speckle image, the mechanical parameters are analyzed based on the random speckle image. Specifically, the random speckle image is first subjected to sub-pixel corner detection to obtain the coordinates of the feature points of the speckle pattern. Then, the random speckle image obtained during the last shooting is retrieved, and the random speckle images of the same area are compared based on the loading state of the bridge during the last shooting and the loading state of the bridge this time. The displacement vector field of the feature points is calculated by matching the feature points, and the strain distribution of the area is obtained. According to Hooke's law, combined with the elastic modulus of concrete and steel bars, the strain is converted into stress, and the mechanical parameters such as the principal stress magnitude, direction, and shear stress of the area are obtained. At the same time, by analyzing the change of displacement over time, the dynamic parameters such as the vibration frequency and damping ratio of the structure are calculated, and these mechanical parameters are associated with the corresponding coordinates of the external damage image.

[0041] The mechanical parameters, geometric feature parameters, and corresponding structural coordinates at the structural coordinates of each visible damage image are input into a pre-established first damage attention model for importance ranking, identifying key visible damage images. The first damage attention model utilizes a Transformer-based architecture. The input layer normalizes the mechanical parameters, geometric feature parameters, and structural coordinates and concatenates them into a feature vector as the model input. The model's intermediate layers include multiple multi-head attention mechanisms and feedforward neural network layers. The multi-head attention mechanism captures the correlations between different features and highlights key features by calculating attention weights for each feature. The feedforward neural network layer performs further nonlinear transformations and feature extraction. The output layer outputs an importance score for each visible damage image through a fully connected layer and a softmax function. The model is trained using the cross-entropy loss function and the Adam optimizer, with labels used. Training is iterated until the loss function converges. All visible damage images are ranked based on the importance scores output by the model, and a certain percentage of images with high scores, such as the top 30%, are selected as key visible damage images.

[0042] When there is no random speckle image, the geometric feature parameters and corresponding structural coordinates of each external damage image at the structural coordinates are input into the pre-established second damage attention model for importance sorting to determine the key external damage images. The structure of the second damage attention model is similar to that of the first damage attention model, but the input layer only contains geometric feature parameters and structural coordinates. The input model is also normalized and spliced, and feature processing is performed through the multi-head attention mechanism and the feedforward neural network layer. The output layer outputs the importance score of each external damage image. During the training process, the damage development speed predicted based on geometric features and historical damage data is used as the label, and the mean square error loss function and Adam optimizer are used for training until the loss converges. The key external damage images are selected according to the importance score sorting.

[0043] Finally, the mechanical parameters, geometric characteristic parameters and damage locations at the corresponding coordinates of the key external damage images (mapped to specific parts of the bridge through structural coordinates, such as the mid-span web, support bottom plate, etc.) are used as the first life influencing parameters.

[0044] Embodiments of the present invention provide a machine vision-based method for concrete bridge damage detection and remaining life prediction. This method not only focuses on the geometric characteristic parameters of visible damage but also incorporates random speckle image analysis to analyze mechanical parameters, breaking through the limitations of single geometric analysis. Furthermore, based on the presence of random speckle images, it flexibly utilizes primary and secondary damage attention models to prioritize importance, avoiding assessment bias caused by missing data and significantly improving assessment efficiency and accuracy. Compared to traditional methods that rely solely on manual experience to determine damage severity, this solution combines geometric characteristics, mechanical parameters, and intelligent algorithms to provide more scientific and reliable primary life-influencing parameters for concrete bridge remaining life assessment.

[0045] As an optional embodiment, the marker image generated by the manual test is a color image generated by the target position hole encountering the target reagent. Based on each marker image and the corresponding structural coordinates, each second life influencing parameter is determined, including:

[0046] Perform image preprocessing on the labeled image to obtain a target image; extract color areas in the target image, and divide the color areas into grades according to different color values ​​to obtain various color grade areas; perform statistics on each color grade area to determine the proportion of each color grade area in the entire hole; input each color grade area and the proportion of each color grade area in the entire hole into a pre-trained machine learning model to predict the carbonization depth of the benchmark concrete and the benchmark steel bar corrosion rate; obtain the temperature and humidity of the environment where the concrete bridge is located; input the environmental temperature and humidity and the carbonization depth of the benchmark concrete into a pre-established concrete carbonization depth correction model to obtain the corrected concrete carbonization depth; input the environmental temperature and humidity and the benchmark steel bar corrosion rate into a pre-established steel bar corrosion rate correction model to obtain the corrected steel bar corrosion rate; the corrected concrete carbonization depth, the corrected steel bar corrosion rate and the corresponding structural coordinates are used as the second life influencing parameters.

[0047] For example, the marked image is preprocessed by using image processing software to convert the image from RGB color space to HSV color space. Gaussian blur is used to reduce noise, and then an adaptive threshold method is used for binarization to highlight the marked area. Then, contour detection is used to extract the hole outline and filter out interference, and the region of interest (ROI) of the hole area is cropped for subsequent analysis. The color area is extracted within the ROI area of ​​the target image, and all pixels in the ROI are traversed and the RGB values ​​are extracted and converted into the hue value H in the HSV space. Based on the typical color characteristics of concrete carbonization and steel bar corrosion, multiple color grade intervals are pre-set, for example, reddish brown for severe rust, yellowish brown for medium rust, blue-gray for uncarbonized, and grayish white for carbonized. This can be achieved by using a threshold-based segmentation method or clustering algorithm (such as K-means) to generate a color grade mask image. The number of pixels in each color grade area is counted and divided by the total number of pixels in the ROI to obtain the proportion of holes in each color grade area.

[0048] The proportion of each color grade area is used as a feature vector and input into a pre-trained machine learning model, such as a random forest regression model. The model is trained based on laboratory standard specimen data. The specimens contain different carbonization depths and steel corrosion rates. The model outputs the benchmark concrete carbonization depth and benchmark steel corrosion rate through the correlation between image color features and the measured carbonization depth and corrosion rate.

[0049] At the same time, temperature and relative humidity are collected by temperature and humidity sensors deployed at the bridge site. The temperature and humidity data and the baseline carbonization depth are input into the carbonization depth correction model. For example, based on the empirical formula of Fick's diffusion law, the corrected carbonization depth is obtained. Specifically, the carbonization depth correction model is as follows:

[0050] d corr =d0·f T ·f RH ;

[0051] Among them, d corr is the corrected carbonization depth, d0 is the reference carbonization depth, f T is the temperature correction factor, f RH is the humidity correction factor;

[0052]

[0053] E a is the activation energy of carbonization reaction, R is the gas constant, generally 40 kJ / mol, T0 is the standard temperature, and T is the actual ambient temperature;

[0054]

[0055] RH0 is the standard humidity, RH is the actual ambient humidity, and n is the empirical coefficient, which is generally between 1.5 and 2.0.

[0056] Input the temperature and humidity data and the benchmark steel corrosion rate into the steel corrosion rate correction model. For example, combined with electrochemical theory, the corrected steel corrosion rate is obtained. Specifically:

[0057] r corr =r0·g T ·g RH ;

[0058] Among them, r corr is the corrected corrosion rate, r0 is the baseline corrosion rate, g T is the temperature correction factor, g RH is the humidity correction factor.

[0059]

[0060] β is an empirical parameter, generally 5000~6000, T0 is the standard temperature, and T is the actual temperature.

[0061]

[0062] RH0 is the standard humidity, RH is the actual humidity, and m is the empirical coefficient, which is generally 2.0 to 3.0.

[0063] Finally, the corrected carbonization depth, corrected steel corrosion rate and their corresponding structural coordinates are integrated as the second life influencing parameters.

[0064] An embodiment of the present invention provides a method for damage detection and remaining life prediction of concrete bridges based on machine vision. It realizes non-destructive detection through the color features of marked images, avoids the destructiveness of traditional detection, quantifies the relationship between color grade ratio and material degradation parameters, and combines real-time temperature and humidity to dynamically correct benchmark parameters. It accurately reflects the actual impact of the environment on concrete carbonization and steel corrosion, and improves the authenticity of the prediction results. At the same time, it integrates marked images and external damage image data to construct a complete life influencing parameter system from the two aspects of apparent characteristics and material degradation mechanism, which is more comprehensive.

[0065] As an optional implementation, before estimating the remaining life of a concrete bridge based on each first life influencing parameter and each second life influencing parameter, the method includes: acquiring three-dimensional point cloud data of the concrete bridge; reconstructing a digital model of the concrete bridge based on the three-dimensional point cloud data; comparing the digital model of the concrete bridge with a standard model of the concrete bridge to determine multiple structural difference points; estimating the remaining life of the concrete bridge based on each first life influencing parameter and each second life influencing parameter, including: estimating the remaining life of the concrete bridge based on each first life influencing parameter, each second life influencing parameter and multiple structural difference points.

[0066] For example, the three-dimensional point cloud data of a concrete bridge can be obtained through laser scanning. The point cloud data is imported into professional 3D modeling software, which automatically identifies the spatial distribution characteristics of the point cloud and connects the discrete points into a triangular mesh through a triangulation algorithm to preliminarily construct a bridge surface model. For complex structural parts of the bridge, such as the connection between the pier and the main beam, and the prestressed anchor area, the point cloud position is manually adjusted and the missing point data is supplemented to ensure the completeness of the model details. In order to make the model more realistic, the digital model is given corresponding material properties based on the real color and texture of the bridge concrete. At the same time, the continuity and smoothness of the model surface are checked, and areas with holes or overlaps are repaired and optimized to finally generate a digital model that accurately reflects the actual geometric shape of the bridge.

[0067] Then, the digital model of the concrete bridge is compared with the standard model. Specifically, as an optional implementation, the digital model of the concrete bridge is compared with the standard model of the concrete bridge to determine multiple structural differences, such as Figure 2 As shown, including:

[0068] S201, for each point in the point cloud, determine a local area with the point as the center, and represent the point cloud in the local area as a graph structure;

[0069] S202, using a graph convolutional network to encode the graph structure and learn a local feature descriptor of the point cloud, where the local feature descriptor includes local geometric features;

[0070] S203, fusing the local feature descriptor of the point cloud with the semantic label information of the point cloud to obtain a context feature descriptor, where the semantic label information of the point cloud is used to represent the environmental information of the bridge where the point cloud is located;

[0071] S204, performing coarse matching based on feature similarity by calculating the local feature descriptors of the points, and establishing a preliminary point pair correspondence relationship;

[0072] S205, calculating the similarity between the context feature descriptors of the corresponding point pairs, where the similarity includes local geometric feature similarity and semantic label consistency;

[0073] S206, in the concrete bridge digital model and the concrete bridge standard model, respectively obtaining local geometric features of corresponding point clouds in point pairs whose similarity exceeds a first threshold;

[0074] S207, determining geometric statistical features of the local area where the point cloud is located based on the local geometric features of the point cloud, where the geometric statistical features include at least one of a regional average normal vector, a curvature distribution, and a point density;

[0075] S208, combining geometric statistical features corresponding to the point cloud in the digital concrete bridge model with geometric statistical features corresponding to the point cloud in the standard concrete bridge model, determining a matching degree of the point cloud pair, and obtaining a matching point cloud pair;

[0076] S209: Determine a plurality of structural difference points based on the matching point cloud pairs.

[0077] For example, when comparing a digital model of a concrete bridge with a standard model to identify structural differences, a spherical region with a radius of r is first defined around each point in the digital model's point cloud as a local region. This radius, typically between 0.1 and 0.5 meters, is set based on the point cloud density and the complexity of the bridge structure, ensuring that the local region contains a sufficient number of point clouds and reflects structural details. The point cloud within the local region is then constructed as a graph, with each point serving as a node. Edge connectivity is determined by calculating the Euclidean distance between nodes. Points with a distance less than a threshold value, d, are connected. The value of d is based on the average spacing of the point clouds.

[0078] The constructed graph structure is then encoded using a graph convolutional network. The network input is the three-dimensional coordinates of the nodes. Through multi-layer convolution operations, the spatial relationship between each node and its neighboring nodes is learned, and the local feature descriptors of the point cloud are output. These descriptors contain geometric feature information such as the geometric shape and convexity of the local area. To further enrich the feature expression, semantic label information of the point cloud is incorporated. Semantic label information is obtained through manual annotation or semantic segmentation model prediction. It covers environmental information such as the part of the bridge where the point cloud is located (such as the main beam, piers, and guardrails) and functional areas (such as the load-bearing area and the non-load-bearing area). The semantic label is encoded into a vector form and spliced ​​with the local feature descriptor to obtain the contextual feature descriptor.

[0079] Then, by calculating the Euclidean distance or cosine similarity between the local feature descriptors of the point cloud, a coarse matching based on feature similarity is performed. Coarse matching can quickly establish preliminary point pair correspondences in large amounts of point cloud data, narrowing the search range for subsequent fine matching and improving overall matching efficiency. Specifically, a similarity threshold T1 is set, and preliminary correspondences are established for point pairs with similarities greater than T1. After the coarse matching is completed, fine matching is performed on the corresponding point pairs to calculate the similarity between their contextual feature descriptors. The similarity here consists of two parts: the similarity of local geometric features, calculated by comparing the differences in geometric parameters in the local feature descriptors; and the consistency of semantic labels.

[0080] The method for determining the consistency of semantic labels may include: inputting the semantic label of any point into a pre-established semantic library of concrete bridge standard models for query; when the standard semantic label is matched, the semantic label similarity is the first similarity, and the semantic library of the concrete bridge standard model contains standard semantic labels and semantic association maps; when the standard semantic label is not matched, the semantic label is split into minimum semantic units, and the associated terms of each minimum semantic unit are queried in the semantic association map; the associated terms of each semantic unit are exhaustively combined and matched twice with the standard semantic label; when the standard semantic label is matched twice, the semantic label similarity is set to the second similarity; when the standard semantic label is not matched twice, the semantic label similarity is set to the third similarity.

[0081] Specifically, when performing semantic label similarity calculations, the semantic label of any point is first input into the semantic library of the standard model of concrete bridges for query. For example, if the semantic label "precast T-beam" is input, and if there is a completely identical standard semantic label "precast T-beam" in the semantic library, the semantic label similarity is directly determined as the first similarity (for example, set to 1, indicating a complete match). If the input semantic label "steel-concrete composite beam flange" does not find a match in the semantic library, it needs to be split into minimum semantic units, namely "steel-concrete", "composite beam", and "flange". Subsequently, the associated terms of each minimum semantic unit are queried in the semantic association graph. For example, the associated terms of "steel-concrete" include "reinforced concrete", "composite beam" is associated with "composite beam", and "flange" is associated with "edge member". Then, these associated terms are exhaustively combined to generate a series of combination results such as "reinforced concrete composite beam edge member" and "steel-concrete composite beam flange", and then a secondary match is performed with the standard semantic label. If the combination result "reinforced concrete composite beam flange" has a corresponding standard semantic label in the semantic library, the semantic label similarity is set to the second similarity (for example, set to 0.8, indicating a high match). If all exhaustive combinations fail to match the standard semantic label, such as if the generated combination result has no corresponding item in the semantic library, the semantic label similarity is set to the third similarity (for example, set to 0, indicating no match). Through this method of first splitting, then exhaustively combining, and finally matching twice, the similarity between the input semantic label and the standard semantic label is systematically and comprehensively determined.

[0082] For local geometric features, they can be characterized by calculating geometric properties such as the normal vector and curvature of the points. The Euclidean distance, cosine similarity and other measurement methods are used to compare the degree of difference in properties such as the normal vector and curvature between the point pairs. In the calculation of geometric feature similarity, the degree of difference and the similarity are inversely mapped. The numerical difference is converted into a similarity value in the range of 0-1 through mathematical transformation. Specifically, for continuous scalar features such as curvature and point density, the Euclidean distance can be used to calculate the feature value difference d = |x1-x2| between the point pairs, and the maximum possible difference value D is pre-set. max Normalize and get the similarity as For example, if the curvature of point A is 0.05 and the curvature of point B is 0.03, then D max = 0.1, the similarity is 0.8; for vector features such as normal vectors, the cosine similarity is used to calculate the cosine value of the vector angle θ. The result directly reflects the degree of directional proximity. The value range is [-1, 1], and the absolute value is usually taken or calculated by Adjust to the interval [0,1], for example, the cosine similarity of the normal vector of point A (0,0,1) and the normal vector of point B (0,0.6,0.8) is 0.8; In addition, a tolerance threshold δ can be set for specific features. If the difference d≤δ, the similarity is 1, otherwise Linear attenuation. For example, if the point cloud density difference threshold is set to 5 points / cubic meter, the similarity of a point pair with 3 different points is 1. This method quantifies the degree of difference in geometric features between point pairs into a similarity value in the interval [0, 1]. The similarity values ​​of various geometric features are then averaged to obtain the local geometric feature similarity. Local geometric feature similarity is calculated by comparing geometric features at the single-point level (such as normal vector direction and curvature), quickly eliminating clearly mismatched point pairs and avoiding subsequent invalid calculations.

[0083] The similarity between the contextual feature descriptors of corresponding point pairs can be expressed as the weighted sum of the similarity of local geometric features and semantic consistency. Then, based on the overall similarity obtained by the weighted sum, point pairs with a similarity exceeding T2 are screened out from the digital model of the concrete bridge and the standard model, and the local geometric features of the corresponding point clouds in these point pairs are obtained.

[0084] Furthermore, based on the local geometric features of the point cloud, the geometric statistical features of the local area where the point cloud is located are calculated. On the basis of the local geometric feature similarity meeting the standards, this embodiment further analyzes the statistical laws of the regional-level geometric structure to avoid mismatching due to single-point noise or local anomalies. Specifically, the average normal vector of the region can be calculated by principal component analysis, the curvature distribution can be calculated using quadratic surface fitting, and the point density can be obtained by counting the number of points in the unit volume. Combined with the geometric statistical features of the corresponding point clouds in the digital model of the concrete bridge and the standard model, the degree of matching of the point cloud pairs is comprehensively evaluated. Specifically, by setting weights for different geometric statistical features, the weighted total difference value is calculated, and the point pairs with a difference value less than the set threshold T3 are determined as matching point cloud pairs. Finally, based on the matching point cloud pairs, the position differences, geometric shape differences, etc. of the corresponding points in the two models are analyzed to determine multiple structural difference points.

[0085] Traditional methods are based only on the comparison of geometric coordinates or simple shape features, and are suitable for scenarios with relatively low matching accuracy requirements. When the matching accuracy requirements are high, they are prone to ignoring local subtle features and semantic environmental influences, and may misjudge similar geometric shapes in different functional areas as differences. The embodiment of the present invention provides a machine vision-based concrete bridge damage detection and remaining life prediction method. By fusing features and semantics, it focuses on both structural details and environmental semantics. Through multi-level matching and filtering, it first performs coarse screening and then combines contextual features to eliminate false matches. Finally, it accurately locates substantial differences based on geometric statistical features to avoid misjudgments.

[0086] As an optional implementation, the first life-influencing parameter, the second life-influencing parameter, and the structural difference point all include corresponding structural coordinates. Based on each first life-influencing parameter, each second life-influencing parameter, and multiple structural difference points, the remaining life of the concrete bridge is estimated, such as Figure 3 As shown, including:

[0087] S301, associating the first life influencing parameter, the second life influencing parameter, and the structural difference point according to the structural coordinates;

[0088] S302, analyzing the coupling relationship between the first life influencing parameter, the second life influencing parameter, and the structural difference point based on the correlation relationship to obtain a coupling relationship feature;

[0089] S303, combining the first life influencing parameter, the second life influencing parameter, the structural difference point, and the coupling relationship characteristics, and based on the linear degradation assumption, calculating the instantaneous rate of damage expansion and material degradation;

[0090] S304, inputting the instantaneous rate of damage expansion and material degradation into the nonlinear model to simulate the acceleration process of long-term damage accumulation and generate a prediction sequence of each parameter changing over time;

[0091] S305 , comparing the predicted sequence with a preset structural failure threshold, and determining the time when each parameter reaches the threshold by combining the synergistic effect of the coupling relationship characteristics on the threshold;

[0092] S306: Take the minimum value as the remaining service life prediction value of the concrete bridge.

[0093] For example, when determining the predicted value of the remaining life of a concrete bridge, the association between different parameters and structural coordinates must first be established. Based on the first life-influencing parameters (such as the mechanical parameters, geometric characteristic parameters, and damage locations corresponding to the key external damage images), the second life-influencing parameters (corrected concrete carbonization depth, steel corrosion rate, and corresponding structural coordinates) and structural difference points (obtained by comparing the digital model of the bridge with the standard model) obtained in the early stage, all parameters are stored in the same data table with the structural coordinates (X, Y, Z) as the index. For example, for the bridge part at the coordinates (12.34, 5.67, 3.21), the crack width, carbonization depth, design and actual size deviation and other data of this position are recorded one by one to ensure that each coordinate point is associated with complete parameter information.

[0094] Next, the coupling relationships between the various parameters are analyzed. The data table is traversed, and for each set of coordinates corresponding to the parameters, the interactions between the primary life-influencing parameter, the secondary life-influencing parameter, and the structural difference point are studied. For example, a structural difference point at a certain location indicates that the beam size is smaller than the design value. Simultaneously, there is a wider crack (primary life-influencing parameter) and a higher carbonation depth (secondary life-influencing parameter). By analyzing multiple similar data points, it is found that the undersizing leads to stress concentration, which in turn accelerates crack propagation. Increased carbonation depth, in turn, weakens concrete strength, further exacerbating crack development and structural deformation. This method identifies coupling patterns under different parameter combinations and defines coupling relationship characteristics. For example, a positive correlation exists between structural size deviation, crack width, and carbonation depth, indicating that changes in one parameter accelerate changes in the others. This analysis process can first observe the distribution patterns between the variables using scatter plots. If a linear trend is observed, the Pearson coefficient is used to verify the strength of the linear positive correlation. If a curved or clustered distribution is observed, the mutual information is used to calculate the linear / nonlinear correlations between the parameters as coupling relationship characteristics.

[0095] Then, based on the linear degradation assumption, for each parameter combination at each coordinate point, the initial degradation rate of each parameter is first determined. For the carbonization depth of concrete and the steel corrosion rate, the previously corrected instantaneous rate is used as the initial value. Then, the initial rate is adjusted based on the coupling relationship characteristics. For example, if the coupling relationship shows that there is a strong positive correlation between the carbonization depth and the steel corrosion rate, and the carbonization depth at the current coordinate point is large, then the initial value of the steel corrosion rate is increased accordingly. Specifically, if the Pearson correlation coefficient between the carbonization depth and the steel corrosion rate is 0.8, it means that for every 1mm increase in carbonization depth, the steel corrosion rate may accelerate proportionally. Therefore, a multiplication correction can be used. For example, if the initial rate of a parameter A is v A , the current value of its coupling parameter B is x B , and the correlation coefficient between the two is ρ, then the correction rate of parameter A is:

[0096]

[0097] Among them, μ B and σ B is the mean and standard deviation of parameter B. Through this formula, the influence of the current state of parameter B on A is standardized and included in the correction. In this formula, the correlation coefficient is used as a proportional factor to reflect the linkage relationship between variables. This is to eliminate the dimension effect and make the coupling effects of different parameters comparable.

[0098] For nonlinear coupling (such as exponential acceleration), a nonlinear correction factor needs to be introduced. For example, in the calculation of the instantaneous rate of material degradation, when the crack width exceeds the critical value, the steel corrosion rate may increase exponentially. In this case, the correction factor can be set as:

[0099] Correction factor = e k·MI(A,B)·f(B) ;

[0100] Where MI(A,B) is the mutual information value of parameters A and B (range [0,1], the larger the value, the stronger the dependence); f(B) is the nonlinear function of parameter B, such as It represents the rate of change of B relative to the reference value B0, k is the calibration coefficient, which can be obtained by fitting historical data. Finally, the initial rate is multiplied by the correction factor to obtain the corrected steel corrosion rate.

[0101] In the above formula, when B exceeds the baseline value, the mutual information value exponentially amplifies its influence on A. For example, if B is the crack width and A is the steel corrosion rate, when the crack width exceeds the critical value B0, the corrosion rate accelerates exponentially with the crack width, and the degree of acceleration is positively correlated with the mutual information value (i.e., the coupling strength) between the two. This method calculates the instantaneous rate of damage growth and material degradation at each coordinate point by comprehensively considering various parameters and their coupling relationships.

[0102] The instantaneous rate is input into a nonlinear model to simulate the long-term damage accumulation process. A model that can reflect the nonlinear changes of parameters is selected, such as a neural network model or a nonlinear equation model based on physical and chemical principles. The instantaneous rate and coupling relationship characteristics of each coordinate point are used as model inputs. The model predicts the changes in each parameter over a period of time based on a set time step (such as in years), generating a prediction sequence of each parameter changing over time. For example, the model predicts the values ​​of parameters such as crack width, carbonization depth, and steel corrosion rate at different time points in the next 5, 10, and 15 years.

[0103] The predicted sequence is compared with preset structural failure thresholds. Failure thresholds for various parameters are pre-set based on bridge design specifications and safety standards, such as crack width exceeding 1.5mm, carbonization depth reaching the steel bar cover thickness, and steel bar corrosion rate exceeding 20%. The timing of each parameter reaching the threshold is determined by combining the synergistic effects of coupling relationship characteristics on the thresholds. For example, the coupled relationship between crack width and carbonization depth may accelerate steel bar corrosion, causing the corrosion rate to reach the threshold earlier than expected. By analyzing the predicted sequence, the specific time when each parameter reaches the corresponding threshold is determined.

[0104] Finally, among the times when all parameters reach their thresholds, the minimum value is selected as the predicted remaining life of the concrete bridge. In other words, if even one key parameter reaches the failure threshold, it could seriously impact the safety of the bridge structure, rendering it unusable. This approach comprehensively considers multiple parameters and their interrelationships, achieving a relatively accurate prediction of the remaining life of concrete bridges, providing an important basis for bridge maintenance and repair decisions.

[0105] The present invention provides a device for detecting damage and predicting the remaining life of a concrete bridge based on machine vision, comprising: an acquisition module for acquiring a surface image of a concrete bridge and its corresponding structural coordinates; a coordinate module for explicitly embedding the structural coordinates into the surface image of the concrete bridge to obtain a coordinated surface image; an identification module for inputting the coordinated surface image into an image recognition model to identify the image type and the corresponding structural coordinates, wherein the image type includes an explicit damage image and a marked image, wherein the marked image is a marked image generated by artificial testing; a first life influencing parameter determination module for determining each first life influencing parameter based on each explicit damage image and the corresponding structural coordinates; a second life influencing parameter determination module for determining each second life influencing parameter based on each marked image and the corresponding structural coordinates; and a remaining life determination module for estimating the remaining life of the concrete bridge based on each first life influencing parameter and each second life influencing parameter.

[0106] The present application also provides an electronic device, such as Figure 4 As shown, a processor 501 and a memory 502 , wherein the processor 501 and the memory 502 may be connected via a bus or other means.

[0107] The processor 501 may be a central processing unit (CPU). The processor 501 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0108] Memory 502, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs, and modules, such as the program instructions / modules corresponding to the machine vision-based concrete bridge damage detection and remaining life prediction method in the embodiments of the present invention. The processor executes the non-transitory software programs, instructions, and modules stored in memory to perform various processor functions and data processing.

[0109] The memory 502 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 502 may optionally include a memory remotely located relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0110] The one or more modules are stored in the memory 502 and when executed by the processor 501, perform the following steps: Figure 1 The machine vision-based concrete bridge damage detection and remaining life prediction method in the illustrated embodiment.

[0111] For details of the above electronic equipment, please refer to Figure 1 The corresponding descriptions and effects in the embodiments shown can be understood and will not be repeated here.

[0112] This embodiment further provides a computer storage medium storing computer-executable instructions capable of executing the machine vision-based concrete bridge damage detection and remaining life prediction method described in any of the above method embodiments. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD); the storage medium may also include a combination of the aforementioned types of memory.

[0113] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A method for concrete bridge damage detection and remaining life prediction based on machine vision, characterized in that: include: Obtain concrete bridge surface images and their corresponding structural coordinates; The structural coordinates are embedded into the concrete bridge surface image to obtain a coordinate surface image; Input the coordinate surface image into the image recognition model to identify the image type and the corresponding structural coordinates. The image type includes an external damage image and a marked image. The marked image is a marked image generated by artificial testing. Determine each first life influencing parameter according to each external damage image and the corresponding structural coordinates; Determine each second life influencing parameter according to each marked image and the corresponding structure coordinates; The remaining service life of the concrete bridge is estimated according to the first service life influencing parameters and the second service life influencing parameters.

2. The method for concrete bridge damage detection and remaining life prediction based on machine vision according to claim 1 is characterized in that: According to each external damage image and the corresponding structural coordinates, each first life influencing parameter is determined, including: Extracting geometric feature parameters of external damage images; Determining whether there is an identified random speckle image at the structural coordinates of the explicit damage image, where the random speckle image is one of the marked images; When there is a random speckle image, the mechanical parameters are analyzed based on the random speckle image; The mechanical parameters, geometric characteristic parameters and corresponding structural coordinates of each external damage image at the structural coordinates are input into a pre-established first damage attention model for importance sorting to determine the key external damage image; When there is no random speckle image, the geometric characteristic parameters and corresponding structural coordinates of each visible damage image are input into the pre-established second damage attention model for importance sorting to determine the key visible damage image; The mechanical parameters, geometric characteristic parameters and damage location at the corresponding coordinates of the key external damage image are taken as the first life influencing parameters.

3. The method for concrete bridge damage detection and remaining life prediction based on machine vision according to claim 1, characterized in that: The marker image generated by the artificial test is the color image generated by the target position hole encountering the target reagent. Based on each marker image and the corresponding structural coordinates, various second life influencing parameters are determined, including: Perform image preprocessing on the labeled image to obtain the target image; Extract color areas from the target image, divide the color areas into different levels according to different color values, and obtain different color level areas; Count the areas of each color grade and determine the proportion of each color grade area in the entire hole; The color grade areas and the proportion of each color grade area in the entire hole are input into a pre-trained machine learning model to predict the carbonation depth of the baseline concrete and the baseline steel corrosion rate; Obtain the temperature and humidity of the environment where the concrete bridge is located; Inputting the ambient temperature and humidity and the carbonation depth of the reference concrete into a pre-established concrete carbonation depth correction model to obtain a corrected concrete carbonation depth; Input the ambient temperature and humidity, and the baseline steel corrosion rate into the pre-established steel corrosion rate correction model to obtain the corrected steel corrosion rate; The corrected carbonization depth of concrete, the corrected steel corrosion rate and the corresponding structural coordinates are used as the second life influencing parameters.

4. A machine vision-based method for concrete bridge damage detection and remaining life prediction according to any one of claims 1 to 3, characterized in that: Before estimating the remaining life of a concrete bridge based on various first life influencing parameters and various second life influencing parameters, the following are included: Obtain 3D point cloud data of concrete bridges; Reconstruct the digital model of the concrete bridge based on the 3D point cloud data; Compare the digital model of the concrete bridge with the standard model of the concrete bridge to identify multiple structural differences; Estimate the remaining life of a concrete bridge based on various first life influencing parameters and various second life influencing parameters, including: The remaining life of a concrete bridge is estimated according to various first life-influencing parameters, various second life-influencing parameters, and a plurality of structural difference points.

5. The method for concrete bridge damage detection and remaining life prediction based on machine vision according to claim 4 is characterized in that: The digital concrete bridge model was compared with a standard concrete bridge model to identify several structural differences, including: For each point in the point cloud, a local area is determined with the point as the center, and the point cloud in the local area is represented as a graph structure; Graph convolutional networks are used to encode the graph structure and learn the local feature descriptors of the point cloud, which contain local geometric features. The local feature descriptor of the point cloud is fused with the semantic label information of the point cloud to obtain the context feature descriptor. The semantic label information of the point cloud is used to represent the environmental information of the bridge where the point cloud is located. By calculating the local feature descriptors of the points, a rough matching based on feature similarity is performed to establish the preliminary point pair correspondence; Calculate the similarity between the contextual feature descriptors of corresponding point pairs. The similarity includes local geometric feature similarity and semantic label consistency. In the concrete bridge digital model and the concrete bridge standard model, local geometric features of corresponding point clouds in point pairs whose similarity exceeds a first threshold are respectively obtained; Determining geometric statistical features of a local area where the point cloud is located based on local geometric features of the point cloud, where the geometric statistical features include at least one of a regional average normal vector, a curvature distribution, and a point density; Combining the geometric statistical features of the point cloud corresponding to the concrete bridge digital model with the geometric statistical features of the point cloud corresponding to the concrete bridge standard model, the matching degree of the point cloud pair is determined to obtain the matching point cloud pair; Based on the matching point cloud pairs, multiple structural difference points are determined.

6. The method for concrete bridge damage detection and remaining life prediction based on machine vision according to claim 4 is characterized in that: The first life influencing parameter, the second life influencing parameter, and the structural difference point all include corresponding structural coordinates. Based on each first life influencing parameter, each second life influencing parameter, and the plurality of structural difference points, the remaining life of the concrete bridge is estimated, including: According to the structural coordinates, the first life influencing parameter, the second life influencing parameter and the structural difference point are associated; According to the correlation relationship, the coupling relationship between the first life influencing parameter, the second life influencing parameter and the structural difference point is analyzed to obtain the coupling relationship characteristics; Combining the first and second life-influencing parameters, structural differences, and coupling relationship characteristics, the instantaneous rate of damage expansion and material degradation is calculated based on the linear degradation assumption. The instantaneous rates of damage growth and material degradation are input into the nonlinear model to simulate the accelerated process of long-term damage accumulation and generate a prediction series of each parameter changing over time; Compare the predicted sequence with the preset structural failure threshold, and determine the time when each parameter reaches the threshold by combining the synergistic effect of the coupling relationship characteristics on the threshold; The minimum value is taken as the predicted value of the remaining life of the concrete bridge.

7. The method for concrete bridge damage detection and remaining life prediction based on machine vision according to claim 5 is characterized in that: Calculate the similarity between the contextual feature descriptors of corresponding point pairs. The similarity includes local geometric feature similarity and semantic label consistency, including: The semantic label of any point is input into the pre-established semantic library of concrete bridge standard models for query. When a standard semantic label is matched, the semantic label similarity is the first similarity. The semantic library of concrete bridge standard models contains standard semantic labels and semantic association maps. When no standard semantic tag is matched, the semantic tag is split into minimum semantic units, and the associated terms of each minimum semantic unit are queried in the semantic association graph; The associated terms of each semantic unit are exhaustively combined and matched with the standard semantic labels for a second time. When the standard semantic label is matched for the second time, the semantic label similarity is set to the second similarity. When the secondary matching fails to find a standard semantic tag, the semantic tag similarity is set to the third similarity.

8. A device for concrete bridge damage detection and remaining life prediction based on machine vision, characterized in that: include: An acquisition module, used to obtain the surface image of the concrete bridge and its corresponding structural coordinates; A coordinate module is used to embed the structural coordinates into the concrete bridge surface image to obtain a coordinate surface image; A recognition module is used to input the coordinate surface image into the image recognition model to identify the image type and the corresponding structural coordinates. The image types include external damage images and marked images. The marked images are marked images generated by manual testing. A first lifespan influencing parameter determination module, configured to determine each first lifespan influencing parameter based on each external damage image and the corresponding structural coordinates; A second lifespan influencing parameter determination module, configured to determine each second lifespan influencing parameter based on each labeled image and the corresponding structural coordinates; The remaining life determining module is used to estimate the remaining life of the concrete bridge according to each first life influencing parameter and each second life influencing parameter.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the steps of the machine vision-based concrete bridge damage detection and remaining life prediction method according to any one of claims 1 to 7.

10. A computer storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by the processor, the steps of the method for concrete bridge damage detection and remaining life prediction based on machine vision are implemented.

Images