Steel box girder bottom damage identification and synchronous positioning method based on distributed camera network

Abstract

The invention discloses a steel box girder bottom damage identification and synchronous positioning method based on a distributed camera network. Comprising the following steps: S1, forming a series of imaging equipment units according to a spatially parallel structure, and connecting the imaging equipment units together through Bluetooth and a wireless area network so as to perform image data acquisition on any position of the bottom of a steel box girder; s2, transversely dividing the bottom area of the steel box girder according to the shooting coverage range of each imaging equipment unit; s3, performing two-stage synchronous identification and positioning based on deep learning driving, in the first stage, performing lightweight reconstruction on a panoramic image by taking MobilenetV4 as a key feature extraction tool, and performing disease area identification and positioning on a global level; and in the second stage, a YOLOv9 target detection framework is utilized to analyze a disease area, and disease information is provided in a local level. The design of the invention can be widely applied to inspection of the bottom of a long-span bridge and efficient analysis of diseases.

Description

Technical Field

[0001] The present invention belongs to the field of structural health detection in civil engineering, and more specifically, to a software and hardware system combining advanced equipment and intelligent algorithms, aiming to achieve accurate identification and rapid location of beam bottom defects. Background Art

[0002] Bridges occupy a pivotal position in the modern infrastructure system, serving as key nodes in the transportation network and core supporting structures for economic development and resource scheduling. During their service life, they face multiple problems such as structural fatigue, material aging, environmental erosion, and overload stress, which will affect their service life and safety performance. In order to effectively meet these challenges, regular health checks on bridges have become an indispensable measure. However, with the advancement of technology and the continuous improvement of testing standards, traditional manual inspection methods have gradually become unable to meet the efficiency and accuracy requirements of modern inspections. In particular, for inspection scenarios at the bottom of bridges, the difficulty of inspection is significantly increased due to their hidden location and complex environment. Therefore, the task of bridge bottom inspection urgently needs to introduce more advanced technologies.

[0003] In recent years, many scholars have developed customized detection equipment to address the bottleneck problems of the above detection scenarios. These intelligent detection devices are favored for their portability and flexibility, and can quickly reach the detection area, providing convenience for bridge detection. However, a core challenge is that in the complex beam bottom environment, most intelligent detection devices still rely on manual operation. This reliance not only limits the degree of automation of the detection process, but also the detection results are largely affected by the ability and experience of technicians, resulting in the inability to ensure the comprehensiveness and accuracy of the results. In addition, the equipment has limited endurance, which directly reduces the efficiency of the detection task. The detection system based on mobile vehicles realizes image acquisition of a large area of ​​the bridge. Relying on this type of detection platform can reduce the dependence on manual control and achieve stable and comprehensive detection of bridge bottom diseases. However, they are not portable and flexible enough in practical applications, and have certain requirements for the working environment, such as the need to occupy the emergency lane. In particular, in the detection of long-span bridges, there are blind areas that are difficult to cover.

[0004] Due to differences in personal experience, different technicians may come up with different evaluation results when analyzing the same test data. This analysis method based on subjective judgment not only introduces uncertainty in the results, but also affects the efficiency and reliability of the test. As a driving force for innovation in various fields, the development of deep learning technology has brought changes to traditional data processing methods. The deep learning model, with its simulation of the structure of the human brain, can automatically extract and learn complex features in the data by building a multi-layer neural network.

[0005] At present, the research on bridge bottom detection technology is mainly focused on small and medium-sized bridges. Existing equipment mostly relies on manual operation and has not yet been automated. Facing the detection needs of long-span bridges, especially the hard-to-reach areas at the bottom, comprehensive and efficient data collection is particularly difficult, which poses a major technical challenge. Summary of the invention

[0006] In response to the above technical problems, the present invention proposes a method for identifying and synchronously locating the bottom damage of steel box girders based on a distributed camera network, which breaks through the barriers of automated analysis of massive data and realizes efficient data processing. Through an innovative two-stage analysis method, efficient and intuitive guidance of the diseased area is achieved at the macro level, and the detailed development of the disease is accurately captured at the micro level. This analysis method achieves the optimal fusion between the global and local levels, providing a new solution for comprehensive inspection of the bottom of the bridge.

[0007] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0008] A method for bottom damage identification and synchronous positioning of a steel box girder based on a distributed camera network comprises the following steps:

[0009] S1. A series of imaging device units are formed in a spatially parallel structure, and the imaging device units are connected together via Bluetooth and wireless area network to collect image data at any position of the bottom of the steel box girder;

[0010] S2. According to the shooting coverage of each imaging device unit, the bottom area of ​​the steel box girder is divided horizontally, and the bottom area of ​​the steel box girder is decomposed into a number of sub-areas, the number of which corresponds to the imaging device unit, and a unique position code is assigned to each sub-area;

[0011] The bottom area of ​​the steel box girder is divided into blocks in the longitudinal direction, and the structural segments in each sub-area are defined as independent blocks. On the basis of the position coding of the original sub-area, the structural segment information is introduced to create the position coding of the block. According to the block concept defined above, the correlation between the images within each block range of the bottom space of the steel box girder is constructed to realize the generation of block images; finally, the block images of the local level are integrated according to the original spatial order using the position coding of each block to form a complete panoramic view of the bottom of the steel box girder;

[0012] S3, a two-stage simultaneous recognition and positioning method driven by deep learning, where:

[0013] In the first stage, the recognition tool based on the MobilenetV4 architecture was used to quickly screen out the key features related to the defects, and a fusion method based on the diffusion model was introduced to improve the generalization ability of the recognition tool. Then, according to the identified defect features, the panorama of the bottom of the steel box girder was lightweight reconstructed, and then the feature fine-grained parameters were introduced to adjust the information richness of the lightweight panorama. Finally, the position-related digital codes were marked on each area of ​​the lightweight panorama to identify and locate the defect area at the bottom of the steel box girder.

[0014] In the second stage, the target detection tool is used to conduct a detailed analysis of the number, type, and local location of the defect areas at the bottom of the steel box girder identified in the first stage.

[0015] In step S2, the block image generation specifically includes the following sub-steps:

[0016] S21, dividing the image into a feature extraction region and a retention region along the longitudinal direction; the feature extraction region is further subdivided into two left and right staple regions in the transverse direction, and using a parallel computing method, in the matching process, the left and right staple regions of the image at adjacent moments are fixed by analogy with using a stapler, that is, the corresponding relationship of feature points is searched for them respectively;

[0017] S22, using homography matrix calculation to obtain a stable transformation relationship matrix, grouping the image sequence contained in the block image, and splicing every two adjacent images as a group at the same time;

[0018] S23, looping steps S21 to S22 until the generation of the block image is completed.

[0019] In step S21, the division ratio between the feature extraction area and the retention area is 0.5.

[0020] The method also includes step S24, in which the image is decomposed into multiple levels according to the frequency through Gaussian filter and sampling operation, and then the high and low frequency components of each level are smoothed, weighted and superimposed. Finally, the components of each frequency band are re-added to effectively eliminate edge discontinuities.

[0021] In step S3, in the first stage of the two-stage simultaneous recognition and positioning method driven by deep learning, during the model training stage, a diffusion model is introduced to simulate and expand the real disease scenario, thereby improving the generalization ability of the model.

[0022] In step S3, in the fusion method based on the diffusion model, the Poisson fusion method is used to rotate, scale, and adjust the color and brightness of the generated disease image before the fusion operation;

[0023] Next, a fused area is randomly selected in the target image, and the gradient information of the area is calculated to obtain the trend of image detail changes;

[0024] Finally, the Poisson equation is used to take the boundary features of the source image and the gradient information of the target image as constraints.

[0025] In step S3, in the first stage of the two-stage synchronous recognition and positioning method driven by deep learning:

[0026] Using global macro information as guidance, the diseased areas in the lightweight panorama are traced back to the source. First, the diseased areas are identified in the panorama through macro analysis, and then the specific locations of these areas are accurately located.

[0027] These areas are then extracted from the lightweight panorama and restored to a more detailed image-based patch representation, allowing for a fine-scale analysis of the diseased areas without losing the original details.

[0028] Finally, the extracted detail features are mapped with the global information through the recorded position coding mapping to ensure the consistency and comprehensiveness of the disease analysis.

[0029] In the second stage of the two-stage simultaneous recognition and positioning method driven by deep learning, step S3 uses the YOLOv9 target detection framework to accurately analyze the identified defective area at the bottom of the steel box girder.

[0030] The beneficial effects of the present invention are:

[0031] In view of the complexity and challenges of detecting the bottom area of ​​long-span bridges, the present invention proposes and implements a hardware and software system that combines advanced equipment with intelligent algorithms. The proposed method has shown remarkable efficiency and accuracy in the detection of the bottom area of ​​long-span bridges. Through automated data collection and accurate defect identification, the system effectively optimizes the detection process. Detailed defect information supports engineering personnel in formulating accurate targeted maintenance strategies and improves the health management level of bridges. The innovative solution of the present invention provides an efficient global perspective for the detection of the bottom area of ​​long-span bridges, significantly improving the detection efficiency and accuracy.

[0032] 1. Dynamic visual perception equipment: The developed equipment has successfully broken through the limitations of traditional control, collection and transmission links, allowing technical experts to operate through a remote interactive interface and automatically complete comprehensive data collection of the beam bottom area. This design significantly improves the efficiency and comprehensiveness of data collection and reduces the need for on-site manual intervention.

[0033] 2. Panoramic image generation and optimization: By introducing the concept of blocks, the panoramic image stitching process is simplified and the image processing accuracy is improved. The orderly division of space and the extraction of feature points in overlapping areas effectively reduce the processing complexity and optimize the image alignment process. Although panoramic images may have complexity issues in information presentation, this method provides the advantage of a global perspective.

[0034] 3. Two-stage recognition and positioning driven by deep learning: Phase I: Use the recognition tool of the MobilenetV4 architecture to quickly screen out key features. In the model training phase, the diffusion model is introduced to simulate and expand the real disease scene, thereby improving the generalization ability of the model. The MobilenetV4 architecture is used to quickly screen key features, and the diffusion model is combined to simulate and expand the disease scene, thereby improving the generalization ability of the model. Through lightweight reconstruction and fine-grained feature adjustment, the recognition and positioning of diseased areas in panoramic images are significantly optimized. Phase II: The YOLOv9 target detection framework is used to accurately analyze the identified diseased areas, realizing the recognition of detailed information on the diseased areas. The refined processing at this stage further enhances the accuracy of disease analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the hardware module structure of the steel box girder bottom damage identification and synchronous positioning method based on a distributed camera network of the present invention;

[0036] Figure 2 A schematic diagram of a series of imaging device units of the present invention arranged in parallel in space;

[0037] Figure 3 The panorama generation process diagram based on the block concept proposed in the present invention;

[0038] Figure 4 The diffusion model and fusion technology proposed in the present invention generate realistic damage scenarios;

[0039] Figure 5 Schematic diagram of the two-phase method for simultaneous identification and location of bottom damage of a steel box girder proposed in the present invention;

[0040] Figure 6 A schematic diagram of the architecture of the feature description tool proposed in the present invention;

[0041] Figure 7 Schematic diagram of the YOLOv9 model framework proposed in the present invention;

[0042] Figure 8 The structural diagram of the generalized efficient layer aggregation network proposed by the present invention;

[0043] Fig. 9 The bridge specifications and equipment layout drawings mentioned in the example;

[0044] Fig.10 The application of dynamic perception in bottom panoramic imaging;

[0045] Fig.11 Detailed information of the block image mentioned in the example is displayed;

[0046] Fig.12 The example proposes a lightweight reconstructed panoramic image;

[0047] Fig.13 The damage analysis results of the proposed example are shown. DETAILED DESCRIPTION

[0048] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only for illustration purposes.

[0049] The present invention is not intended to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art all fall within the scope defined by the appended claims of this application.

[0050] The framework of this method is shown in the figure Figure 1 As shown. From the system level, this method can be divided into three parts:

[0051] S1. Dynamic visual perception equipment. This equipment breaks through the boundaries of control, collection and transmission, allowing technical experts to remotely operate through an interactive interface, thereby automatically completing comprehensive data collection of the beam bottom without on-site manual intervention.

[0052] S2. Block-based panorama generation strategy. The association between images is constructed through orderly division, and the block images are processed independently and in parallel. In the feature extraction and matching stage, the staple principle is used to extract the feature points of the overlapping areas to simplify the image alignment process and reduce the processing complexity.

[0053] S3, a two-stage synchronous recognition and positioning method driven by deep learning. In the first stage, the recognition tool of the MobilenetV4 architecture is used to quickly screen out key features. In the model training stage, the diffusion model is introduced to simulate and expand the real disease scene, thereby improving the generalization ability of the model. Then, based on the key features identified, the research team lightweight reconstructs the panorama. At the same time, the feature fine-grained adjustment parameters and position encoding technology are introduced to optimize the representation form of global information and improve the accuracy of recognition and positioning of diseased areas. In the second stage, the focus is turned to in-depth analysis of the diseased areas identified in the first stage. Through position encoding technology, it is possible to quickly trace back to the relevant original block image. Furthermore, the advanced target detection framework YOLOv9 is introduced to accurately identify the detailed information of the diseased area.

[0054] The main components of the dynamic visual perception device in S1 are:

[0055] The device is mainly composed of three parts: perception module, control transmission module and motion module. Figure 2 The perception module uses distributed cameras to replace the technicians' visual system and is connected via Bluetooth and wireless area network.

[0056] Edge computers are used to control the transmission module, and technicians can easily set and adjust various parameters of the acquisition phase, including time, acquisition mode, focal length and other parameters, as well as start and stop acquisition actions.

[0057] The beam bottom inspection vehicle serves as a motion module, giving the distributed camera the ability to collect data at any position on the beam bottom.

[0058] The specific content of the block-based panorama generation strategy in S2 is:

[0059] Traditional methods have problems such as high resource consumption, high time cost and strong manual dependence when processing large-scale image data. In addition, these methods are limited to local analysis of a single image, and it is difficult to consider the correlation between images, resulting in an inability to effectively conduct comprehensive analysis of state information. In order to overcome these challenges, the present invention is based on the concept of panoramic image generation and associates discrete image data. Then, the panoramic image of the beam bottom is reorganized with blocks as the construction unit.

[0060] The traditional panoramic image construction method, that is, splicing discrete images into a complete image containing all contents, has obvious limitations for the complex structure of long-span bridges, especially the bottom of the bridge. Therefore, the present invention divides the bottom area of ​​the beam horizontally according to the coverage of each unit in the dynamic perception chain. The process aims to decompose the bottom area of ​​the beam into several sub-areas, the number of which corresponds to the perception unit. At the same time, a unique position code is assigned to each sub-area. Furthermore, the bottom area of ​​the beam is divided into blocks in detail in the longitudinal direction. Based on the characteristics of the structural design, the structural segments in each sub-area are defined as independent blocks, thereby realizing the orderly discretization of the entire bottom space of the beam. On the basis of the position coding of the original sub-area, the position coding of the block is created by introducing the structural segment information. According to the block concept defined above, the correlation between the images within each block range of the bottom space of the beam is constructed to realize the generation of block images. Finally, the block images of the local level are integrated according to the original spatial order using the position coding of each block to form a complete panoramic image of the bottom of the beam. The introduction of the block concept allows each block to be processed independently and simultaneously, such as Figure 3 At the same time, each block is given a unique code to quickly locate any area in the global space.

[0061] To generate the block image, the image is first divided into a feature extraction area and a retention area along the vertical direction. Considering the maximum speed of the motion module and the overlap of the images at adjacent moments, the division ratio between the two areas is set to 0.5. The feature extraction area is further subdivided horizontally into two left and right staple areas, and parallel computing is used. In the matching process, the left and right staple areas of the image at adjacent moments are fixed by analogy with a stapler, that is, the correspondence of feature points is searched for them respectively. Finally, the homography matrix is ​​calculated to obtain a stable transformation relationship matrix. Under the condition of sufficient computing resources, the image sequence contained in the block image can be grouped, and every two adjacent images can be spliced ​​as a group at the same time. This grouping strategy enables the splicing of multiple pairs of images at the same time, thereby effectively improving the overall processing efficiency. Subsequently, the cycle operation is performed according to this strategy until the generation of the block image is completed.

[0062] The overlapping parts of the initial block images at adjacent moments may show obvious boundaries or transition areas. In order to solve this problem, the present invention adopts a pyramid expansion method, which decomposes the image into multiple levels according to the frequency through Gaussian filters and sampling operations, and then performs smooth weighting and superposition processing on the high and low frequency components of each level. Finally, the components of each frequency band are re-added to effectively eliminate edge discontinuities.

[0063] The specific contents of the two-stage simultaneous recognition and positioning method driven by deep learning in S3 are as follows:

[0064] The present invention targets the main defect characteristics of the bottom of the steel box girder - coating peeling and rust, and uses a diffusion model with high image generation quality and easy training as a defect synthesis tool. The present invention does not directly train and produce defects in a large scene, but focuses on local defect areas. Resources are concentrated on the synthesis of defect data to ensure that the main content of the synthetic image is concentrated on the information of the coating deterioration area. Subsequently, the synthesized defect information is mapped and fused into the real detection scene through methods such as Poisson fusion, such as Figure 4 shown.

[0065] The diffusion model mainly consists of two key stages: forward diffusion and reverse denoising. Figure 4 As shown in Figure 1, the forward diffusion stage is a Markov chain process with typical characteristics. In this process, Gaussian noise plays a key role in the input sample image in a continuous manner. ( The orderliness of the input image is gradually lost, and then gradually transformed into a series of noise distribution images. Each state change represents the advancement of a stage, from Z1 to Z2, and then to the subsequent stages, until the final state is formed. Variance of Gaussian noise distribution in Represents the mean coefficient. The state value at each moment t Only with the previous moment of Related. And with The increase of The closer the distribution state of is to pure noise, that is,

[0066] In the reverse denoising stage, the original image distribution is restored Use a neural network architecture with parameter θ to perform the reverse distribution In this way, with the help of the powerful learning and fitting ability of neural networks, we try our best to approximate and achieve effective prediction and estimation of the inverse distribution.

[0067] Neural networks need to be inversely distributed The mean coefficient of and variance During the network training process, The variance value of is explicitly specified as Just for There is a direct relationship between it and the noise distribution ∈. Therefore, the neural network is a tool specifically used for noise prediction, and its purpose is to make the output noise It can be as close to the normal distribution as possible. By training the noise predictor, the model tries to find an optimal way to approximate the normal distribution through continuous learning and adjustment. When the training is completed, a random sampling operation can be performed from the standard normal distribution to obtain a noise sample. Then, the neural network involved in the reverse process is used to restore and reshape it into a complete image based on this noise sample.

[0068] In further research, the present invention adopts advanced Poisson fusion technology. Before the fusion operation, the generated disease image needs to be rotated, scaled, and the color and brightness adjusted. These enhanced samples have different viewing angles, sizes, and lighting conditions, providing a richer data basis for the simulation process. Next, the fused area is arbitrarily selected in the target image, and the gradient information of the area is calculated to obtain the trend of image detail changes. Finally, a mathematical model is constructed based on the Poisson equation, taking the boundary features of the source image and the gradient information of the target image as constraints.

[0069] The present invention proposes an innovative two-stage disease simultaneous identification and location method, such as Figure 5As shown in the figure. In the first stage, the panoramic image is optimized and reconstructed through a feature representation method based on deep learning. In the second stage, based on the learning features of the defect area in the first stage, the detailed defect information is further learned in depth to provide a complete service status of the bridge: the number, type and specific location of the defect.

[0070] The block image contains a lot of non-critical background information, which does not contribute substantially to the operation and maintenance analysis. To address this problem, the present invention introduces the MobilenetV4-S network architecture to perform a preliminary analysis of the block image of the beam bottom area. The superior performance of MobilenetV4-S is due to the special module design, namely the Universal Inverted Bottleneck (UIB) module design. Figure 6 As shown. A concise digital coding system is used to characterize these features. In this system, the '0' code indicates that no disease is detected in the block image, suggesting that the block is in normal service status; in contrast, the '1' code indicates that a disease is found in the block image, suggesting that the block may be in abnormal service status. The present invention extends the visualization properties of lightweight panoramas by introducing color labels. Specifically, the 0 code is assigned a red label, while the 1 code is assigned a green label. In order to further highlight the diseased area and provide spatial information, the present invention also proposes a multi-dimensional feature representation method. The method is extended on the basis of the original 0-1 coding. On the two-dimensional plane, '0' means that no disease is detected in the block image, while '1' indicates that there is a disease. This information is replicated in three-dimensional space by assigning an additional attribute value to each block in the third dimension perpendicular to the two-dimensional plane. The value of the third dimension can be set to 1 for the diseased area, while the value of this dimension remains 0 for the normal area.

[0071] The present invention introduces a feature fine-grained adjustment parameter to optimize the expression of disease features in the global reconstruction stage. Combined with the structural characteristics of most bridge engineering projects, the present invention focuses on scenes with a high aspect ratio, works in the longitudinal direction of the block image, and evenly subdivides the image into multiple sub-blocks. A more refined feature extraction is performed on each sub-block. In addition, the position encoding of the third section is improved in combination with fine-grained parameters to form a new encoding system. In the process of block map refinement, clear position information is encoded for each sub-block.

[0072] To address the challenge of effectively combining global and local information, the present invention uses global-level macro information as a guide to trace the diseased areas in the lightweight panorama. First, the diseased areas are identified in the panorama through macro analysis, and then the specific locations of these areas are located. Next, these areas are extracted from the lightweight panorama and restored to a more detailed image-based block representation. The diseased areas are analyzed in detail without losing the original details. Finally, the extracted detail features are mapped to the global information through the recorded position coding mapping.

[0073] The fine detection phase uses the advanced target detection framework YOLOv9 to establish the mapping relationship between network output and input data. The model structure is as follows Figure 7 shown.

[0074] Example

[0075] The Nansha Bridge project consists of 7 approach bridges, 3 interchanges and 2 long-span suspension bridges, each with a length of more than 1 km. The focus of this study is the Dasha Strait Bridge, which has a double-tower, single-span suspension design and spans 360+1200+480m. The main span uses a steel box girder with a total width of 49.7m. In the longitudinal direction, sections 3 to 84 were selected for testing, each with a length of 12.8m and a total length of 1049.6m. For the transverse inspection, panels 5 to 11 were selected as targets. Fig. 9 This is the general layout diagram of the bridge.

[0076] In the dynamic perception system of the present invention, the layout strategy of the perception modules is carefully designed, such as Fig.10 As shown. The perception submodule is installed on the outside of the platform through a customized external bracket. In addition, considering the impact of extreme weather conditions such as typhoons and long-term vibration on the stability of the system, the design of the external bracket has taken targeted measures to firmly fix it to the inspection vehicle with high-strength bolt fasteners. The layout position of each perception module has been precisely calculated and is located directly below the center of the responsible panel. The field of view it covers is approximately 3.5 meters. This design ensures that a single module can fully cover the inspection needs of the panel, such as Fig.10 As shown. In terms of the computing control module, a dual module configuration is adopted, one controls 3 cameras and the other controls 4 cameras. This configuration is used to evaluate the stability of the control module in multi-tasking. Before the braking module is started, the acquisition instructions are sent to each perception unit synchronously through the software interface in the edge calculator. After the acquisition task is completed, the control software is responsible for securely transmitting the data in the perception module to the central control system through wireless transmission. Subsequently, the data is stably transmitted to the remote server through the intranet, and the computing resources of the server are used for in-depth data analysis and processing.

[0077] According to the panoramic generation strategy based on the block concept described in the present invention, the collected discrete data of the bottom of the beam are carefully integrated. The physical dimensions of a single plate in a standard beam section are 3 meters wide and 12.8 meters long. The resolution of the output image is 3000×12800 pixels through innovative splicing technology combined with the correction method of the homography matrix, and each pixel corresponds to an actual physical dimension of 1 mm. In order to verify the accuracy of the block image, the experimental steps are carefully designed. Specifically, red rectangular labels of various sizes are first artificially produced, and the sizes of these labels are accurately measured and recorded as reference standards in the subsequent verification process. Subsequently, these labels are accurately attached to the manually designed inspection points. The position information of these points is determined in advance by manual measurement. After the image acquisition task is completed, the research team analyzed the collected block images and compared the red rectangular labels in the images with the previously recorded size and position information. The conclusion shows that the dimensional measurement accuracy of the entire system can be controlled within 2mm, and the positioning accuracy error can be controlled within 3mm. This result proves the reliability and effectiveness of the image acquisition and processing method used in practical applications. Fig.11 was displayed in .

[0078] In the panorama generation stage, a total of 574 independent block images are generated, each of which corresponds to a part of the actual beam section, and the total area covers a beam section length of 1,049.6 meters. The present invention adopts a block-based positioning system. These block images can be simply combined to reconstruct a complete panorama of the beam bottom according to the established position coding rules, such as Fig.11 (Only part of the content is shown). The strategy of simply combining block images into a panoramic image is not feasible in practice. In order to meet this challenge, the present invention adopts the proposed two-stage disease simultaneous identification and location method to divide the problem into two levels: global and local. At the global level, focus on the characterization of the main features to avoid the complexity of displaying all detailed information in a global scope. Then focus on the important local disease areas and use high-precision processing methods to analyze these details.

[0079] In the first phase, the proposed feature characterization method was used to conduct in-depth feature analysis on 574 panel images. The total pixels of the panorama after lightweight reconstruction were , which was 99.99% smaller than the image-based representation. Due to factors such as the age of the bridge and the timeliness of the maintenance of the steel box bridge, the data collected on site had certain limitations in verifying the effectiveness of the method proposed in this invention.

[0080] Fig.12The maintenance of the bottom defects of steel box girders is shown in Figure 1. To overcome these limitations, the present invention designs a rehearsal experiment of simulated defects based on a synthetic database (covering real and synthetic defect morphology). First, a complete beam bottom data set is established using fine-grained real scene data. The data set contains 7 lateral target areas, each of which is further divided into 492 sub-blocks, forming a total of 3444 sub-block images. In order to simulate the random distribution of real defects, digital codes of 40 defect areas are randomly generated in each sub-area, and these codes correspond to the position codes in the block coordinates. Then, the defect morphology data is randomly extracted from the synthetic database, the number of extractions is set to 280 times, and it is ensured that the data extracted each time is not reused in subsequent operations. The extracted defect morphology data is fused into the sub-block image of the corresponding position code. The entire process adopts an end-to-end processing mode, with 3444 sub-block images as input and 3444 sub-block images containing defect areas as output. The processing process maintains the "black box" characteristics. The 3444 sub-block images containing the diseased areas were input into the first-stage feature characterization tool for preliminary analysis. A total of 277 diseased areas were identified, but 3 areas were not identified. After analyzing these missed areas, it was found that the disease scale was small and the morphology was not obvious. Therefore, the ability of the disease feature characterization tool in small target identification needs to be further improved in the future. Based on the output disease feature description results (including the 3 manually added missed areas) and position coding, a lightweight panorama was generated, such as Fig.12 In the block coordinate system, each subblock is not expanded along the longitudinal segment coordinate axis, but is added in the direction of the plate coordinate axis to form a subblock coordinate axis.

[0081] Phase 2 automatically extracts the location code of the diseased area and maps the feature description back to the original sub-block image. Using the detection tool, the disease category, quantity and location of all sub-block images were identified in detail. The specific results are as follows: Fig.13 As shown. In the figure, the first column shows the position encoding of the sub-block image, and the second column shows the disease detection effect of the corresponding area. The results show that peeling and rust diseases can be accurately identified. The third column records in detail the coordinate information of the disease in the block image. In order to locate the position of the disease more accurately, a coordinate system is established with the upper left corner of each block image as the origin. The local positioning coordinates in the sub-block are combined with the S-coded information of the sub-block image to convert it into global block coordinates. The present invention applies a thermal map to the estimation of the disease-affected area to provide a result that is more consistent with the disease morphology. The last column in the figure shows the visualization result of the disease-affected area estimation through a thermal map.

[0082] In summary, the specific embodiments verify the effectiveness of the solution proposed in the present invention and its applicability to complex projects.

[0083] What is disclosed above is only a typical embodiment of the present invention, but the embodiment of the present invention is not limited thereto. Any homogeneous modification made to this patent by any technician in this field after reading the patent should fall within the protection scope of the present invention.

Claims

1. A method for identifying and synchronously locating bottom damage of a steel box girder based on a distributed camera network, characterized in that: The following steps are involved: S1. A series of imaging device units are formed in a spatially parallel structure, and the imaging device units are connected together via Bluetooth and wireless area network to collect image data at any position of the bottom of the steel box girder; S2. According to the shooting coverage of each imaging device unit, the bottom area of ​​the steel box girder is divided horizontally, and the bottom area of ​​the steel box girder is decomposed into a number of sub-areas, the number of which corresponds to the imaging device unit, and a unique position code is assigned to each sub-area; The bottom area of ​​the steel box girder is divided into blocks in the longitudinal direction, and the structural segments in each sub-area are defined as independent blocks. On the basis of the position coding of the original sub-area, the structural segment information is introduced to create the position coding of the block. According to the block concept defined above, the correlation between the images within each block range of the bottom space of the steel box girder is constructed to realize the generation of block images; finally, the block images of the local level are integrated according to the original spatial order using the position coding of each block to form a complete panoramic view of the bottom of the steel box girder; S3, a two-stage simultaneous recognition and positioning method driven by deep learning, where: In the first stage, the recognition tool based on the MobilenetV4 architecture was used to quickly screen out the key features related to the defects, and a fusion method based on the diffusion model was introduced to improve the generalization ability of the recognition tool. Then, according to the identified defect features, the panorama of the bottom of the steel box girder was lightweight reconstructed, and then the feature fine-grained parameters were introduced to adjust the information richness of the lightweight panorama. Finally, the position-related digital codes were marked on each area of ​​the lightweight panorama to identify and locate the defect area at the bottom of the steel box girder. In the second stage, the target detection tool is used to conduct a detailed analysis of the number, type, and local location of the defect areas at the bottom of the steel box girder identified in the first stage.

2. The method for bottom damage identification and synchronous positioning of steel box beams based on a distributed camera network according to claim 1 is characterized in that: In step S2, the block image generation specifically includes the following sub-steps: S21, dividing the image into a feature extraction region and a retention region along the longitudinal direction; the feature extraction region is further subdivided into two left and right staple regions in the transverse direction, and using a parallel computing method, in the matching process, the left and right staple regions of the image at adjacent moments are fixed by analogy with using a stapler, that is, the corresponding relationship of feature points is searched for them respectively; S22, using homography matrix calculation to obtain a stable transformation relationship matrix, grouping the image sequence contained in the block image, and splicing every two adjacent images as a group at the same time; S23, looping steps S21 to S22 until the generation of the block image is completed.

3. The method for identifying and synchronously locating bottom damage of a steel box girder based on a distributed camera network according to claim 2 is characterized in that: In step S21, the division ratio between the feature extraction area and the retention area is 0.

5.

4. The method for identifying and synchronously locating bottom damage of a steel box girder based on a distributed camera network according to claim 2 is characterized in that: The method also includes step S24, in which the image is decomposed into multiple levels according to the frequency through Gaussian filter and sampling operation, and then the high and low frequency components of each level are smoothed, weighted and superimposed. Finally, the components of each frequency band are re-added to effectively eliminate edge discontinuities.

5. The method for identifying and synchronously locating bottom damage of a steel box girder based on a distributed camera network according to claim 1, characterized in that: In step S3, in the first stage of the two-stage simultaneous recognition and positioning method driven by deep learning, during the model training stage, a diffusion model is introduced to simulate and expand the real disease scenario, thereby improving the generalization ability of the model.

6. The method for identifying and synchronously locating bottom damage of a steel box girder based on a distributed camera network according to claim 2, characterized in that: Step S3: In the diffusion model-based fusion method, before using the Poisson fusion method, the generated disease image is rotated, scaled, and the color and brightness are adjusted; Next, a fused area is randomly selected in the target image, and the gradient information of the area is calculated to obtain the trend of image detail changes; Finally, the Poisson equation is used to take the boundary features of the source image and the gradient information of the target image as constraints.

7. The method for identifying and synchronously locating bottom damage of a steel box girder based on a distributed camera network according to claim 1, characterized in that: In the first stage of the two-stage synchronous recognition and positioning method driven by deep learning, step S3: Using global macro information as guidance, the diseased areas in the lightweight panorama are traced back to the source. First, the diseased areas are identified in the panorama through macro analysis, and then the specific locations of these areas are accurately located. These areas are then extracted from the lightweight panorama and restored to a more detailed image-based patch representation, allowing for a fine-scale analysis of the diseased areas without losing the original details. Finally, the extracted detail features are mapped with the global information through the recorded position coding mapping to ensure the consistency and comprehensiveness of the disease analysis.

8. The method for identifying and synchronously locating bottom damage of a steel box girder based on a distributed camera network according to claim 1, characterized in that: In the second stage of the two-stage simultaneous recognition and positioning method driven by deep learning, step S3 uses the YOLOv9 target detection framework to accurately analyze the identified defective area at the bottom of the steel box girder.

Images