Review method and system for rectification of safety hazards in urban rail transit engineering construction

Abstract

The application discloses a review method and system for rectification of hidden dangers in urban rail transit engineering construction safety, which comprises the following steps: firstly, acquiring a field image and labeling hidden danger elements to form a data set, then using knowledge graph technology to build a safety hidden danger rectification knowledge system to realize the mapping of hidden dangers and compliance image elements, and based on the knowledge system, generating a compliance image, i.e., a correction image of hidden danger image elements, through a controllable generation model. Finally, an actual rectification image is acquired and compared with the compliance image for semantic consistency analysis to evaluate the rectification effect. The safety management of urban rail transit engineering construction is more efficient, reliable and intelligent in hidden danger rectification review.

Description

Technical Field

[0001] This invention belongs to the field of application technology of intelligent safety management technology in urban rail transit engineering construction, and more specifically, it relates to the verification method and system for the rectification of safety hazards in urban rail transit engineering construction. Background Technology

[0002] In the construction of urban rail transit projects, the rectification and verification of safety hazards is a crucial step. Currently, this process mainly relies on manual operation. Verifiers must meticulously compare the status information at the time of hazard discovery with the actual situation after rectification, and comprehensively assess the effectiveness and compliance of rectification measures in accordance with national and industry standards such as the "Safety Code for Construction of Urban Rail Transit Projects." However, due to the variability of human subjective judgment, the enormous workload, and the complexity of the site environment, traditional manual verification methods face challenges such as low efficiency, unstable accuracy, and inconsistent standard implementation. These factors severely restrict the efficient and intelligent development of safety management and cannot meet the high standards of safety management required by modern urban rail transit projects.

[0003] While some technologies have attempted to assist or replace manual review through intelligent means, most of these technologies are still in the exploratory stage, and their accuracy, practicality, and universality have not yet met the needs of actual application. For example, existing image recognition-based construction hazard detection systems, BIM model-based construction site safety early warning methods and systems, drone-based inspection methods and systems, and intelligent construction equipment operation management platforms and control methods at construction sites, while improving the efficiency and intelligence level of hazard detection and recording, do not directly address the key aspects of hazard rectification and verification. The limitations of these technologies lie in their insufficient ability to identify complex hazard rectification, their inability to effectively combine standards and specifications for comprehensive judgment, and their limited intelligent decision-making capabilities, making it difficult to achieve comprehensive, accurate, and reliable verification of safety hazard rectification in urban rail transit engineering construction.

[0004] Therefore, those skilled in the art urgently need an intelligent verification method and system that combines urban rail transit engineering construction safety regulations and rules to improve the intelligence level and safety assurance capabilities of hidden danger rectification verification. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention can identify and analyze complex safety hazards, and can also conduct precise rectification assessments based on relevant safety regulations and standards, thereby providing strong technical support for the safety management of urban rail transit projects. Specifically, in the first aspect, this invention provides a method for reviewing the rectification of safety hazards in urban rail transit construction projects, the method comprising:

[0006] Step S1: Acquire images of complex scenes at the construction site, and classify and label the hidden danger image elements in the images to form a hidden danger image dataset;

[0007] Step S2: Generate a knowledge system for safety hazard rectification based on knowledge graph technology; the knowledge system for safety hazard rectification is a knowledge system that forms a mapping relationship with the compliant image elements corresponding to the target hazard image in the hazard image;

[0008] Step S3: Based on the safety hazard rectification knowledge system, a compliant image of the same scene as the hazard image is generated through a controllable generation model; the compliant image is an image after the elements of the hazard image have been corrected according to the safety hazard knowledge system.

[0009] Step S4: Obtain the actual rectified image of the hidden danger image and the compliant image corresponding to the hidden danger image, and compare and analyze them based on the image semantic consistency metric to display the analysis and evaluation results.

[0010] In the first aspect, step S2 further includes:

[0011] Step S21: The standard elements of the compliant image are parsed using natural language processing technology to extract key information units, including hazard type, rectification requirements, construction process, material specifications, and safety distance, and a map structure of target hazard image elements and compliant image elements is established.

[0012] Step S22: Set up a query engine to query the rectification standards for the target hidden dangers; the query engine includes graph database storage and retrieval for fast indexing based on the type of hidden danger;

[0013] Step S23: Form the knowledge system for rectifying the safety hazards.

[0014] In the first aspect, step S3 further includes:

[0015] Step S31: Using the conditional diffusion model as the basic framework, prepare the hazard image and rectification specification vector as input;

[0016] Step S32: The rectification standard vector is encoded using a text encoder, and the input image is encoded using a pre-trained variational autoencoder. The text encoding and image encoding are fused in the conditional diffusion model through mechanisms such as cross attention to form a unified conditional vector.

[0017] Step S33: Minimize the diffusion prediction error. Use the weighted mean square error loss function as the training objective. Through the forward process: gradually add noise to the input image to simulate the process from a clean image to pure noise. Through the backward process: based on the unified conditional vector, predict the noise residual at each step through the conditional denoising network to guide the denoising process, so that the generated latent space gradually tends to the image that conforms to the rectification specifications.

[0018] In the first aspect, step S31 includes performing a forward diffusion process using Formula 1, wherein the forward diffusion process is a process of gradually adding noise from a compliant image to noise, and Formula 1 includes:

[0019]

[0020] Where, x t For the image representation at time step t; x t-1 The image representation at time step t-1; To represent a normal distribution; β t I is the noise gain coefficient at time step t, used to control the amount of noise added at each step. I is the identity matrix used to maintain the independence of each dimension of the image.

[0021] In the first aspect, step S32 includes performing a reverse generation process using Formula 2, wherein the reverse generation process is a process of gradually restoring a compliant image from noise, and Formula 2 includes:

[0022]

[0023] Among them, P θ To represent the probability distribution controlled by parameter θ; x t-1 X represents the image at time step t-1, i.e., the denoised image; t Let be the image representation at time step t, i.e., the current noisy image; s is the rectification specification vector, containing the standard information required to convert the hazard image into a compliance image; μ θ Σ is the mean function, determined by the model parameter θ, used to predict the mean of the denoised image; θ θ is the covariance function, determined by the model parameter θ, used to predict the covariance of the denoised image; t is the time step.

[0024] In the first aspect, step S33 includes using Formula 3 to set a training objective, the training objective being to minimize the diffusion prediction error, employing a weighted mean square error loss, wherein Formula 3 includes:

[0025]

[0026] Among them, L diffusionLet be the loss function of the diffusion model; E is the expectation operator, representing the average over all possible x0, t, and ∈; X0 is the original compliant image; t is the time step; ∈ is the noise sampled from the standard normal distribution N(0, I); ∈ θ The noise predicted by the model is determined by the model parameter θ; |.| 2 The squared error is used to calculate the difference between the predicted noise and the actual noise.

[0027] In the first aspect, step S4, the comparative analysis based on image semantic consistency measurement includes: a consistency measurement step based on perceptual loss, which is implemented through formula four:

[0028]

[0029] Among them, I pred To predict the image (after rectification), I ref To utilize the rectified image generated by the preceding generative model, φ l C represents the feature map of the l-th layer extracted by the VGG model. l H l W l Let represent the number of channels and size of the feature map at layer l, respectively, and L represent the set of all intermediate features.

[0030] In the first aspect, step S4, the comparative analysis based on image semantic consistency measurement includes: a measurement based on segmentation semantic consistency, which is implemented through formula five:

[0031] SCS(I rect ,I ref )=λ1·IoU(S rect ,S ref )+λ2·SSIM(I rect ,I ref (Formula 5)

[0032] Among them, IoU measures the intersection-union ratio of the segmentation between the reworked image and the reference image, SSIM measures the structural similarity between the reworked image and the reference image, and λ1 and λ2 are the weight coefficients of the two scores, which can be set to 1 by default.

[0033] In the first aspect, step S4 further includes: introducing a local relaxation index to adjust the intersection-union ratio calculation formula six to adapt to reasonable local changes during the image segmentation process, wherein formula six includes:

[0034]

[0035] Among them, |S rect ∩S ref| represents the number of pixels in the intersection region of Srect and Sref, |S rect ∪S ref | represents the number of pixels in the union region of Srect and Sref, and δ is a non-negative relaxation parameter used to control the tolerance for local changes.

[0036] Secondly, the present invention provides a verification system for the rectification of safety hazards in urban rail transit engineering construction, wherein the verification system is applied to the verification method for the rectification of safety hazards in urban rail transit engineering construction.

[0037] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0038] 1. The verification method for safety hazard rectification in urban rail transit construction provided by this invention significantly improves the efficiency and accuracy of hazard rectification verification by automatically acquiring and processing images of urban rail transit construction sites. First, the method rapidly constructs a high-quality hazard image dataset by classifying and labeling hazard image elements, laying a solid foundation for subsequent intelligent verification. Second, the safety hazard rectification knowledge system established using knowledge graph technology ensures accurate mapping between hazard images and compliant image elements, thereby improving the accuracy and compliance of rectification measures.

[0039] 2. Furthermore, this method generates compliant images based on a knowledge system for safety hazard rectification using a controllable generative model, achieving intelligent and automated hazard rectification and effectively reducing errors caused by human factors. Finally, by comparing and analyzing the actual rectified images with the compliant images through image semantic consistency measurement, it can not only accurately identify rectification compliance but also intuitively display the analysis and evaluation results, further improving the intelligence level and safety assurance capability of hazard rectification review. The application of this method enables more efficient, reliable, and intelligent hazard rectification review in the safety management of urban rail transit engineering construction. Attached Figure Description

[0040] Figure 1 This is a flowchart of a verification method for rectifying safety hazards in urban rail transit construction, as described in an embodiment of the present invention. Figure 1 ;

[0041] Figure 2 This is a flowchart of a verification method for rectifying safety hazards in urban rail transit construction, as described in an embodiment of the present invention. Figure 2 ;

[0042] Figure 3 This is a flowchart of a verification method for rectifying safety hazards in urban rail transit construction, as described in an embodiment of the present invention. Figure 3 . Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0044] Example 1:

[0045] Please see Figure 1-3 This embodiment provides a verification method for the rectification of safety hazards in urban rail transit construction. This invention, by integrating standard knowledge to generate compliant images and employing a semantic-based consistency metric, overcomes the limitations of traditional pixel-based comparison methods, achieving a more intelligent verification technology that better reflects the actual complexity of rectification efforts. Specifically, it includes:

[0046] Step S1: Acquire images of complex scenes at the construction site, and classify and label the hazard image elements in the images to form a hazard image dataset; Step S2: Generate a safety hazard rectification knowledge system based on knowledge graph technology; the safety hazard rectification knowledge system is a knowledge system that forms a mapping relationship with the compliant image elements corresponding to the target hazard image in the hazard images; Step S3: Based on the safety hazard rectification knowledge system, generate a compliant image of the same scene corresponding to the hazard image through a controllable generation model; the compliant image is an image after the hazard image elements have been corrected according to the safety hazard knowledge system; Step S4: Acquire the rectified image of the hazard image and the compliant image corresponding to the hazard image, and compare and analyze them based on the image semantic consistency metric to display the analysis and evaluation results.

[0047] Specifically, the verification method provided in this embodiment significantly improves the efficiency and accuracy of hazard rectification verification by automating the acquisition and processing of images from urban rail transit construction sites. First, the method rapidly constructs a high-quality hazard image dataset by classifying and labeling hazard image elements, laying a solid foundation for subsequent intelligent verification. Second, the safety hazard rectification knowledge system established using knowledge graph technology ensures accurate mapping between hazard images and compliant image elements, thereby improving the accuracy and compliance of rectification measures. Furthermore, this method generates compliant images based on the safety hazard rectification knowledge system using a controllable generation model, achieving intelligent and automated hazard rectification and effectively reducing errors caused by human factors. Finally, by comparing and analyzing actual rectified images with compliant images through image semantic consistency measurement, it can not only accurately identify rectification compliance but also intuitively display the analysis and evaluation results, further improving the intelligence level and safety assurance capabilities of hazard rectification verification. The application of this method enables more efficient, reliable, and intelligent hazard rectification verification for the safety management of urban rail transit construction projects.

[0048] Furthermore, to implement an intelligent verification method for hazard rectification, a high-quality hazard image dataset must first be established. For the complex scenarios at urban rail transit construction sites, high-resolution imaging equipment (such as industrial cameras, LiDAR combined with RGB-D cameras, etc.) should be used to acquire images from multiple angles and at multiple time periods, ensuring coverage of hazard scenarios under different construction techniques and environmental conditions. Image acquisition must simultaneously record metadata such as location information, construction stage, and timestamps to support subsequent hazard tracing and dynamic analysis.

[0049] The collected raw images need to be classified and annotated with fine granularity by professionals in accordance with the "Safety Code for Construction of Urban Rail Transit Engineering" and other national industry standards. The annotation adopts a multi-label system, specifically including but not limited to: template detachment, improper scaffolding erection, missing fire-fighting facilities, exposed electrical wiring, and other hazard categories. The annotation is not limited to the hazard area itself, but also needs to annotate the hazard's impact range, adjacent structures, and other related elements to facilitate subsequent semantic modeling.

[0050] After image acquisition, the hazard type needs to be manually specified or automatically identified by a hazard identification model. To enhance the consistency and standardization of hazard categories, a standardized hazard naming system needs to be established to ensure the accuracy of hazard categories and provide a foundation for subsequently finding the correct hazard rectification standards.

[0051] Furthermore, step S2 also includes:

[0052] Step S21: The standard elements of the compliant image are parsed using natural language processing technology to extract key information units, including hazard type, rectification requirements, construction process, material specifications, and safety distance, and a map structure of target hazard image elements and compliant image elements is established.

[0053] Step S22: Set up a query engine to query the rectification standards for the target hidden dangers; the query engine includes graph database storage and retrieval for fast indexing based on the type of hidden danger;

[0054] Step S23: Form the knowledge system for rectifying the safety hazards.

[0055] Specifically, for step two, in order to achieve intelligent generation of hazard images into compliant images, the system needs to have a structured understanding and query capability of current standards and regulations. This module constructs a safety hazard rectification knowledge system based on Knowledge Graph (KG) technology.

[0056] First, the code clauses are parsed using Natural Language Processing (NLP) technology to extract key information units, including hazard types, rectification requirements, construction techniques, material specifications, and safety distances, and an entity and relation graph structure is established. For example, the rectification requirement for "scaffolding not erected according to specifications" is "the erection height does not exceed the specified limit, and standard fasteners are used for fixing nodes."

[0057] The query engine uses a graph database (such as Neo4j) for storage and retrieval, and supports fast indexing based on hazard type. For example, for a hazard H, the query for its rectification standard requirement R(H) can be defined as:

[0058] R(H) = Retrieve(KG, type(H))

[0059] Here, type(H) is the standard classification label for the potential hazard. Furthermore, to support the dynamic invocation of standard constraints during training of the generative model, a standard encoder needs to be designed to embed the rectification requirements as vectorized inputs for generation. Specifically, a Transformer encoder can be used, outputting a standard vector s∈R. d This module enables the system to automatically map potential hazards to compliance requirements, ensuring that the generated images strictly conform to standards and specifications.

[0060] In a preferred embodiment, step S3 further includes:

[0061] Step S31: Using a conditional diffusion model as the basic framework, such as InstructPix2Pix, prepare the hazard image and use the CLIP text encoder rectification specification vector as input.

[0062] Step S32: The rectification standard vector is encoded using a text encoder, and the input image is encoded using a pre-trained VAE (variational autoencoder). The text encoding and image encoding are fused in the conditional diffusion model through mechanisms such as cross attention to form a unified conditional vector.

[0063] For the text encoder in step S32, a CLIP text encoder can be used, which uses a pre-trained VAE variational autoencoder to encode the input image.

[0064] Step S33: Minimize the diffusion prediction error. Use the weighted mean square error loss function as the training objective. Through the forward process: gradually add noise to the input image to simulate the process from a clean image to pure noise. Through the backward process: based on the unified conditional vector, predict the noise residual at each step through a conditional denoising network (such as U-Net) to guide the denoising process, so that the generated latent space gradually tends to the image that conforms to the rectification specifications.

[0065] For the denoising network in step S33, this embodiment preferably uses U-Net.

[0066] Specifically, the core of this module lies in generating a compliant image from a potential hazard image based on a Conditional Diffusion Model (CDM), integrating rectification standard information to ensure the generated result conforms to the specifications. The CDM structure is used, with the input being a potential hazard image x0 and the corresponding rectification standard vector s. Noise is added during the diffusion process, and the reverse diffusion process combines the standard vector to gradually restore the compliant image x0. T .

[0067] To enhance compliance correction in key rectification areas (i.e., areas with potential hazards), a Local Attention mechanism is introduced. This mechanism uses a saliency map to guide the model to more accurately recover details at the hazard locations. Ultimately, the trained Diffusion generation model can generate compliant images that meet the required standards based on the hazard images and rectification criteria. This will provide a reference for subsequent review.

[0068] Further, step S31 includes performing a forward diffusion process using Formula 1, wherein the forward diffusion process is a process of gradually adding noise from a compliant image, and Formula 1 includes:

[0069]

[0070] Where, x t For the image representation at time step t; x t-1 The image representation at time step t-1; To represent a normal distribution; β t I is the noise gain coefficient at time step t, used to control the amount of noise added at each step. I is the identity matrix used to maintain the independence of each dimension of the image.

[0071] Further, step S32 includes performing a reverse generation process using Formula 2. This reverse generation process is the process of gradually restoring a compliant image from noise. Formula 2 includes:

[0072]

[0073] Among them, P θ To represent the probability distribution controlled by parameter θ; x t-1 X represents the image at time step t-1, i.e., the denoised image; t Let be the image representation at time step t, i.e., the current noisy image; s is the rectification specification vector, containing the standard information required to convert the hazard image into a compliance image; μ θ Σ is the mean function, determined by the model parameter θ, used to predict the mean of the denoised image; θ θ is the covariance function, determined by the model parameter θ, used to predict the covariance of the denoised image; t is the time step.

[0074] Further, step S33 includes using Formula 3 to set a training objective, wherein the training objective is to minimize the diffusion prediction error, and a weighted mean square error loss is used. Formula 3 includes:

[0075]

[0076] Among them, L diffusion Let be the loss function of the diffusion model; E is the expectation operator, representing the average over all possible x0, t, and ∈; X0 is the original compliant image; t is the time step; ∈ is the noise sampled from the standard normal distribution N(0, I); ∈ θ The noise predicted by the model is determined by the model parameter θ; |.| 2 The squared error is used to calculate the difference between the predicted noise and the actual noise.

[0077] In a preferred embodiment, to address the issue of natural differences between the actual rectified image and the generated standard compliant image, this module proposes an image semantic consistency measurement method. Image semantic consistency measurement comprises two aspects: a consistency measurement based on perceptual loss and a measurement based on segmentation semantic consistency. The consistency measurement based on perceptual loss does not directly calculate differences in pixel space (such as L2 or L1), but rather calculates the differences between images in the feature space of a deep neural network, specifically by using a pre-trained image classification network (such as VGG) to extract semantic feature representations of the image and comparing the differences between these representations. Specifically, the intermediate layer features of the VGG-16 or VGG-19 network, such as the outputs of `relu1_2`, `relu2_2`, `relu3_3`, `relu4_3`, etc., are used as the "perceptual representation" of the image. The L2 loss (Euclidean distance) is calculated between the feature maps extracted from these layers. This step is implemented through Equation 4:

[0078]

[0079] Among them, I pred To predict the image (after rectification), I ref To utilize the rectified image generated by the preceding generative model, φ l C represents the feature map of the l-th layer extracted by the VGG model. l H l W l Let represent the number of channels and size of the feature map at layer l, respectively, and L represent the set of all intermediate features.

[0080] Perceptual loss takes into account the high-level semantic consistency of an image (such as structure and object category). It is insensitive to small pixel offsets but sensitive to differences in texture, shape, and boundaries.

[0081] In a preferred embodiment, in step S4, regarding semantic consistency in segmentation, firstly, the rectified image I... rect With reference compliance image I ref Perform semantic segmentation to obtain the corresponding semantic segmentation graph S. rect and S ref The segmentation model employs a lightweight architecture, including but not limited to DeepLabV3+, and is trained on urban rail transit engineering construction image data. After obtaining the segmentation map, a Semantic Consistency Score (SCS) is defined, which is a comprehensive evaluation based on category matching and spatial distribution similarity. This step is achieved through Equation 5:

[0082] SCS(I rect ,I ref )=λ1·IoU(Srect ,S ref )+λ2·SSIM(I rect ,I ref (Formula 5)

[0083] Among them, IoU measures the intersection-union ratio of the segmentation between the reworked image and the reference image, SSIM measures the structural similarity between the reworked image and the reference image, and λ1 and λ2 are the weight coefficients of the two scores, which can be set to 1 by default.

[0084] In a preferred embodiment, in addition to adapting to reasonable local changes (such as the addition of temporary support structures), a flexible matching mechanism is designed to tolerate changes in small-scale non-critical areas. A local relaxation index δ is introduced to adjust the IoU calculation, specifically calculated using Formula Six:

[0085]

[0086] Among them, |S rect ∩S ref | represents the number of pixels in the intersection region of Srect and Sref, |S rect ∪S ref | represents the number of pixels in the union region of Srect and Sref, and δ is a non-negative relaxation parameter used to control the tolerance for local changes.

[0087] Finally, fusion And SCS, such as summation fusion, average fusion, etc. If the fused result is greater than a certain threshold T, the rectification result is considered compliant; otherwise, it is not compliant.

[0088] Overall, this invention offers several advantages over existing technologies. In terms of technological advancement, by integrating standard knowledge to generate compliant images and employing semantic-based consistency metrics, it overcomes the limitations of traditional pixel-based comparison methods, achieving a more intelligent and realistically complex review technology. Regarding the accuracy of rectification review, by introducing image semantic understanding, the review system can tolerate reasonable scene changes and accurately identify rectification compliance, significantly improving the accuracy and reliability of hazard rectification review results. In terms of rectification and review efficiency, by pre-generating standard compliant images and combining them with intelligent semantic comparison algorithms, it greatly reduces the subjective judgment and workload of manual review, significantly improving the processing speed and efficiency of large-scale hazard rectification review.

[0089] Example 2:

[0090] This embodiment provides a verification system for the rectification of safety hazards in urban rail transit engineering construction, and the verification system is applied to the verification method for the rectification of safety hazards in urban rail transit engineering construction.

[0091] Specifically, the review system utilizes review methods to automate the acquisition and processing of images from urban rail transit construction sites, significantly improving the efficiency and accuracy of hazard rectification review. First, this method rapidly constructs a high-quality hazard image dataset by classifying and labeling hazard image elements, laying a solid foundation for subsequent intelligent review. Second, the safety hazard rectification knowledge system established using knowledge graph technology ensures a precise mapping between hazard images and compliant image elements, thereby improving the accuracy and compliance of rectification measures.

[0092] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for reviewing the rectification of safety hazards in urban rail transit construction projects, characterized in that, The method includes: Step S1: Acquire images of complex scenes at the construction site, and classify and label the hidden danger image elements in the images to form a hidden danger image dataset; Step S2: Generate a knowledge system for safety hazard rectification based on knowledge graph technology; the knowledge system for safety hazard rectification is a knowledge system that forms a mapping relationship with the compliant image elements corresponding to the target hazard image in the hazard image; Step S3: Based on the safety hazard rectification knowledge system, a compliant image of the same scene as the hazard image is generated through a controllable generation model; the compliant image is an image after the elements of the hazard image have been corrected according to the safety hazard knowledge system. Step S4: Obtain the actual rectified image of the hidden danger image and the compliant image corresponding to the hidden danger image, and compare and analyze them based on the image semantic consistency metric to display the analysis and evaluation results.

2. The method for reviewing the rectification of safety hazards in urban rail transit construction projects according to claim 1, characterized in that, Step S2 further includes: Step S21: The standard elements of the compliant image are parsed using natural language processing technology to extract key information units, including hazard type, rectification requirements, construction process, material specifications, and safety distance, and a map structure of target hazard image elements and compliant image elements is established. Step S22: Set up a query engine to query the rectification standards for the target hidden dangers; the query engine includes graph database storage and retrieval for fast indexing based on the type of hidden danger; Step S23: Form the knowledge system for rectifying the safety hazards.

3. The method for reviewing the rectification of safety hazards in urban rail transit construction projects according to claim 1, characterized in that, Step S3 further includes: Step S31: Using the conditional diffusion model as the basic framework, prepare the hazard image and rectification specification vector as input; Step S32: The rectification specification vector is encoded using a text encoder, and the input image is encoded using a pre-trained variational autoencoder. The text encoding and image encoding are fused in the conditional diffusion model through a cross-attention mechanism to form a unified conditional vector. Step S33: Minimize the diffusion prediction error. Use the weighted mean square error loss function as the training target. Through the forward process: gradually add noise to the input image to simulate the process from a clean image to pure noise. Through the backward process: based on the unified conditional vector, predict the noise residual at each step through the conditional denoising network to guide the denoising process, so that the generated latent space gradually tends to the image that conforms to the rectification specifications.

4. The method for reviewing the rectification of safety hazards in urban rail transit construction projects according to claim 3, characterized in that, Step S31 includes performing a forward diffusion process using Formula 1. The forward diffusion process is a process of gradually adding noise from a compliant image to a noisy image. Formula 1 includes: (Formula 1) in, The image representation at time step t; The image representation at time step t-1; To represent the normal distribution; The noise gain coefficient at time step t is used to control the amount of noise added at each step. It is an identity matrix used to maintain the independence of each dimension of the image.

5. The method for reviewing the rectification of safety hazards in urban rail transit construction projects according to claim 4, characterized in that, Step S32 includes performing a reverse generation process using Formula 2. The reverse generation process is a process of gradually restoring a compliant image from noise. Formula 2 includes: (Formula 2) Among them, P θ To represent the probability distribution controlled by parameter θ; x t-1 The image representation at time step t-1, i.e., the denoised image; x t Let be the image representation at time step t, i.e., the current noisy image; s is the rectification specification vector, containing the standard information required to convert the hazard image into a compliance image; μ θ The mean function, determined by the model parameter θ, is used to predict the mean of the denoised image; Σ θ θ is the covariance function, determined by the model parameter θ, used to predict the covariance of the denoised image; t is the time step.

6. The method for reviewing the rectification of safety hazards in urban rail transit construction projects according to claim 3, characterized in that, Step S33 includes setting a training objective using Formula 3, wherein the training objective is to minimize the diffusion prediction error, and a weighted mean square error loss is used. Formula 3 includes: (Formula 3) Among them, L diffusion Let be the loss function for the diffusion model; E is the expectation operator, representing the expectation value for all possible x0, t, and t. The average; x0 is the original compliant image; t is the time step; The noise sampled from the standard normal distribution N(0,I); The noise predicted by the model is determined by the model parameter θ; |.| 2 The squared error is used to calculate the difference between the predicted noise and the actual noise.

7. The method for reviewing the rectification of safety hazards in urban rail transit construction projects according to claim 1, characterized in that, In step S4, the comparative analysis based on image semantic consistency measurement includes: a consistency measurement step based on perceptual loss, which is implemented through formula four: (Formula 4) in, The predicted image, i.e., the image after rectification; To generate a rectified image using a preceding generative model, This represents the feature map of the l-th layer extracted using the VGG model. , , These represent the number of channels and the size of the feature map at layer l, respectively. This represents the set of all intermediate features.

8. The method for reviewing the rectification of safety hazards in urban rail transit construction projects according to claim 1, characterized in that, In step S4, the comparative analysis based on image semantic consistency measurement includes: a measurement based on segmentation semantic consistency, which is implemented through formula five: (Formula 5) in, The cross-union ratio (CUI) of segmentation between the rectified image and the reference image is measured. Measure the structural similarity between the rectified image and the reference image. and The weighting coefficients for the two scores can be set to 1 by default.

9. The method for reviewing the rectification of safety hazards in urban rail transit construction projects according to claim 1, characterized in that, Step S4 further includes: introducing a local relaxation index to adjust the intersection-union ratio calculation formula six, in order to adapt to reasonable local changes during the image segmentation process, wherein formula six includes: (Formula 6) in, express and The number of pixels in the intersection region, express and The number of pixels in the union region, where δ is a non-negative relaxation parameter used to control the tolerance for local variations.

Images