BIM-based construction safety supervision method and system

Abstract

This invention provides a BIM-based method and system for supervising construction safety, relating to the field of construction safety supervision technology. The method includes: acquiring images of the construction site to be processed; obtaining a BIM building model in its current state; determining the first location information of the building structure within the BIM building model in its current state; obtaining risk status information of the building structure in the image to be processed; setting risk areas; acquiring the second location information of the risk areas in the BIM building model in its current state; determining the risk level of the area under construction in the current construction cycle; and determining a safety supervision plan for the next construction cycle. According to this invention, the risk level of a construction area can be determined based on captured images, thereby determining a safety supervision plan. The safety supervision plan for the next construction cycle can be determined based on the risk status of the construction area in the current construction cycle, improving the accuracy and relevance of the determination of the construction safety supervision plan.

Description

Technical Field

[0001] This invention relates to the field of construction safety supervision technology, and in particular to a BIM-based method and system for supervising construction safety. Background Technology

[0002] CN120013233A discloses a closed-loop management method for construction safety risks based on BIM technology, relating to the field of construction safety management technology. This method establishes a BIM model for construction safety risk management, acquires on-site construction data for each construction section, and binds it to the BIM model. An AI intelligent recognition algorithm is used to determine whether there are safety risks in the on-site construction data, extracts the risk data for construction sections with safety risks, matches corresponding rectification content, and sends it to the responsible person. The responsible person completes the rectification according to the content and uploads the rectified data. The rectified data is then judged using an AI intelligent recognition algorithm or manually. If the rectification is complete, the distinguished display is removed; if the rectification is incomplete, an early warning is issued. This solves the problem that existing construction safety risk management methods cannot achieve closed-loop management of risk sources. This invention is applicable to closed-loop management of construction safety.

[0003] CN119831167A discloses a BIM-based method and system for predicting safety risks at construction sites. By optimizing the data flow processing, it ensures that data acquired from data acquisition devices can be quickly and accurately transmitted to the central processing unit and matched with static information in the multidimensional information model. This method not only improves the timeliness of updating the location, status, and relationships of construction elements in the multidimensional information model but also enhances the accuracy of analyzing potential safety hazards. Ultimately, by embedding the established risk mitigation measures into the multidimensional information model, it can more effectively guide on-site operations, ensuring that all activities are carried out in accordance with the latest safety standards while flexibly responding to dynamic changes. This method significantly improves the real-time performance and effectiveness of construction site safety management and reduces the likelihood of safety accidents.

[0004] According to relevant technologies, intelligent supervision of construction safety can be carried out, but there is a lack of effective means to determine the degree of safety risk in the construction area, making it impossible to accurately judge the risk level of the construction area and generate corresponding safety supervision plans.

[0005] The information disclosed in the background section of this application is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] This invention provides a BIM-based method for supervising building construction safety, which can solve the technical problem that related technologies have difficulty in accurately determining the risk level of a construction area.

[0007] According to a first aspect of the present invention, a BIM-based method for supervising construction safety is provided, comprising:

[0008] Acquire images of the construction site to be processed at the end of the current construction cycle;

[0009] The building structure in the image to be processed is associated with the BIM building model to obtain the BIM building model in the current state.

[0010] Determine the first location information of the building structure in the current state of the BIM building model;

[0011] The image to be processed is processed using a structural image processing model to obtain risk information about the building structure in the image.

[0012] Based on the risk status information of the building structure and the first location information of the building structure in the current state of the BIM building model, risk areas are set in the current state of the BIM building model.

[0013] Obtain the second location information of the risk area in the current state of the BIM building model;

[0014] Based on the risk status information and the second location information of the risk area, the risk level of the area where construction is being carried out in the current construction cycle is determined;

[0015] Based on the risk level of the area where construction is being carried out during the current construction cycle, and the second location information, a safety supervision plan for the next construction cycle is determined.

[0016] According to the present invention, obtaining risk status information of building structures in an image to be processed includes:

[0017] Group images with the same structure into a processing image group;

[0018] Based on the shooting angle of each image in the image group, obtain the angle code of the image to be processed;

[0019] The first feature vector of each image in the image group to be processed is obtained by using the feature extraction layer of the structural image processing model.

[0020] The angle code of the image to be processed is concatenated with the first feature vector to obtain the second feature vector of the image to be processed.

[0021] The connection weights between nodes corresponding to each image to be processed are obtained through the attention mechanism of the structural image processing model.

[0022] Based on the connection weights, the second feature vectors corresponding to each image to be processed are aggregated to obtain the third feature vector of each image to be processed.

[0023] The third feature vector is input into multiple first fully connected layers to obtain multiple risk description values;

[0024] Multiple risk description values ​​are input into the first activation layer to obtain the probability information of the existence of various security risks.

[0025] Based on the existence probability information, determine the risk type information corresponding to each image to be processed;

[0026] Obtain the union of risk type information corresponding to each image in the image group to be processed, and obtain the risk status information of the building structure.

[0027] According to the present invention, the training steps of the structural image processing model include:

[0028] Training images with the same structure are grouped into training image groups, and the angular encoding of each training image in the training image group is obtained;

[0029] The first training feature vector of each training image is obtained through the feature extraction layers of the structured image processing model.

[0030] The angle code is concatenated with the first training feature vector to obtain the second training feature vector for each training image;

[0031] By using the attention mechanism of the structural image processing model, the connection weights between nodes corresponding to each training image are obtained;

[0032] Based on the connection weights, the second training feature vectors corresponding to each training image are aggregated to obtain the third training feature vector of each training image.

[0033] The third training feature vector is input into multiple first fully connected layers to obtain multiple training risk description values;

[0034] Multiple training risk description values ​​are input into the first activation layer to obtain training probability information for various security risks.

[0035] Based on the angle encoding, the risk type labeling information of the training images, and the training probability information, the first loss function of the structured image processing model is determined;

[0036] The structure image processing model is trained based on the first loss function to obtain the trained structure image processing model.

[0037] According to the present invention, determining the first loss function of the structural image processing model includes:

[0038] According to the formula ,

[0039] Determine the first loss function for the structural image processing model. ,in, This provides the training probability information for the j-th security risk in the k-th training image. This represents the maximum probability of the existence of the j-th security risk, determined based on risk type annotations from multiple training images. Encode the angle of the k-th training image. An angular encoding for the training image that maximizes the probability of the existence of the j-th security risk, determined based on risk type annotation information from multiple training images. for and The maximum similarity, where N is the number of types of security risks, M is the number of training images, j≤N, k≤M, and j, N, k, and M are all positive integers.

[0040] According to the present invention, determining the risk level of an area where construction is being carried out during the current construction cycle includes:

[0041] Based on the second location information of the risk area, determine the centroid location information of the risk area;

[0042] Based on the risk status information, centroid location information, and area of ​​the risk region, determine the input information of the node corresponding to each risk region, as well as the adjacency matrix of the graph structure composed of multiple nodes;

[0043] Using a graph neural network model, the output information of the node corresponding to each risk region is obtained from the input information and the adjacency matrix;

[0044] The output information of each node is weighted and summed to obtain the risk characteristics information of the construction area.

[0045] The risk characteristic information of the construction area is input into the second fully connected layer and the second activation layer to obtain the risk characteristic value;

[0046] The risk characteristic value is rounded up to obtain the risk level of the area where construction is being carried out in the current construction cycle.

[0047] According to the present invention, determining the input information of the node corresponding to each risk region, and the adjacency matrix of the graph structure composed of multiple nodes, includes:

[0048] The risk status information, centroid location information, and area of ​​the risk region are combined to form a risk description vector for the risk region.

[0049] The risk description vector is input into the third fully connected layer to obtain the input information of the node corresponding to each risk region;

[0050] Based on the risk status information, centroid location information, and second location information of the two risk areas, determine the weight data between the nodes corresponding to the two risk areas.

[0051] The adjacency matrix is ​​obtained based on the weight data.

[0052] According to the present invention, determining the weight data between nodes corresponding to two risk areas includes:

[0053] According to the formula ,

[0054] Determine the weight data between the node corresponding to the s-th risk region and the node corresponding to the t-th risk region. and ,in, For the s-th risk area, For the t-th risk area, for and The area of ​​their intersection. for and The area of ​​the union of the two sets. For the risk status information of the s-th building structure, For the risk status information of the t-th building structure, for and similarity, Let be the distance between the centroid locations of the s-th risk region and the t-th risk region. Let s and t be the average distance between the centroid locations of each risk area, where s and t are positive integers.

[0055] According to a second aspect of the present invention, a BIM-based construction safety monitoring system is provided, comprising:

[0056] The image to be processed module acquires images of the construction site to be processed at the end of the current construction cycle;

[0057] The building model module associates the building structure in the image to be processed with the BIM building model to obtain the BIM building model in the current state.

[0058] The first location information module determines the first location information of the building structure in the current state of the BIM building model;

[0059] The risk status information module processes the image to be processed using a structural image processing model to obtain risk status information of the building structure in the image to be processed.

[0060] The risk area module sets risk areas in the current state of the BIM building model based on the risk status information of the building structure and the first location information of the building structure in the current state of the BIM building model.

[0061] The second location information module obtains the second location information of the risk area in the current state of the BIM building model;

[0062] The risk level module determines the risk level of the area to be constructed in the current construction cycle based on the risk status information and the second location information of the risk area.

[0063] The safety supervision plan module determines the safety supervision plan for the next construction cycle based on the risk level of the area under construction in the current construction cycle and the second location information.

[0064] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0065] According to the present invention,

[0066] At the end of the current construction cycle, images of the construction site awaiting processing can be acquired and associated with the BIM building model to obtain the current state of the BIM building model. Risk areas can be set in the BIM building model based on the images to determine the risk level of the construction area, thereby determining a safety supervision plan. Based on the risk status of the construction area in the current construction cycle, a safety supervision plan for the next construction cycle can be determined, improving the accuracy and relevance of the safety supervision plan and enhancing its real-time performance and efficiency. Furthermore, the images of the construction site awaiting processing at the end of the construction cycle can be associated with the BIM building model to obtain the current state of the BIM building model, thus determining the primary location of the building structure within the current state of the BIM building model. The safety status of the building structure can be monitored in real time, providing basic data for determining the safety supervision plan for the next construction cycle. When training a structural image processing model, the similarity of shooting angles between images can be determined by encoding the angles of the training images. This leads to the determination of the similarity of safety risks based on the images. Weights are then assigned to the cross-entropy loss function corresponding to the training probability information of various safety risks for each training image, resulting in the loss function of the structural image processing model. This process improves the accuracy and relevance of the training, thus enhancing the performance of the structural image processing model. Furthermore, weight data between nodes can be determined based on risk status information and the second location information of risk areas. Through graph neural network (GNN) model processing, the risk level of the area under construction in the current construction cycle can be determined. During GNN model training, the loss function is determined based on the rounding-up characteristic. This process further enhances the performance of the GNN model, improving the accuracy, comprehensiveness, and objectivity of risk level determination. Moreover, based on the risk level of the area under construction in the current construction cycle and the second location information, a safety supervision plan for the next construction cycle can be determined. This improves the accuracy and relevance of the construction safety supervision plan, enhancing the real-time performance and efficiency of safety supervision.

[0067] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Other features and aspects of the invention will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0069] Figure 1 An exemplary flowchart of a BIM-based construction safety supervision method according to an embodiment of the present invention is shown.

[0070] Figure 2 An exemplary flowchart illustrates a process for obtaining risk status information of a building structure in an image to be processed according to an embodiment of the present invention;

[0071] Figure 3 A block diagram of a BIM-based building construction safety supervision system according to an embodiment of the present invention is shown as an example. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0074] Figure 1 An exemplary flowchart of a BIM-based construction safety supervision method according to an embodiment of the present invention is shown, the method comprising:

[0075] Step S1: At the end of the current construction cycle, acquire the image of the construction site to be processed;

[0076] Step S2: Associate the building structure in the image to be processed with the BIM building model to obtain the BIM building model in the current state.

[0077] Step S3: Determine the first location information of the building structure in the current state of the BIM building model;

[0078] Step S4: Process the image to be processed using a structural image processing model to obtain risk status information of the building structure in the image to be processed.

[0079] Step S5: Based on the risk status information of the building structure and the first location information of the building structure in the current state of the BIM building model, set up risk areas in the current state of the BIM building model.

[0080] Step S6: Obtain the second location information of the risk area in the current state of the BIM building model;

[0081] Step S7: Based on the risk status information and the second location information of the risk area, determine the risk level of the area where construction is being carried out in the current construction cycle;

[0082] Step S8: Based on the risk level of the area under construction in the current construction cycle and the second location information, determine the safety supervision plan for the next construction cycle.

[0083] The BIM-based construction safety supervision method according to embodiments of the present invention can acquire images of the construction site to be processed at the end of the current construction cycle, associate them with the BIM building model to obtain the current state of the BIM building model, and set risk areas in the BIM building model based on the images to be processed to determine the risk level of the construction area, thereby determining a safety supervision plan. Based on the risk situation of the construction area in the current construction cycle, a safety supervision plan for the next construction cycle can be determined, improving the accuracy and relevance of the determination of the construction safety supervision plan, and enhancing the real-time performance and efficiency of safety supervision.

[0084] According to an embodiment of the present invention, in step S1, images of the construction site to be processed are acquired at the end of the current construction cycle. A construction cycle can be the time period required to complete a partial construction target, such as the time period required to construct a building structure or a floor. The lengths of each construction cycle can be different. At the end of the current construction cycle, images of various building structures (e.g., reinforced concrete structures, steel structures, etc.) at the construction site from various angles can be captured using photographic equipment such as handheld cameras or cameras mounted on drones; these images are the images to be processed. For example, after a room is completed, images of the walls, beams, columns, and stairs within the room from multiple angles can be captured using a camera to obtain the images to be processed. The present invention does not limit this.

[0085] According to an embodiment of the present invention, in step S2, the building structure in the image to be processed is associated with the BIM building model to obtain the BIM building model in its current state. For example, the image to be processed is identified to determine the type of the component in the image, so that the component can be found in the component database of the BIM model, and the BIM building model is updated using the component. This associates the building structure in the image to be processed with the BIM building model. Furthermore, after updating the BIM building model, the BIM building model at the end of the current construction cycle can be obtained, which is the BIM building model in its current state.

[0086] According to an embodiment of the present invention, in step S3, the first position information of the building structure in the current state of the BIM building model is determined. The first position information may be the coordinate information of the building structure in the model coordinate system of the BIM building model, and the first position information is the virtual position information of the building structure in the BIM building model.

[0087] In this way, the images of the construction site at the end of the construction cycle can be linked with the BIM building model to obtain the current state of the BIM building model, thereby determining the primary location of the building structure within the current state of the BIM building model. This allows for real-time monitoring of the building structure's safety status, providing fundamental data for determining the safety supervision plan for the next construction cycle.

[0088] Figure 2 An exemplary flowchart illustrating the process of obtaining risk status information of a building structure in an image to be processed according to an embodiment of the present invention is shown.

[0089] According to an embodiment of the present invention, in step S4, the image to be processed is processed using a structural image processing model to obtain risk status information of the building structure in the image to be processed, including: step S41, grouping images of the same structure into a group of images to be processed; step S42, obtaining the angle code of the image to be processed based on the shooting angle of each image in the group of images to be processed; step S43, obtaining the first feature vector of each image in the group of images to be processed through the feature extraction layers of the structural image processing model; step S44, concatenating the angle code of the image to be processed with the first feature vector to obtain the second feature vector of the image to be processed; step S45, through the annotation of the structural image processing model... The process involves several steps: Step S46: Using a force-based mechanism, the connection weights between nodes corresponding to each image to be processed are obtained; Step S47: Based on the connection weights, the second feature vectors corresponding to each image to be processed are aggregated to obtain the third feature vectors of each image to be processed; Step S48: The third feature vectors are input into multiple first fully connected layers to obtain multiple risk description values; Step S49: The multiple risk description values ​​are input into a first activation layer to obtain the probability information of the existence of various safety risks; Step S40: Based on the probability information, the risk type information corresponding to each image to be processed is determined; Step S410: The union of the risk type information corresponding to each image to be processed in the image group is obtained to obtain the risk status information of the building structure.

[0090] According to an embodiment of the present invention, in step S41, images of the same structure are grouped into a processing image group. Images of the same building structure taken from various angles are grouped into a processing image group. For example, images of the same staircase taken from various angles are grouped into a processing image group corresponding to that staircase.

[0091] According to an embodiment of the present invention, in step S42, an angle code for the image to be processed is obtained based on the shooting angle of each image in the image group to be processed. The shooting angle can be determined based on the angle between the shooting direction and the x-axis, y-axis, and z-axis in a three-dimensional coordinate system. The angle code can be represented in vector form. For example, when photographing a pillar in a three-dimensional coordinate system, if the angles between the camera's optical axis and the x-axis, y-axis, and z-axis are 30°, 60°, and 45° respectively, these angles can be used as the shooting angle. , , This refers to the angle code of the image to be processed. Of course, the angle code of the image to be processed can also be determined based on information such as the sine, cosine, and tangent values ​​of the shooting angle corresponding to each image in the image group. This invention does not impose any limitations on this.

[0092] According to an embodiment of the present invention, in step S43, the first feature vector of each image to be processed in the image group is obtained through the feature extraction layer of the structured image processing model. By extracting features from the image to be processed through the feature extraction layer (e.g., a convolutional layer) of the structured image processing model, a long vector (e.g., a 64-dimensional vector or a 128-dimensional vector) is obtained, which is the first feature vector of the image to be processed, thereby obtaining the first feature vector of each image to be processed in the image group. The first feature vector can describe the features of the corresponding image to be processed.

[0093] According to an embodiment of the present invention, in step S44, the angle code of the image to be processed is concatenated with the first feature vector to obtain the second feature vector of the image to be processed. Concatenating the angle code of the image to be processed (e.g., a 3-dimensional vector) with the first feature vector (e.g., a 64-dimensional vector) yields a new vector (i.e., a 67-dimensional vector), which is the second feature vector.

[0094] According to an embodiment of the present invention, in step S45, the connection weights between nodes corresponding to each image to be processed are obtained through the attention mechanism of the structural image processing model. Each image to be processed can be considered a node in the graph structure, and the connection weights describe the degree of closeness of the relationship between the images to be processed corresponding to two nodes. When determining the connection weights, the graph structure consists of multiple nodes and edges, each node corresponds to an image to be processed, and each edge can represent whether there is a connection relationship between the nodes (i.e., whether the images to be processed corresponding to the nodes are related) and the degree of closeness of the relationship. The input vectors of two nodes with a connection relationship (i.e., the second feature vector of the image to be processed corresponding to the node) are concatenated and processed by a fully connected layer to obtain the corresponding value, which can be used as the connection weight between the two nodes. The parameters of the fully connected layer can be determined through training.

[0095] According to an embodiment of the present invention, in step S46, the second feature vectors corresponding to each image to be processed are aggregated according to the connection weights to obtain the third feature vector of each image to be processed. The second feature vector of each image to be processed and the second feature vectors of images related to it (i.e., the input vectors of nodes) are weighted and summed using the corresponding connection weights to obtain the aggregated vector of the second feature vectors corresponding to each image to be processed, which is the third feature vector (i.e., the output vector of nodes). The third feature vector of the image to be processed aggregates its own features and the features of images related to it, providing a more comprehensive description of the features of the image to be processed. By fusing image features from various angles, it can more comprehensively describe the risks present in the components of the image to be processed. The above processing can be performed on each image to be processed to obtain the third feature vector of each image to be processed.

[0096] According to an embodiment of the present invention, in step S47, the third feature vector is input into multiple first fully connected layers to obtain multiple risk description values. After inputting the third feature vector of the image to be processed into the first fully connected layer, a risk description value corresponding to the safety risk can be obtained, which can describe the probability that the image to be processed has a corresponding safety risk. Each fully connected layer is used to output a risk description value for one type of safety risk. For example, after inputting the third feature vector of the image to be processed into the first fully connected layer corresponding to the fracture risk, a fracture risk description value can be obtained, which can describe the probability that the image to be processed has fractured. Of course, the first fully connected layer may also include a first fully connected layer corresponding to the collapse risk, a first fully connected layer corresponding to the falling object risk, etc., and after inputting the third feature vector into each first fully connected layer, a risk description value for each type of risk can be obtained.

[0097] According to an embodiment of the present invention, in step S48, multiple risk description values ​​are respectively input into the first activation layer to obtain the probability information of the existence of multiple safety risks. By inputting multiple risk description values ​​into the first activation layer, the sigmoid function can be used to map the risk description values ​​to the range of [0,1], thereby obtaining the probability of the existence of various safety risks in the image to be processed, which is the probability information of the existence of multiple safety risks in the image to be processed. For example, by inputting the risk description values ​​of fracture, collapse, and falling objects from height of a certain image to be processed into the first activation layer, the probability information of the existence of fracture, collapse, and falling objects from height in the image to be processed can be obtained.

[0098] According to an embodiment of the present invention, in step S49, risk type information corresponding to each image to be processed is determined based on the existence probability information. When the existence probability information of a security risk is greater than 0.5, it can be considered that the component in the image to be processed has that security risk, and all security risks existing in the component in the image to be processed can be determined as the risk type information corresponding to the image to be processed.

[0099] According to an embodiment of the present invention, in step S410, the union of risk type information corresponding to each image to be processed in the image group to be processed is obtained to obtain the risk status information of the building structure. Since multiple images to be processed are taken for each building structure, and the risk type information corresponding to each image to be processed may be different, the risk status information of the building structure can be determined according to the union of risk type information corresponding to each image to be processed in the image group to be processed (i.e., all safety risks existing in the building structure in the image group to be processed). The risk status information can be represented in vector form, specifying the risk type information corresponding to each element in the vector. For example, a group of images to be processed for a pillar includes three images, and the union of the risk type information corresponding to the three images is collapse or fracture, excluding falling objects from heights. 1 represents a present safety risk, and 0 represents a non-existent safety risk. Therefore, the union of the risk type information corresponding to each image in the group of images to be processed for that pillar can be represented as {1, 1, 0}. When represented in vector form, the first element indicates whether there is a collapse risk, the second element indicates whether there is a fracture risk, and the third element indicates whether there is a falling object risk. Furthermore, this union will be represented as a vector, i.e., (1, 1, 0). This invention does not impose any limitations on this.

[0100] According to an embodiment of the present invention, the training steps of the structural image processing model include: grouping training images of the same structure into a training image group and obtaining the angle code of each training image in the training image group; obtaining a first training feature vector of each training image through the feature extraction layer of the structural image processing model; concatenating the angle code with the first training feature vector to obtain a second training feature vector of each training image; obtaining the connection weights between the nodes corresponding to each training image through the attention mechanism of the structural image processing model; aggregating the second training feature vectors corresponding to each training image according to the connection weights to obtain a third training feature vector of each training image; inputting the third training feature vectors into multiple first fully connected layers to obtain multiple training risk description values; inputting the multiple training risk description values ​​into a first activation layer to obtain training probability information of multiple security risks; determining a first loss function of the structural image processing model according to the angle code, the risk type labeling information of the training images, and the training probability information; and training the structural image processing model according to the first loss function to obtain a trained structural image processing model.

[0101] According to embodiments of the present invention, similar to determining the probability information of multiple risks, training images of the same building structure (images of the building structure taken by a camera from multiple angles) can be grouped into a training image group. Based on the angle between the shooting direction of each training image in the training image group and the coordinate axes of the three-dimensional coordinate system, an angle code for each training image is determined. This angle code can also be represented in vector form. Further, a first training feature vector for each training image can be obtained through the feature extraction layer of the structural image processing model. The angle code is then concatenated with the first training feature vector to obtain a corresponding long vector for each training image, which is the second training feature vector. By using the attention mechanism of the structural image processing model, the connection weights between nodes corresponding to each training image are obtained. Based on the connection weights, the second training feature vectors corresponding to each training image are weighted and summed (i.e., aggregated) to obtain the third training feature vectors of each training image. The third training feature vectors are then input into the first fully connected layer responsible for outputting risk description values ​​for different security risks to obtain multiple training risk description values. These multiple training risk description values ​​are then input into the first activation layer to map the risk description values ​​to the range [0,1]. This can describe the probability of various security risks existing in the training image, i.e., the training probability information of various security risks of the training image.

[0102] According to an embodiment of the present invention, determining the first loss function of the structural image processing model based on the angle encoding, the risk type labeling information of the training image, and the training probability information includes: determining the first loss function of the structural image processing model according to formula (1). , (1),

[0103] in, This provides the training probability information for the j-th security risk in the k-th training image. This represents the maximum probability of the existence of the j-th security risk, determined based on risk type annotation information from multiple training images. Encode the angle of the k-th training image. An angular encoding for the training image that maximizes the probability of the existence of the j-th security risk, determined based on risk type annotation information from multiple training images. for and The maximum similarity, where N is the number of types of security risks, M is the number of training images, j≤N, k≤M, and j, N, k, and M are all positive integers.

[0104] According to an embodiment of the present invention, in formula (1), The maximum value of the probability of the existence of the j-th security risk determined based on the risk type annotation information of multiple training images can be used as the probability of the existence of the j-th security risk determined based on the risk type annotation information of multiple training images. It is either 0 or 1. Since multiple training images were obtained for a single building structure, and each training image was taken from a different direction, when the building structure has a type j safety risk, it may not be possible to determine the existence of the type j safety risk using the risk type annotation information of the k-th training image. Instead, the existence of the type j safety risk is determined using the risk type annotation information of other training images in the corresponding training image group. Among the probabilities of the existence of the type j safety risk determined based on the risk type annotation information of multiple training images, at least one probability is 1 (i.e., ...). When the probability of the j-th safety risk is 1, the building structure can be considered to have the j-th safety risk. For example, for a staircase, if the probability of the fracture risk is 0 based on the risk type annotation information of the first two training images in the corresponding training image group, and the probability of the fracture risk is 1 based on the risk type annotation information of the third training image in the corresponding training image group, then the staircase can be considered to have the fracture risk. When the probability of the j-th safety risk is 0 based on the risk type annotation information of multiple training images, the building structure can be considered not to have the j-th safety risk.

[0105] According to an embodiment of the present invention, in formula (1), when the probability of the existence of the j-th security risk determined based on the risk type annotation information of multiple training images is 1, The value is 1 when the probability of the j-th security risk, determined based on risk type annotation information from multiple training images, is 0. The value is 0. For based on and The cross-entropy loss function can be adjusted during training to minimize the parameters of the structured image processing model, thereby improving its performance. and The error between them is reduced.

[0106] According to an embodiment of the present invention, in formula (1), The similarity (e.g., cosine similarity) between the angular code of the k-th training image and the angular code of the training image that maximizes the probability of the j-th security risk determined from risk type annotations of multiple training images can be used as a candidate value for the weight coefficient of the cross-entropy loss function corresponding to the k-th training image. Each candidate weight coefficient can describe the importance of the k-th training image. Since there may be multiple training images in the training image set where the probability of the j-th security risk is 1, the maximum similarity between the angular code of the k-th training image and the angular code of the training image that maximizes the probability of the j-th security risk determined from risk type annotations of multiple training images is used as the weight coefficient of the cross-entropy loss function corresponding to the k-th training image. The weight coefficients described above can be used to describe the importance of the k-th training image. The larger the value of the weight coefficients, the more important the k-th training image is. That is, the greater the maximum similarity between the angle code of the k-th training image and the angle code of the training image that achieves the maximum probability of the existence of the j-th security risk determined based on the risk type annotation information of multiple training images, the closer the shooting angles of the k-th training image and the training image that achieves the maximum probability of the existence of the j-th security risk are, and the higher their similarity is. Therefore, the closer the probability of the existence of the j-th security risk determined based on the training probability information of the k-th training image is to the maximum probability of the existence of the j-th security risk determined based on the risk type annotation information of multiple training images, the higher the weight coefficients should be. The first loss function can be obtained by weighted summing the cross-entropy loss functions corresponding to various types of risks in each training image using the above weight coefficients.

[0107] According to an embodiment of the present invention, a structure image processing model is trained based on a first loss function to obtain a trained structure image processing model. Backpropagation can be performed using the loss function, and the parameters of the structure image processing model can be adjusted using gradient descent to train the model. After multiple training iterations (i.e., training using multiple training images), training is complete, resulting in a trained structure image processing model.

[0108] In this way, when training a structural image processing model, the similarity of shooting angles between images can be determined by encoding the angles of the training images, thereby determining the similarity of security risks based on the images. Weights are then set for the cross-entropy loss function corresponding to the training probability information of various security risks for each training image, resulting in the loss function of the structural image processing model. The trained structural image processing model is then obtained, improving the accuracy and relevance of the training and enhancing the performance of the structural image processing model.

[0109] According to an embodiment of the present invention, in step S5, a risk area is set in the current state BIM building model based on the risk status information of the building structure and the first location information of the building structure in the current state BIM building model. For example, if the risk information of the building structure includes the risk of falling objects from the construction scaffold, a circular area with a preset radius directly below the first location information of the scaffold in the current state BIM building model can be determined as the risk area in the current state BIM building model. The radius of the circular area can be determined according to the scope of the safety risk of the building structure. For example, a circular area with a radius of 3 meters directly below the scaffold with the risk of falling objects from the building is the corresponding risk area in the current state BIM building model. The present invention does not limit this.

[0110] According to an embodiment of the present invention, in step S6, second location information of the risk area in the current state of the BIM building model is obtained. The second location information may be the coordinate information of the risk area in the model coordinates of the BIM building model, and the second location information is the virtual location information of the risk area in the BIM building model.

[0111] According to an embodiment of the present invention, in step S7, determining the risk level of the area under construction in the current construction cycle based on the risk status information and the second location information of the risk area includes: determining the centroid location information of the risk area based on the second location information of the risk area; determining the input information of the node corresponding to each risk area and the adjacency matrix of the graph structure composed of multiple nodes based on the risk status information, the centroid location information and the area of ​​the risk area; obtaining the output information of the node corresponding to each risk area through the graph neural network model using the input information and the adjacency matrix; performing weighted summation on the output information of each node to obtain the risk feature information of the construction area; inputting the risk feature information of the construction area into the second fully connected layer and the second activation layer to obtain the risk feature value; and rounding the risk feature value up to obtain the risk level of the area under construction in the current construction cycle.

[0112] According to an embodiment of the present invention, the centroid location information of the risk area is determined based on the second location information of the risk area. When determining the centroid location information of the risk area, since the risk area may be an irregular shape, it can be divided into multiple simple shapes (e.g., circles, rectangles, etc.), and a weighted average is performed on the centroid of each simple shape and its area to obtain the centroid location of the risk area. The centroid of the simple shape can be determined based on the coordinate information of each point in the simple shape in the model coordinates of the BIM building model. The centroid location information of the risk area is the virtual location information of the centroid in the BIM building model.

[0113] According to an embodiment of the present invention, determining the input information of the node corresponding to each risk region and the adjacency matrix of the graph structure composed of multiple nodes based on the risk status information, centroid location information, and area of ​​the risk region includes: forming a risk description vector of the risk region from the risk status information, centroid location information, and area of ​​the risk region; inputting the risk description vector into a third fully connected layer to obtain the input information of the node corresponding to each risk region; determining the weight data between the nodes corresponding to the two risk regions based on the risk status information, centroid location information, and second location information of the two risk regions; and obtaining the adjacency matrix based on the weight data.

[0114] According to an embodiment of the present invention, each risk region can be used as a node in a graph structure. The risk status information, centroid location information, and area of ​​the risk region are combined to form a risk description vector for the risk region. By representing the risk status information, centroid location information, and area of ​​the risk region as vectors and concatenating them, a short vector is obtained, which is the risk description vector of the risk region. For example, the risk status information, centroid location information, and area of ​​a certain risk region can be represented by vectors (1, 1, 0), (2, 5), and (10). After concatenation, the risk description vector of the risk region (1, 1, 0, 2, 5, 10) is obtained. After inputting the risk description vector into the third fully connected layer, a long vector is obtained, which is the input information of the node corresponding to each risk region.

[0115] According to an embodiment of the present invention, the connection relationship between nodes can be determined by an adjacency matrix. Based on the risk status information, centroid location information, and second location information of two risk regions, the weight data between the nodes corresponding to the two risk regions is determined, including: determining the weight data between the node corresponding to the s-th risk region and the node corresponding to the t-th risk region according to formula (2). and , (2),

[0116] in, For the s-th risk area, For the t-th risk area, for and The area of ​​their intersection. for and The area of ​​the union of the two sets. For the risk status information of the s-th building structure, For the risk status information of the t-th building structure, for and similarity, Let be the distance between the centroid locations of the s-th risk region and the t-th risk region. Let s and t be the average distance between the centroid locations of each risk area, where s and t are positive integers.

[0117] According to an embodiment of the present invention, in formula (2), The area of ​​the intersection of the s-th risk region and the t-th risk region is the ratio of the area of ​​the union of the s-th risk region and the t-th risk region. It can be considered as the degree of overlap between the s-th risk region and the t-th risk region. This represents the similarity (e.g., cosine similarity) between the risk status information of the s-th building structure and the risk status information of the t-th building structure. The maximum value of the two values ​​represents the similarity between the s-th risk region and the t-th risk region, and is used to describe the closeness of the relationship between the s-th risk region and the t-th risk region.

[0118] According to an embodiment of the present invention, in formula (2), This represents the ratio of the distance between the centroid locations of the s-th and t-th risk regions to the average distance between the centroid locations of all risk regions. This ratio can be used to adjust the similarity between the s-th and t-th risk regions. When When the value is greater than 1, the s-th risk region is far from the t-th risk region, which can reduce the similarity between the s-th and t-th risk regions. Conversely, when the value is less than 1, the similarity between the s-th and t-th risk regions decreases. When the distance is less than 1, the s-th risk region is closer to the t-th risk region, which can increase the similarity between the s-th and t-th risk regions. Therefore... The similarity between the adjusted s-th risk region and the t-th risk region can be represented by the weight data between the nodes corresponding to the s-th and t-th risk regions. The larger the weight data, the closer the relationship between the s-th and t-th risk regions. Since the graph structure of the risk regions is an undirected graph, therefore... Furthermore, the graph structure of the risk region differs from that of the image to be processed.

[0119] According to an embodiment of the present invention, the adjacency matrix is ​​obtained based on the weight data between each node. Since the weight data of all s-th risk regions and themselves in the adjacency matrix is ​​0 (i.e., the diagonal data of the adjacency matrix is ​​0), the adjusted adjacency matrix can be obtained by summing the adjacency matrix and the bias matrix, which can enhance the expressive power of the graph neural network model. In the adjacency matrix, the element in the s-th row and t-th column is the weight data between the node corresponding to the s-th risk region and the node corresponding to the t-th risk region, and each node in this graph structure corresponds to a risk region.

[0120] According to an embodiment of the present invention, the output information of the node corresponding to each risk region is obtained by processing the input information and the adjacency matrix using a graph neural network model. The degree matrix, which is a diagonal matrix, is obtained from the adjacency matrix. Since the graph structure corresponding to the risk region is an undirected graph, and the adjacency matrix is ​​a symmetric matrix, the degree of the i-th vertex in the degree matrix is ​​the sum of the elements in the i-th row or i-th column of the adjacency matrix. For example, in the degree matrix, the degree of the first vertex is the sum of the elements in the first row of the adjacency matrix, the degree of the second vertex is the sum of the elements in the second row of the adjacency matrix, and so on, thus obtaining the degree matrix corresponding to the adjacency matrix. By processing the input information, the adjacency matrix, and the degree matrix using a graph neural network model (e.g., a GCN model), the output information of the node corresponding to each risk region can be obtained. Features of risk regions that are related to themselves can be aggregated to provide a more comprehensive description of the characteristics of the risk regions.

[0121] According to an embodiment of the present invention, the output information of each node is weighted and summed to obtain the risk characteristic information of the construction area. The weights in the weighted summation process can be determined based on the ratio of the number of safety risks in each risk area to the total number of safety risks in the construction area. For example, if a risk area has 5 safety risks and the total number of safety risks in the construction area is 10, the weight of the output information of the node corresponding to that risk area can be set as follows: Furthermore, by performing weighted summation on the output information of each node, the risk characteristic information of the construction area can be obtained. This allows for the aggregation of characteristics of all risk areas within the construction area, providing a more comprehensive description of the safety risks in the construction area.

[0122] According to an embodiment of the present invention, the risk characteristic information of the construction area is input into a second fully connected layer and a second activation layer to obtain risk characteristic values. The risk characteristic information of the construction area is then processed by the second fully connected layer and the second activation layer (e.g., a ReLU function) to obtain the risk characteristic values. The second activation layer can adjust risk characteristic values ​​less than zero to 0, while retaining the original values ​​of risk characteristic values ​​greater than zero. The parameters of the second fully connected layer can be determined through training.

[0123] According to an embodiment of the present invention, when training a graph neural network model, a loss function can be determined based on the training risk feature values ​​and labeled risk levels of each training construction area. In the loss function, the training risk feature value can be represented by R, and the labeled risk level can be represented by L. Since the corresponding training risk level can be obtained by rounding up the training risk feature value, when R ≤ L and LR is less than 1, that is, when the training risk level determined based on the training risk feature value is the same as the labeled risk level, the training risk feature value can be considered correct, with only errors present. Therefore, the loss function can be determined as follows: When R < L and LR > 1, or R > L, meaning the training risk level determined based on the training risk feature value differs from the labeled risk level, it can be considered that the training risk feature value is incorrect. In this case, the output information of the node corresponding to each risk region can be input into the second fully connected layer and the second activation layer to obtain the risk feature value of each node. Since the maximum value among the risk feature values ​​of each node (which can be used...) The risk characteristic value of the construction area has the greatest impact (i.e., the closest relationship) on the risk feature value of the construction area; therefore, the loss function can be determined as follows: Based on the aforementioned loss function, a graph neural network model is trained. Backpropagation is performed using the loss function, and gradient descent is used to adjust the parameters of the graph neural network model. After multiple training iterations (i.e., training using multiple training regions), the training is complete, yielding the trained graph neural network model.

[0124] According to an embodiment of the present invention, the risk characteristic value is rounded up to obtain the risk level of the area under construction in the current construction cycle. Since the risk characteristic information may be non-integer (e.g., 0.3, 1.2, etc.), the risk characteristic value can be rounded up, and the rounded value is the risk level of the area under construction in the current construction cycle. For example, if the risk characteristic information is 1.2, rounding up yields a risk level of 2 for the area under construction in the current construction cycle. Furthermore, a risk level upper limit can be set. For example, setting the risk level upper limit to 3 means that if the risk level determined based on the risk characteristic information is greater than 3, the risk level is determined to be 3. The present invention does not impose any limitations on this.

[0125] In this way, weight data between nodes can be determined based on risk status information and the secondary location information of the risk area. Then, through processing by a graph neural network model, the risk level of the area under construction in the current construction cycle can be determined. During the training of the graph neural network model, the loss function is determined based on the rounding-up characteristic. This results in a trained graph neural network model, improving its performance and enhancing the accuracy, comprehensiveness, and objectivity of risk level determination.

[0126] According to an embodiment of the present invention, in step S8, a safety supervision plan for the next construction cycle is determined based on the risk level of the area under construction in the current construction cycle and the second location information. The safety supervision plan may include safety supervision measures such as fencing off certain locations to allow only designated workers to enter, and placing safety signs. For example, if the risk level of the area under construction in the current construction cycle is level one, the corresponding location needs to be fenced off; if the risk level is level two, the corresponding location needs to be fenced off and only designated workers are allowed to enter; if the risk level is level three, only designated workers are allowed to enter the construction site.

[0127] This method allows for the determination of safety supervision plans for the next construction cycle based on the risk level of the area under construction during the current construction period and secondary location information. This improves the accuracy and relevance of the determined construction safety supervision plans, and enhances the real-time nature and efficiency of safety supervision.

[0128] The BIM-based construction safety supervision method according to embodiments of the present invention can acquire images of the construction site to be processed at the end of the current construction cycle, associate them with the BIM building model to obtain the current state of the BIM building model, and set risk areas in the BIM building model based on the images to be processed to determine the risk level of the construction area, thereby determining a safety supervision plan. Based on the risk situation of the construction area in the current construction cycle, a safety supervision plan for the next construction cycle can be determined, improving the accuracy and relevance of the determination of the construction safety supervision plan, and enhancing the real-time performance and efficiency of safety supervision. Furthermore, the images of the construction site to be processed at the end of the construction cycle can be associated with the BIM building model to obtain the current state of the BIM building model, thereby determining the first position information of the building structure in the current state of the BIM building model. The safety status of the building structure can be monitored in real time, providing basic data for determining the safety supervision plan for the next construction cycle. When training a structural image processing model, the similarity of shooting angles between images can be determined by encoding the angles of the training images. This leads to the determination of the similarity of safety risks based on the images. Weights are then assigned to the cross-entropy loss function corresponding to the training probability information of various safety risks for each training image, resulting in the loss function of the structural image processing model. This process improves the accuracy and relevance of the training, thus enhancing the performance of the structural image processing model. Furthermore, weight data between nodes can be determined based on risk status information and the second location information of risk areas. Through graph neural network (GNN) model processing, the risk level of the area under construction in the current construction cycle can be determined. During GNN model training, the loss function is determined based on the rounding-up characteristic. This process further enhances the performance of the GNN model, improving the accuracy, comprehensiveness, and objectivity of risk level determination. Moreover, based on the risk level of the area under construction in the current construction cycle and the second location information, a safety supervision plan for the next construction cycle can be determined. This improves the accuracy and relevance of the construction safety supervision plan, enhancing the real-time performance and efficiency of safety supervision.

[0129] Figure 3 An exemplary block diagram of a BIM-based construction safety supervision system according to an embodiment of the present invention is shown, the system comprising:

[0130] The image to be processed module acquires images of the construction site to be processed at the end of the current construction cycle;

[0131] The building model module associates the building structure in the image to be processed with the BIM building model to obtain the BIM building model in the current state.

[0132] The first location information module determines the first location information of the building structure in the current state of the BIM building model;

[0133] The risk status information module processes the image to be processed using a structural image processing model to obtain risk status information of the building structure in the image to be processed.

[0134] The risk area module sets risk areas in the current state of the BIM building model based on the risk status information of the building structure and the first location information of the building structure in the current state of the BIM building model.

[0135] The second location information module obtains the second location information of the risk area in the current state of the BIM building model;

[0136] The risk level module determines the risk level of the area to be constructed in the current construction cycle based on the risk status information and the second location information of the risk area.

[0137] The safety supervision plan module determines the safety supervision plan for the next construction cycle based on the risk level of the area under construction in the current construction cycle and the second location information.

[0138] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0139] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A BIM-based method for supervising construction safety, characterized in that, include: Acquire images of the construction site to be processed at the end of the current construction cycle; The building structure in the image to be processed is associated with the BIM building model to obtain the BIM building model in the current state. Determine the first location information of the building structure in the current state of the BIM building model; The image to be processed is processed using a structural image processing model to obtain risk information about the building structure in the image to be processed. Based on the risk status information of the building structure and the first location information of the building structure in the current state of the BIM building model, risk areas are set in the current state of the BIM building model. Obtain the second location information of the risk area in the current state of the BIM building model; Based on the risk status information and the second location information of the risk area, the risk level of the area where construction is being carried out in the current construction cycle is determined; Based on the risk level of the area where construction is being carried out during the current construction cycle, and the second location information, determine the safety supervision plan for the next construction cycle; The image to be processed is processed using a structural image processing model to obtain risk status information of the building structure in the image, including: Group images with the same structure into a processing image group; Based on the shooting angle of each image in the image group, obtain the angle code of the image to be processed; The first feature vector of each image in the image group to be processed is obtained by using the feature extraction layer of the structural image processing model. The angle code of the image to be processed is concatenated with the first feature vector to obtain the second feature vector of the image to be processed. The connection weights between nodes corresponding to each image to be processed are obtained through the attention mechanism of the structural image processing model. Based on the connection weights, the second feature vectors corresponding to each image to be processed are aggregated to obtain the third feature vector of each image to be processed. The third feature vector is input into multiple first fully connected layers to obtain multiple risk description values; Multiple risk description values ​​are input into the first activation layer to obtain the probability information of the existence of various security risks. Based on the existence probability information, determine the risk type information corresponding to each image to be processed; Obtain the union of risk type information corresponding to each image in the image group to be processed, and obtain the risk status information of the building structure. Based on the risk status information and the second location information of the risk area, the risk level of the area where construction is being carried out in the current construction cycle is determined, including: Based on the second location information of the risk area, determine the centroid location information of the risk area; Based on the risk status information, centroid location information, and area of ​​the risk region, determine the input information of the node corresponding to each risk region, as well as the adjacency matrix of the graph structure composed of multiple nodes; The input information and the adjacency matrix are processed using a graph neural network model to obtain the output information of the node corresponding to each risk area; The output information of each node is weighted and summed to obtain the risk characteristics information of the construction area. The risk characteristic information of the construction area is input into the second fully connected layer and the second activation layer to obtain the risk characteristic value; The risk characteristic value is rounded up to obtain the risk level of the area where construction is being carried out in the current construction cycle.

2. The BIM-based construction safety supervision method according to claim 1, characterized in that, Based on the risk status information, centroid location information, and area of ​​the risk region, determine the input information of the node corresponding to each risk region, as well as the adjacency matrix of the graph structure composed of multiple nodes, including: The risk status information, centroid location information, and area of ​​the risk region are combined to form a risk description vector for the risk region. The risk description vector is input into the third fully connected layer to obtain the input information of the node corresponding to each risk region; Based on the risk status information, centroid location information, and second location information of the two risk areas, determine the weight data between the nodes corresponding to the two risk areas. The adjacency matrix is ​​obtained based on the weight data; Based on the risk status information, centroid location information, and second location information of the two risk areas, the weight data between the nodes corresponding to the two risk areas is determined, including: According to the formula ， Determine the weight data between the node corresponding to the s-th risk region and the node corresponding to the t-th risk region. and ,in, For the s-th risk area, For the t-th risk area, for and The area of ​​their intersection. for and The area of ​​the union of the two sets. For the risk status information of the s-th building structure, For the risk status information of the t-th building structure, for and similarity, Let be the distance between the centroid locations of the s-th risk region and the t-th risk region. s represents the average distance between the centroid locations of each risk area, where s and t are positive integers.

3. The BIM-based construction safety supervision method according to claim 2, characterized in that, When training the graph neural network model, the loss function is determined based on the training risk feature values ​​and labeled risk levels of each training construction area. This includes determining the loss function as follows: when R ≤ L and LR < 1, the loss function is... When R < L and LR > 1, or R > L, the output information of the node corresponding to each risk region is input into the second fully connected layer and the second activation layer to obtain the risk feature value of each node, and the loss function is determined as follows: R represents the training risk feature value, and L represents the labeled risk level. This represents the maximum value among the risk characteristic values ​​of each node.

4. The BIM-based construction safety supervision method according to claim 1, characterized in that, The training steps of the structured image processing model include: Training images with the same structure are grouped into training image groups, and the angular encoding of each training image in the training image group is obtained; The first training feature vector of each training image is obtained through the feature extraction layers of the structured image processing model. The angle code is concatenated with the first training feature vector to obtain the second training feature vector for each training image; By using the attention mechanism of the structural image processing model, the connection weights between nodes corresponding to each training image are obtained; Based on the connection weights, the second training feature vectors corresponding to each training image are aggregated to obtain the third training feature vector of each training image. The third training feature vector is input into multiple first fully connected layers to obtain multiple training risk description values; Multiple training risk description values ​​are input into the first activation layer to obtain training probability information for various security risks. Based on the angle encoding, the risk type labeling information of the training images, and the training probability information, the first loss function of the structured image processing model is determined; The structure image processing model is trained based on the first loss function to obtain the trained structure image processing model; Based on the angle encoding, risk type labeling information of the training images, and training probability information, the first loss function of the structured image processing model is determined, including: According to the formula ， Determine the first loss function for the structural image processing model. ,in, This provides the training probability information for the j-th security risk in the k-th training image. This represents the maximum probability of the existence of the j-th security risk, determined based on risk type annotation information from multiple training images. Encode the angle of the k-th training image. An angular encoding for the training image that maximizes the probability of the existence of the j-th security risk, determined based on risk type annotation information from multiple training images. for and The maximum similarity, where N is the number of types of security risks, M is the number of training images, j≤N, k≤M, and j, N, k, and M are all positive integers.

5. A BIM-based construction safety supervision system, used to execute the method as described in any one of claims 1-4, characterized in that, include: The image to be processed module acquires images of the construction site to be processed at the end of the current construction cycle; The building model module associates the building structure in the image to be processed with the BIM building model to obtain the BIM building model in the current state. The first location information module determines the first location information of the building structure in the current state of the BIM building model; The risk status information module processes the image to be processed using a structural image processing model to obtain risk status information of the building structure in the image to be processed. The risk area module sets risk areas in the current state of the BIM building model based on the risk status information of the building structure and the first location information of the building structure in the current state of the BIM building model. The second location information module obtains the second location information of the risk area in the current state of the BIM building model; The risk level module determines the risk level of the area to be constructed in the current construction cycle based on the risk status information and the second location information of the risk area. The safety supervision plan module determines the safety supervision plan for the next construction cycle based on the risk level of the area under construction in the current construction cycle and the second location information.

Images