AI Safety Identification Method and System Based on BIM and Smart Construction Site

Abstract

The present application provides an AI safety recognition method and system based on BIM and intelligent construction sites. By effectively processing partially missing safety recognition data, especially for the safety index annotation data of BIM models with missing features, the safety index estimation data of the BIM model is introduced to generate the prior data of the BIM model safety index. This not only improves the robustness and accuracy of the AI safety optimization network when dealing with incomplete data, but also significantly optimizes the network parameter learning process, enabling the generated AI safety optimization network to efficiently and accurately output BIM model data with target safety recognition indicators, thereby enhancing the application efficiency of the BIM model in safety recognition and providing data support for the safety management of intelligent construction sites.

Description

Technical Field

[0001] This application relates to the technical field of intelligent construction sites. Specifically, it relates to an AI safety recognition method and system based on BIM and intelligent construction sites. Background Technique

[0002] With the wide application of Building Information Modeling (BIM) technology in the construction industry, BIM models have become an indispensable tool in the building design, construction, and management processes. BIM models not only contain the geometric information of buildings but also cover a large amount of safety index data, which is crucial for ensuring the safety of building construction. However, in practical applications, due to various reasons (such as incomplete data collection, annotation errors, etc.), there are often partial missing safety index data in BIM models, which poses great challenges to subsequent safety management and risk assessment.

[0003] Traditional solutions mainly rely on manual inspections and empirical judgments. This method is not only inefficient but also easily affected by human factors, resulting in low accuracy and reliability of safety recognition. In recent years, with the rapid development of artificial intelligence (AI) technology, applying AI technology to the safety recognition of BIM models has become a trend. However, existing AI safety recognition methods often cannot achieve satisfactory results when dealing with partially missing safety recognition data. Summary of the Invention

[0004] In view of the problems mentioned above, in combination with the first aspect of this application, embodiments of this application provide an AI safety recognition method based on BIM and intelligent construction sites. The method includes:

[0005] Obtain partial missing safety recognition data of a sample BIM model sequence learned by an AI safety optimization network. The partial missing safety recognition data includes X BIM model safety index annotation data. Each BIM model safety index annotation data includes annotation safety index parameters of a BIM model for Y safety indicators, and there is at least one BIM model safety index annotation data with feature missing among the X BIM model safety index annotation data. Among the annotation safety index parameters of the Y safety indicators included in the BIM model safety index annotation data with feature missing, there is at least one missing annotation safety index parameter of a safety indicator.

[0006] For each sample BIM model, if the BIM model safety index annotation data corresponding to the sample BIM model in the partially missing safety identification data has missing features, then based on the BIM model safety index estimation data and the BIM model safety index annotation data of the sample BIM model, generate the BIM model safety index prior data of the sample BIM model. The BIM model safety index estimation data of the sample BIM model includes: the estimated safety index parameters of the sample BIM model generated for at least one safety index;

[0007] Based on the BIM model safety index prior data of the sample BIM model, perform network parameter learning on the AI safety optimization network to generate an AI safety optimization network with completed network parameter learning. The AI safety optimization network with completed network parameter learning is used to optimize and generate BIM model data with target safety identification indexes.

[0008] On the other hand, an embodiment of the present application further provides a smart construction site service system, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or codes, and the processor is used to execute the programs, instructions, or codes in the machine-readable storage medium to implement the above method.

[0009] Based on the above aspects, the embodiment of the present application effectively processes partially missing safety identification data, especially for the BIM model safety index annotation data with missing features, introduces BIM model safety index estimation data to generate BIM model safety index prior data, not only improves the robustness and accuracy of the AI safety optimization network when processing incomplete data, but also significantly optimizes the network parameter learning process, so that the generated AI safety optimization network can efficiently and accurately output BIM model data with target safety identification indexes, thereby enhancing the application efficiency of the BIM model in safety identification and providing data support for the safety management of smart construction sites. Description of the Drawings

[0010] Figure 1 is a schematic execution flowchart of the AI safety identification method based on BIM and smart construction site provided by the embodiment of the present application.

[0011] Figure 2 is a schematic hardware architecture diagram of the smart construction site service system provided by the embodiment of the present application. Detailed Embodiments

[0012] The present application will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1It is a schematic flowchart of an AI security recognition method based on BIM and intelligent construction site provided by an embodiment of the present application. The AI security recognition method based on BIM and intelligent construction site will be introduced in detail below.

[0013] Step S110, obtain partial missing security recognition data of the sample BIM model sequence learned by the AI security optimization network. The partial missing security recognition data includes X BIM model security index annotation data. Each BIM model security index annotation data includes annotation security index parameters of a BIM model for Y security indexes. And there is at least one BIM model security index annotation data with feature missing in the X BIM model security index annotation data. Among the annotation security index parameters of the Y security indexes included in the BIM model security index annotation data with feature missing, there is at least one missing annotation security index parameter of a security index.

[0014] Specifically, the AI security optimization network is a network model based on artificial intelligence (AI) technology, which is specifically used to optimize the security recognition indexes of BIM models. It can automatically adjust model parameters by learning and analyzing a large amount of BIM model data to generate BIM models that meet specific security requirements.

[0015] The sample BIM model sequence is a series of BIM models used to train and test the AI security optimization network, which usually covers different design stages, design teams and building structure parts to ensure that the network can learn extensive security knowledge and features.

[0016] The partial missing security recognition data refers to the situation where some data in the collected BIM model security recognition data is incomplete or missing. These data may not contain all necessary security index parameters due to various reasons (such as incomplete data collection, data transmission errors, etc.).

[0017] The BIM model security index annotation data is the data for annotating each security index in the BIM model. Each BIM model will have a corresponding set of BIM model security index annotation data, which is used to describe the security features and performance of the model. The BIM model security index annotation data with feature missing refers to the situation where there is at least one missing annotation parameter of a security index in the BIM model security index annotation data. This kind of missing may be caused by problems in the data collection or transmission process.

[0018] For example, in this embodiment, during the process of building a BIM (Building Information Modeling) security management system for a large construction project, the data processing and model optimization of the whole process are managed.

[0019] For example, in this construction project, there are multiple building structure parts of different types, such as the main structure, auxiliary facilities, etc., and each part has a corresponding BIM model. These BIM models cover rich information, including safety indicators in multiple aspects such as structural safety, fire safety, and electrical safety.

[0020] In this embodiment, multiple sample BIM model sequences from different design stages and different design teams have been collected. For example, from the preliminary design stage, there is a sequence of 10 sample BIM models, and from the detailed design stage, there is another sequence of 8 sample BIM models, etc.

[0021] For each sample BIM model sequence, the server obtains its complete safety identification data. Taking the sequence of 10 sample BIM models in the preliminary design stage as an example, the complete safety identification data of each sample BIM model may include labeled safety indicator parameters for 5 safety indicators (assuming Y = 5, namely structural stability, fire protection level, good electrical grounding, rationality of evacuation routes, and ventilation safety).

[0022] For example, for the safety indicator of structural stability, the labeled safety indicator parameter may be a safety factor calculated according to specific structural mechanics; for the fire protection level, it may be a level value labeled according to the building fire protection code (such as level 1, level 2, etc.).

[0023] Among them, this embodiment can take each BIM model as a unit to merge the complete safety identification data of multiple sample BIM model sequences. During the merging process, data missing may occur.

[0024] Suppose the complete safety identification data from 3 different sample BIM model sequences are merged, and a total of 20 BIM models' safety identification data are obtained (i.e., X = 20). During this process, due to reasons such as incomplete data collection or data transmission errors for some models in specific safety indicators, there may be characteristic missing in the safety indicator labeled data of some BIM models.

[0025] For example, for one of the BIM models, due to a measurement device failure, no accurate labeled safety indicator parameter is obtained for the safety indicator of good electrical grounding, which forms the safety indicator labeled data of the BIM model with characteristic missing. Among the partially missing safety identification data of these 20 BIM models, there may be 5 such safety indicator labeled data of BIM models with characteristic missing.

[0026] Step S120. For each sample BIM model, if the BIM model safety index annotation data corresponding to the sample BIM model in the partially missing safety identification data has missing features, then based on the BIM model safety index estimation data and the BIM model safety index annotation data of the sample BIM model, generate the BIM model safety index prior data of the sample BIM model. The BIM model safety index estimation data of the sample BIM model includes: the estimated safety index parameters of the sample BIM model generated for at least one safety index.

[0027] Specifically, the BIM model safety index estimation data is the estimation data generated by a dedicated BIM model safety index estimation network. When there are missing features in the BIM model safety index annotation data, these data can be used to supplement or replace the missing annotation parameters. The BIM model safety index prior data is the prior information generated based on the BIM model safety index annotation data and the estimation data, and is used to provide additional guidance and constraints during the learning process of the AI safety optimization network to help the network more accurately generate BIM models that meet safety requirements.

[0028] In this embodiment, a dedicated BIM model safety index estimation network is pre-configured to generate the BIM model safety index estimation data of the sample BIM model. For a BIM model with missing features, this network will make an estimation based on other relevant information of the model.

[0029] For example, for a BIM model with missing features in the ventilation safety index, the BIM model safety index estimation network will estimate the estimated safety index parameters of ventilation safety based on relevant information such as the building space layout and personnel density of the model. Suppose this estimated safety index parameter is an evaluation value of the ventilation effect obtained by learning a large number of similar building models through a machine learning-based algorithm.

[0030] Taking a sample BIM model as an example, assume that the annotation safety index parameter of this sample BIM model for the safety index of the rationality of evacuation routes is missing.

[0031] First, the server will check whether the annotation safety index parameter of the rationality of evacuation routes in the BIM model safety index annotation data of this sample BIM model is missing. Since it is indeed missing, the server will obtain the estimated safety index parameter of the rationality of evacuation routes from the BIM model safety index estimation data of this sample BIM model as the prior safety index parameter of this safety index.

[0032] For other safety indicators in this sample BIM model where there are no missing labeled safety indicator parameters, such as structural stability, the server will directly output the labeled safety indicator parameters of structural stability as the prior safety indicator parameters of this safety indicator.

[0033] Finally, based on the prior safety indicator parameters of these Y safety indicators (assuming Y = 5), the server generates the prior BIM model safety indicator data of this sample BIM model.

[0034] Step S130: According to the prior BIM model safety indicator data of the sample BIM model, perform network parameter learning on the AI security optimization network to generate an AI security optimization network with completed network parameter learning. The AI security optimization network with completed network parameter learning is used to optimize and generate BIM model data with target security recognition indicators.

[0035] In this embodiment, the server generates an iterative BIM model corresponding to the sample BIM model based on the AI security optimization network, the sample BIM model, and the prior BIM model safety indicator data of the sample BIM model.

[0036] Among them, the AI security optimization network includes a first optimization branch and a second optimization branch. The first optimization branch is used to generate a hidden layer expression vector based on the sample BIM model and the prior BIM model safety indicator data of the sample BIM model. For example, for a sample BIM model, the first optimization branch generates a hidden layer expression vector through calculations and conversions of hidden layer neurons according to information such as the structural characteristics and prior safety indicator data of the model. The average data and morphological vector of the distribution model parameters followed by this hidden layer expression vector are determined by the network parameter information of the first optimization branch, the sample BIM model, and the prior BIM model safety indicator data of the sample BIM model.

[0037] Then, the second optimization branch generates an iterative BIM model based on this hidden layer expression vector and the prior BIM model safety indicator data of the sample BIM model. The average data and morphological vector of the distribution model parameters followed by this iterative BIM model are determined by the network parameter information of the second optimization branch, the hidden layer expression vector, and the prior BIM model safety indicator data of the sample BIM model.

[0038] Suppose for a sample BIM model, the annotation safety index parameter of the structural stability in the BIM model safety index annotation data for this annotation safety index is complete, and there is also an estimated safety index parameter for structural stability in the BIM model safety index estimation data. Based on these two parameters, the server calculates the fitness evaluation cost parameter. For example, if the difference between the estimated safety index parameter and the annotation safety index parameter is large, the fitness evaluation cost parameter will be high, which reflects the estimated performance index of the BIM model safety index estimation data.

[0039] Then, based on the iterative BIM model and the sample BIM model, the server determines the first cost parameter. For example, if there are significant differences in the building shape dimensions between the iterative BIM model and the sample BIM model (which is an obvious feature), then the first cost parameter will reflect the distance of this obvious feature between the iterative BIM model and the sample BIM model, and the value will be large.

[0040] Suppose the prior data of the BIM model safety index for the sample BIM model has been generated. The server determines the second cost parameter and the third cost parameter based on the prior data of the BIM model safety index, the collaborative parameter array of the prior data of the BIM model safety index, the average data of the prior data of the BIM model safety index, and the average data of the BIM model safety index estimation data.

[0041] For example, for an annotation safety index (such as the fire protection rating), if the distance between its prior safety index parameter and the overall characteristics of the prior data of the BIM model safety index is large, the second cost parameter will reflect this situation. And for other safety indices among the Y safety indices except the annotation safety index (such as the good electrical grounding), if the distance between its prior safety index parameter and the characteristics of the prior data of the BIM model safety index is large, the third cost parameter will reflect this difference.

[0042] Next, the server obtains the hidden layer expression vector generated for the sample BIM model from the hidden network layer of the AI safety optimization network. This hidden layer expression vector is the result of the sample BIM model being calculated and transformed by the hidden layer neurons. For example, this hidden layer expression vector may contain the representation of the sample BIM model in different abstract feature spaces.

[0043] Then, the mean and shape vector in the distribution model parameters followed by the hidden layer expression vector are determined. The mean reflects the central position of the hidden layer expression vector in the current state distribution, and the shape vector is related to the distribution shape.

[0044] Select a sample set of hidden layer expression vectors based on the mean and shape vectors in the distribution model parameters followed by the hidden layer expression vectors, and obtain a pre-constructed hypothesis distribution model (this hypothesis distribution model represents the distribution pattern that the hidden layer expression vectors should follow under the hypothesis state distribution).

[0045] Use the KL divergence as a measure of the difference between the distribution model parameters of the hidden layer expression vectors in the hidden layer expression vector sample set and the distribution model parameters of the hypothesis distribution model. Determine the relevant parameters for calculating the fourth cost parameter based on the mean and covariance matrix of the distribution model parameters of the hidden layer expression vectors implemented, and the covariance matrix related to the mean and shape vectors corresponding to the hypothesis distribution model.

[0046] Calculate the fourth cost parameter based on the calculation method of the KL divergence and the relevant parameters, integrating the difference between the mean and covariance matrix of the distribution model parameters of the hidden layer expression vectors and the covariance matrix related to the mean and shape vectors corresponding to the hypothesis distribution model. For example, if the distribution of the hidden layer expression vectors is significantly different from the hypothesis distribution model, the fourth cost parameter will be larger. Finally, divide the calculated fourth cost parameter by the reference value calculated based on the prior example BIM model for normalization to generate the normalized fourth cost parameter.

[0047] Thus, the server determines the probability model optimization cost parameter based on the first cost parameter, the second cost parameter, the third cost parameter, and the fourth cost parameter. This probability model optimization cost parameter reflects the feature distance between the iterative BIM model generated based on the prior data of the BIM model safety index and the example BIM model.

[0048] Finally, the server determines the training cost parameter based on the fitness evaluation cost parameter and the probability model optimization cost parameter.

[0049] In this embodiment, the server optimizes the network parameters of the AI security optimization network based on the training cost parameter. For example, if the training cost parameter is high, it indicates that the performance of the current AI security optimization network needs to be improved. The server will adjust parameters such as weights and biases in the network to reduce the training cost parameter. After multiple iterative optimizations, an AI security optimization network that has completed network parameter learning is finally generated. This AI security optimization network that has completed network parameter learning can be used to optimize and generate BIM model data with target security identification indicators. For example, in subsequent construction projects, if it is necessary to generate a BIM model that meets specific target security identification indicators such as structural safety and fire safety, this optimized AI security optimization network can be used for optimization and generation.

[0050] If the number of BIM models in the partially missing safety identification data is greater than the threshold number (assuming the threshold number is 15). For example, in the previously mentioned partially missing safety identification data of 20 BIM models, the BIM model safety index estimation network will call the first deep learning network for which network parameter pre-learning has not been performed. Moreover, during the network parameter learning process of the AI safety optimization network, the network parameters of the first deep learning network are optimized based on the loss function value between the BIM model safety index annotation data of the sample BIM model and the BIM model safety index estimation data of the sample BIM model.

[0051] If the number of BIM models in the partially missing safety identification data is not greater than the threshold number (assuming there are only 10 BIM models in the partially missing safety identification data), the BIM model safety index estimation network will call the second deep learning network after network parameter pre-learning. This second deep learning network is an artificial intelligence network with BIM model safety index estimation performance generated after network parameter pre-learning using the complete safety identification data of the sample BIM model sequence. The complete safety identification data includes various sample BIM models and corresponding safety index parameters.

[0052] Based on the above steps, the embodiment of the present application effectively processes the partially missing safety identification data, especially for the BIM model safety index annotation data with feature missing, and introduces the BIM model safety index estimation data to generate the BIM model safety index prior data, which not only improves the robustness and accuracy of the AI safety optimization network when processing incomplete data, but also significantly optimizes the network parameter learning process, enabling the generated AI safety optimization network to efficiently and accurately output the BIM model data with the target safety identification index, thereby enhancing the application efficiency of the BIM model in safety identification and providing data support for the safety management of smart construction sites.

[0053] In a possible implementation manner, step S120 includes:

[0054] Step S121, for each of the Y safety indicators, if the annotation safety index parameter of the safety indicator covered in the BIM model safety index annotation data of the sample BIM model is missing, obtain the estimated safety index parameter of the safety indicator from the BIM model safety index estimation data of the sample BIM model as the safety index prior parameter of the safety indicator.

[0055] Step S122, if the annotation safety index parameter of the safety indicator covered in the BIM model safety index annotation data of the sample BIM model is not missing, output the annotation safety index parameter of the safety indicator as the safety index prior parameter of the safety indicator.

[0056] Step S123: Generate the prior data of the BIM model safety indicators for the sample BIM model based on the prior parameters of the safety indicators of the Y safety indicators.

[0057] In this embodiment, in the large-scale construction project mentioned above, the server continuously processes and optimizes the safety indicators of the BIM model. Still taking a project containing multiple different building structure parts (such as the main structure, auxiliary facilities, etc.) as an example, each part has a corresponding BIM model, and these BIM models contain safety indicator information in multiple aspects such as structural safety, fire safety, and electrical safety.

[0058] Suppose we focus on a specific sample BIM model, which represents an important functional area in the building. For example, it is a BIM model of an office area containing numerous personnel activity spaces and complex equipment. Assume Y = 5, that is, there are 5 safety indicators, namely structural stability, fire protection level, good electrical grounding, rationality of evacuation routes, and ventilation safety.

[0059] For the structural stability indicator, the server first checks the marked safety indicator parameters of structural stability in the marked data of the BIM model safety indicators of this sample BIM model.

[0060] In this construction project, the marked safety indicator parameters of structural stability are obtained through detailed structural mechanics analysis. For example, they are the safety factors of the structure under different load combinations. Suppose the server finds that the marked safety indicator parameters of this sample BIM model are completely present without any missing situation.

[0061] Then, according to the rules, the server will directly output the marked safety indicator parameters of this structural stability as the prior parameters of the safety indicator of structural stability. This means that for the key safety indicator of structural stability, the existing accurate marked data is used as prior information because it is a reliable result obtained through professional structural analysis.

[0062] For the safety indicator of fire protection level, the server checks the marked data of the BIM model safety indicators of the sample BIM model again.

[0063] The marked safety indicator parameters of the fire protection level are marked according to the building fire protection code. For example, they are numerical values such as level 1 and level 2, which reflect the fire protection capabilities in many aspects such as building materials and fire partitions.

[0064] Suppose in this sample BIM model, due to certain reasons (such as partial loss during data transmission or negligence during original data collection), the marked safety indicator parameters of the fire protection level are missing.

[0065] At this time, the server obtains the estimated safety index parameter of the fire protection level from the BIM model safety index estimation data of the sample BIM model. This estimated safety index parameter is generated by the BIM model safety index estimation network in the server. This network may estimate a fire protection level value, such as estimating it as level 2, based on various factors such as the functional use of the building (office area, relatively high personnel density, high fire protection requirements), and the type of building materials (materials that may be used inferred from other labeled parts) through a pre-learned algorithm. Then, this estimated value of level 2 is used as the prior parameter of the safety index for the fire protection level.

[0066] For the electrical grounding goodness index, the server checks the situation of the electrical grounding goodness safety index in the BIM model safety index annotation data of the sample BIM model.

[0067] The annotated safety index parameter of the electrical grounding goodness is reflected by data such as the grounding resistance value obtained by electrical detection equipment.

[0068] Assume that in this sample BIM model, the annotated safety index parameter of the electrical grounding goodness exists. For example, the grounding resistance value is accurately annotated as 1 ohm, which is an accurate value obtained during the electrical system design and testing process.

[0069] Then, the server directly uses this annotated safety index parameter of 1 ohm as the prior parameter of the safety index for the electrical grounding goodness. This is because accurate annotation data can directly provide reliable prior information without using estimated values.

[0070] For the evacuation route rationality index, when the server checks the evacuation route rationality safety index in the BIM model safety index annotation data of the sample BIM model.

[0071] The annotated safety index parameter of the evacuation route rationality may involve the comprehensive evaluation results of various factors such as the width of the route, the length of the route, the location and quantity of evacuation doors.

[0072] Assume that in this sample BIM model, due to some design change information not being updated to the annotation data in a timely manner, the annotated safety index parameter of the evacuation route rationality is missing.

[0073] Then, the server will obtain the estimated safety index parameters for the rationality of evacuation routes from the BIM model safety index estimation data. This estimated safety index parameter may be an evaluation value estimated through machine learning algorithms based on factors such as the total floor area of the building, the occupancy capacity, and the distribution of different functional areas. For example, if the estimated value for the rationality of the evacuation route is 0.8 (assuming that values between 0 and 1 represent the degree of rationality), this 0.8 is used as the prior safety index parameter for the safety index of the rationality of the evacuation route.

[0074] For the ventilation safety index, the server checks the status of the safety index of ventilation safety in the BIM model safety index annotation data of the sample BIM model.

[0075] The annotated safety index parameters for ventilation safety may be obtained based on factors such as the designed air volume and the number of air changes of the ventilation system.

[0076] Suppose that in this sample BIM model, the annotated safety index parameters for ventilation safety are complete. For example, the number of air changes is annotated as 5 times per hour, which is a value obtained based on the ventilation system design specifications and actual calculations.

[0077] Then, the server will directly use this annotated safety index parameter of 5 times per hour as the prior safety index parameter for the safety index of ventilation safety.

[0078] After the server has respectively determined the prior safety index parameters for these Y = 5 safety indexes (structural stability, fire protection level, good electrical grounding, rationality of evacuation routes, ventilation safety), it will generate the prior data of the BIM model safety index for this sample BIM model based on the prior safety index parameters of these 5 safety indexes.

[0079] This prior data of the BIM model safety index will be a set containing the prior safety index parameters of these 5 safety indexes, which combines accurate annotated safety index parameters (when available) and safety index parameters obtained through estimation (when annotation parameters are missing). This prior data will serve as an important basis for subsequent network parameter learning of the AI safety optimization network. For example, when training the AI safety optimization network, this prior data will be input into the network together with the sample BIM model, and the network will learn and optimize based on this prior information to improve the prediction and optimization ability of the BIM model safety index, and finally generate a network that can optimize the BIM model data with target safety recognition indexes.

[0080] Thus, it is possible to accurately generate the prior data of the BIM model safety index for the sample BIM model, providing a key data basis for the effective operation of the entire BIM model safety optimization system.

[0081] In a possible implementation, the method further includes: generating a missing indication array corresponding to the partially missing safety identification data, where the missing indication array includes X×Y nodes, and each node represents whether the labeled safety index parameter of a safety index of a BIM model is missing.

[0082] In the previously mentioned construction project, the server has obtained the partially missing safety identification data of the sample BIM model sequence learned by the AI safety optimization network. Here, it is assumed that there are X = 20 BIM model safety index annotation data in the partially missing safety identification data, and each annotation data targets Y = 5 safety indexes (structural stability, fire protection level, electrical grounding integrity, evacuation route rationality, ventilation safety).

[0083] The server starts to generate the missing indication array. This missing indication array will have 20×5 = 100 nodes. For the safety index of structural stability of the first BIM model, the server checks whether the labeled safety index parameter in the partially missing safety identification data is missing. If it is missing, then the corresponding node in the missing indication array (the first BIM model corresponds to the 1st - 5th nodes, and structural stability is the 1st node) will be marked with a specific identifier indicating missing (e.g., 1); if there is no missing, it will be marked with an identifier indicating no missing (e.g., 0).

[0084] Next, for the safety index of fire protection level of the first BIM model, repeat the above - mentioned checking process and make corresponding marks on the corresponding node (the 2nd node) in the missing indication array. In this way, check and mark all 5 safety indexes of the first BIM model in turn.

[0085] Then, the server operates on the 5 safety indexes of the second BIM model in the same way, and accurately marks the corresponding nodes of each safety index in the missing indication array. For example, if the labeled safety index parameter of electrical grounding integrity of the second BIM model is missing, mark it as 1 at the corresponding node (the 6 + 3 - 1 = 8th node).

[0086] This process continues until the checking and marking of all 5 safety indexes of all 20 BIM models are completed, and finally a complete missing indication array is formed. This array fully reflects whether the labeled safety index parameters of each safety index of each BIM model in the partially missing safety identification data are missing, provides an intuitive and useful reference for subsequent data processing, analysis, and operations based on the sample BIM model, and helps the server manage and process BIM model safety identification - related data more efficiently.

[0087] In a possible implementation, step S110 includes:

[0088] Step S111: Obtain the complete safety identification data of multiple sample BIM model sequences. The complete safety identification data of each sample BIM model sequence includes the labeled safety index parameters of at least one BIM model for at least one safety index. The safety indexes covered by the complete safety identification data of different sample BIM model sequences are all different, and the BIM models covered by the complete safety identification data of different sample BIM model sequences are all different.

[0089] Step S112: Take each BIM model as a unit, perform a merging process on the complete safety identification data of the multiple sample BIM model sequences to generate the partially missing safety identification data.

[0090] In this embodiment, the server first obtains the complete safety identification data of sample BIM model sequences from different design stages of a construction project. For example, in the preliminary design stage of a building, there is a sample BIM model sequence. Each BIM model in this sequence corresponds to different parts of the building, such as the foundation structure part, the first-floor main structure part, etc. Each BIM model has labeled safety index parameters for specific safety indexes. For the BIM model of the foundation structure part, its safety indexes may include the bearing capacity of the foundation, etc. The labeled safety index parameter of the safety index of the bearing capacity of the foundation is a specific value obtained through geological exploration data and structural design calculations, such as being able to withstand a pressure of 50 tons per square meter.

[0091] In the detailed design stage of the building, there is another sample BIM model sequence. The safety indexes targeted by the BIM models in this sequence are different. For example, for the BIM model of the electrical system part inside the building, the safety indexes include the overload protection ability of the electrical circuit, etc. The labeled safety index parameter of the overload protection ability of the electrical circuit is determined according to the electrical design specifications. For example, the overload protection current is set to 20 amperes.

[0092] For another example, in the decoration design stage of the building, there is another sample BIM model sequence. The safety indexes targeted by the BIM models in this sequence may be the fire resistance performance of the decoration materials, etc. The labeled safety index parameter of the fire resistance performance of the decoration materials is marked according to the fireproof material grade standard, such as reaching the B1 fireproof standard.

[0093] Among them, the BIM models covered by the complete safety identification data of these different sample BIM model sequences are all different. Each BIM model has its unique structural and functional characteristics, corresponding to different safety requirements and evaluation criteria.

[0094] Meanwhile, the safety indicators covered by the complete safety identification data of different sample BIM model sequences are also different. This is because the safety focuses vary in different stages and parts of the building. For example, in the preliminary design stage, the focus is on basic safety indicators such as structural safety, while in the decoration design stage, more attention is paid to safety indicators related to decoration materials.

[0095] Next, the server begins to merge the complete safety identification data of these multiple sample BIM model sequences in units of each BIM model. For example, the server first processes the BIM model data from the preliminary design stage and the detailed design stage. For a BIM model of the basic structure part that exists in both stages but represents different design states, the server merges the safety indicator data of this BIM model in these two stages.

[0096] During the merging process, some problems may occur. For example, in the preliminary design stage, only the safety indicator of the foundation bearing capacity may be marked for the BIM model of the basic structure part, while in the detailed design stage, in addition to the foundation bearing capacity, the seismic performance of the basic structure is added as a safety indicator. However, for some reason, during the merging, the data of the newly added safety indicator of seismic performance may not be fully and correctly merged, resulting in partial data loss.

[0097] As the server continues to merge the complete safety identification data of more sample BIM model sequences, the situation of data loss may become more complex. For example, when merging the BIM model data of the decoration design stage, for some safety indicators related to structure and electricity, due to differences in data formats in different stages or negligence in data management, some data may be lost during the merging.

[0098] After the server merges all the relevant complete safety identification data of the sample BIM model sequences, it finally generates partially missing safety identification data. This partially missing safety identification data contains the safety indicator annotation data of multiple BIM models, but there is at least one BIM model safety indicator annotation data with feature missing, that is, the marked safety indicator parameters of some BIM models for some safety indicators are missing.

[0099] Through the above steps, partially missing safety identification data for subsequent operations is obtained, which provides a data basis for further optimizing the safety identification and management of BIM models.

[0100] In a possible implementation manner, step S130 includes:

[0101] Step S131: Based on the AI security optimization network, the iterative BIM model corresponding to the sample BIM model is generated according to the sample BIM model and the prior data of the BIM model security indicators of the sample BIM model.

[0102] Step S132: Based on the sample BIM model and the iterative BIM model, the training cost parameter of the AI security optimization network is determined, and the training cost parameter reflects the feature distance between the sample BIM model and the iterative BIM model.

[0103] Step S133: Based on the training cost parameter, the network parameters of the AI security optimization network are optimized to generate the AI security optimization network that has completed network parameter learning.

[0104] In this embodiment, in the BIM model security management of a construction project, the AI security optimization network in the server has a specific structure and function. This network includes a first optimization branch and a second optimization branch.

[0105] For a specific sample BIM model, such as the BIM model representing the office area mentioned before, the server first utilizes the first optimization branch of the AI security optimization network. This sample BIM model contains safety information in multiple aspects such as structure, fire protection, and electricity, and there is already prior data of BIM model security indicators generated based on the previous steps.

[0106] The first optimization branch starts to generate a hidden layer expression vector based on the sample BIM model and its prior data of BIM model security indicators. This hidden layer expression vector is the result of the calculation and transformation of the sample BIM model and its prior data by the hidden layer neurons. For example, for the safety indicator of structural stability, if the prior parameter of the safety indicator in the prior data indicates that the structural stability is at a certain specific level, the hidden layer neurons will, based on this information and the structure-related features in the sample BIM model (such as beam-column layout, material strength, etc.), convert this information into a part of the hidden layer expression vector through complex calculations.

[0107] Then, the second optimization branch generates an iterative BIM model based on this hidden layer expression vector and the prior data of the BIM model security indicators of the sample BIM model. This iterative BIM model is, to a certain extent, the result of the optimization or adjustment of the sample BIM model. For example, in terms of fire safety, if the prior data of the sample BIM model shows that there is room for improvement in the fire protection level, the iterative BIM model may adjust the fire separation layout inside the building or use materials with higher fire protection performance, and these adjustments are based on the network's analysis of the prior data and the hidden layer expression vector.

[0108] Next, the server begins to determine the fitness evaluation cost parameter among the training cost parameters. Taking the structural stability safety index in the sample BIM model as an example, in the BIM model safety index annotation data of the sample BIM model, there are accurate annotation safety index parameters for structural stability (such as the structural safety factor), and at the same time, there are also estimated safety index parameters for structural stability in the BIM model safety index estimation data.

[0109] The server determines the fitness evaluation cost parameter by comparing these two parameters. If the estimated safety index parameter differs significantly from the annotation safety index parameter, it indicates that the estimation accuracy is low, and the fitness evaluation cost parameter will be high. For example, if the annotated structural safety factor is 1.5 and the estimated safety factor is 1.2, the difference between the two will be quantified as part of the fitness evaluation cost parameter. This parameter reflects the estimation performance index of the BIM model safety index estimation data.

[0110] Next, the first cost parameter is determined. This first cost parameter reflects the distance of the iterative BIM model from the explicit features of the sample BIM model. For example, in terms of the explicit feature of the building's external dimensions, if the length, width, and height of the sample BIM model are 50 meters, 30 meters, and 20 meters respectively, and the length, width, and height of the iterative BIM model become 52 meters, 30 meters, and 20 meters, the server will calculate the first cost parameter based on the differences in these dimensions. The greater the change in the dimensions, the larger the value of the first cost parameter, which intuitively reflects the degree of difference in explicit features such as the appearance of the two models.

[0111] Regarding the second cost parameter, taking the fire protection grade, which is an annotation safety index in the BIM model safety index prior data of the sample BIM model, as an example. The server determines it based on the BIM model safety index prior data, the collaborative parameter array of the BIM model safety index prior data, the average data of the BIM model safety index prior data, and the average data of the BIM model safety index estimation data. If the prior parameter of the fire protection grade safety index has a large distance from the overall characteristics of the prior data, for example, if the overall prior data tends to have a higher fire protection grade while the prior parameter of this annotation safety index is relatively low, the second cost parameter will reflect this situation.

[0112] Regarding the third cost parameter, consider other safety indexes in the sample BIM model besides the annotation safety indexes (such as the fire protection grade), such as the good electrical grounding. If the prior parameter of the electrical grounding safety index has a large distance from the characteristics of the BIM model safety index prior data, the third cost parameter will reflect this difference. For example, if the prior data indicates that the overall electrical safety is at a high level, but the prior parameter of the electrical grounding shows a large grounding resistance (i.e., poor grounding condition), the third cost parameter will increase accordingly.

[0113] Next, the server obtains the hidden layer expression vector generated for the sample BIM model from the hidden network layer of the AI security optimization network. This hidden layer expression vector has distribution model parameters it follows, including the mean and the shape vector. The mean reflects the central position of the hidden layer expression vector in the current state distribution, and the shape vector is related to the distribution shape.

[0114] The server selects a sample set of the hidden layer expression vector based on the mean and the shape vector in the distribution model parameters followed by the hidden layer expression vector, and obtains the pre-constructed hypothesis distribution model. The hypothesis distribution model represents the distribution pattern that the hidden layer expression vector should follow under the hypothesis state.

[0115] Then, the KL divergence is used as a measure of the difference between the distribution model parameters of the hidden layer expression vectors in the hidden layer expression vector sample set and the distribution model parameters of the hypothesis distribution model. For example, if there is a large difference between the actual distribution of the hidden layer expression vector and the hypothesis distribution model, the KL divergence will be large. The server determines the relevant parameters for calculating the fourth cost parameter based on the mean and covariance matrix of the distribution model parameters of the hidden layer expression vector, and the covariance matrix related to the mean and shape vector corresponding to the hypothesis distribution model. Finally, based on the calculation method of the KL divergence and the relevant parameters, the difference between the mean and covariance matrix of the distribution model parameters of the hidden layer expression vector and the covariance matrix related to the mean and shape vector corresponding to the hypothesis distribution model is comprehensively considered to calculate the fourth cost parameter. The calculated fourth cost parameter is normalized by dividing it by the reference value calculated based on the prior sample BIM model to generate the normalized fourth cost parameter.

[0116] The server determines the probability model optimization cost parameter based on the first cost parameter, the second cost parameter, the third cost parameter, and the fourth cost parameter. This probability model optimization cost parameter reflects the feature distance between the iterative BIM model generated based on the prior data of the BIM model security index and the sample BIM model.

[0117] Finally, the server determines the training cost parameter based on the fitness evaluation cost parameter and the probability model optimization cost parameter. This training cost parameter comprehensively considers various factors such as estimation accuracy and model feature distance.

[0118] For example, the server optimizes the network parameters of the AI security optimization network according to the calculated training cost parameter. If the training cost parameter is high, it indicates that the performance of the current AI security optimization network still needs to be improved. For example, if the fitness evaluation cost parameter is high, it indicates poor estimation accuracy, and the server may adjust network parameters such as weights and biases in the first optimization branch and the second optimization branch to reduce the fitness evaluation cost parameter.

[0119] For the case where the first cost parameter is relatively high, it may mean that the difference in explicit features between the iterative BIM model and the sample BIM model is unreasonable, and the server will adjust the network parameters to reduce this difference. Similarly, for the cases where the second cost parameter, the third cost parameter, and the fourth cost parameter are relatively high, the network parameters will also be adjusted accordingly, such as adjusting the connection weights between neurons in the hidden layer, etc.

[0120] Through multiple iterations of optimization, continuously adjust the network parameters until the training cost parameter reaches a satisfactory level, and finally generate an AI security optimization network that has completed learning of the network parameters. This AI security optimization network that has completed learning of the network parameters can better optimize according to the sample BIM model and its prior data, and can be used to optimize and generate BIM model data with target security identification indicators.

[0121] In a possible implementation manner, the training cost parameter includes a goodness-of-fit evaluation cost parameter and a probability model optimization cost parameter, and the probability model optimization cost parameter includes a first cost parameter, a second cost parameter, a third cost parameter, and a fourth cost parameter.

[0122] Step S132 includes:

[0123] Step S1321, based on the annotation security index parameters of the annotation security index in the BIM model security index annotation data, and the estimated security index parameters of the annotation security index in the BIM model security index estimation data, determine the goodness-of-fit evaluation cost parameter, where the goodness-of-fit evaluation cost parameter reflects the estimated performance index of the BIM model security index estimation data, and the annotation security index is the security index corresponding to the annotation security index parameter without missing values in the BIM model security index annotation data.

[0124] Wherein, when the BIM model security index estimation data of the sample BIM model continues to be updated and changed as the network parameter information of the AI security optimization network is learned, the collaborative parameter array and the average parameter value of the BIM model security index prior data follow the continuous update and change.

[0125] In this embodiment, in the BIM model security management scenario of a construction project, the server processes the relevant data of the sample BIM model. Taking the BIM model of the office area mentioned above as an example, this model has multiple security indicators, such as structural stability, fire protection grade, good electrical grounding, rationality of evacuation routes, and ventilation safety, etc.

[0126] For the safety index of structural stability, in the BIM model safety index annotation data, the annotation safety index parameters are obtained based on detailed structural analysis. Assuming that the structural system of this office area is a frame structure, through the calculation of the bearing capacity of beams and columns and the seismic analysis of the overall structure, etc., the annotation safety index parameter of structural stability is a safety factor of 1.5. This value is determined after considering various factors such as the seismic intensity of the area where the building is located and the design loads of the building (including live loads and dead loads).

[0127] At the same time, the estimated safety index parameters in the BIM model safety index estimation data are generated by a dedicated estimation network. This estimation network may consider factors such as the height of the building, the number of floors, and the type of structural materials, and estimates the safety factor of structural stability to be 1.3.

[0128] The server uses a specific algorithm to calculate the goodness-of-fit evaluation cost parameter. A common method is the mean squared error (MSE). For the annotation safety index of structural stability, the calculation method is as follows:

[0129] Mean squared error (MSE=(annotation safety index parameter - estimated safety index parameter)^2=(1.5 - 1.3)^2 = 0.04)

[0130] For the safety index of fire resistance rating, the annotation safety index parameter is determined to be level 2 according to the building fire protection code. The BIM model safety index estimation data estimates the fire resistance rating to be 1.8 levels based on factors such as the personnel density and fire load in the building. Calculate the mean squared error:

[0131] (MSE=(2 - 1.8)^2 = 0.04)

[0132] Assume that for the safety index of good electrical grounding, the annotation safety index parameter is a grounding resistance of 0.8 ohms, and the estimated safety index parameter is 0.9 ohms. Calculate the mean squared error:

[0133] (MSE=(0.8 - 0.9)^2 = 0.01)

[0134] The server synthesizes these mean squared errors calculated for different annotation safety indices. If simply taking the average, the goodness-of-fit evaluation cost parameter is:

[0135] ((0.04 + 0.04 + 0.01) / 3 = 0.03). This value reflects the estimation performance index of the BIM model safety index estimation data on multiple indices. If the goodness-of-fit evaluation cost parameter is large, it indicates that the accuracy of the estimation network needs to be improved.

[0136] Step S1322: Based on the iterative BIM model and the sample BIM model, determine the first cost parameter, which reflects the distance of the explicit feature between the iterative BIM model and the sample BIM model.

[0137] For the explicit feature of the building's external dimensions, in the sample BIM model, the length of the office area is 50 meters, the width is 30 meters, and the height is 20 meters. In the iterative BIM model, the length becomes 51 meters, the width becomes 30.5 meters, and the height becomes 20 meters.

[0138] Calculate the distances in each dimension:

[0139] The distance in the length direction is (vert51 - 50vert = 1) meter;

[0140] The distance in the width direction is (vert30.5 - 30vert = 0.5) meter;

[0141] The distance in the height direction is (vert20 - 20vert = 0) meter.

[0142] One way for the server to determine the first cost parameter is to simply sum these distances. The first cost parameter is (1 + 0.5 + 0 = 1.5) meters. This value reflects the distance between the iterative BIM model and the sample BIM model in terms of the explicit feature of the building's external dimensions.

[0143] Next, consider the explicit feature of the evacuation route layout. In the sample BIM model, there are two evacuation routes. The width of Route 1 is 2 meters, and the width of Route 2 is 1.8 meters. The positions of the evacuation doors are at both ends and in the middle of the routes respectively. In the iterative BIM model, the width of Route 1 becomes 2.2 meters, the width of Route 2 becomes 1.9 meters, and the position of one of the evacuation doors is slightly adjusted.

[0144] For the change in the width of the routes:

[0145] The distance of the change in the width of Route 1 is (vert2.2 - 2vert = 0.2) meter;

[0146] The distance of the change in the width of Route 2 is (vert1.9 - 1.8vert = 0.1) meter.

[0147] For the change in the position of the evacuation door, according to the pre - defined measurement method, assume the corresponding distance for this change is 0.5 (this value is the quantification result of the impact of the change in the evacuation door position on the evacuation efficiency).

[0148] Taken together, the contribution of this part to the first cost parameter is (0.2 + 0.1 + 0.5 = 0.8).

[0149] The server combines the distance contributions of explicit features such as the building's external dimensions and evacuation route layout to determine the first cost parameter. If there are only these two parts, the first cost parameter is (1.5 + 0.8 = 2.3). This first cost parameter comprehensively reflects the distance between the iterative BIM model and the sample BIM model in multiple explicit features.

[0150] Step S1323: Based on the BIM model safety index prior data, the collaborative parameter array of the BIM model safety index prior data, the average data of the BIM model safety index prior data, and the average data of the BIM model safety index estimated data, determine the second cost parameter and the third cost parameter. The second cost parameter reflects the feature distance of the safety index prior parameter of the marked safety index in the BIM model safety index prior data compared to the BIM model safety index prior data. The third cost parameter reflects the feature distance of the safety index prior parameters of the other safety indices except the marked safety index among the Y safety indices covered by the BIM model safety index prior data compared to the BIM model safety index prior data.

[0151] Assume that the fire protection level is used as the marked safety index. In the BIM model safety index prior data of the sample BIM model, the safety index prior parameter of the fire protection level is level 2.

[0152] The collaborative parameter array of the BIM model safety index prior data contains the relationship information between each safety index. For example, the degree of association between the fire protection level and safety indices such as building materials and evacuation routes. Assume that the collaborative parameter array indicates that the correlation coefficient between the fire protection level and building materials is 0.8 (indicating a high correlation), and the correlation coefficient with the evacuation route is 0.6.

[0153] The average data of the BIM model safety index prior data reflects the average level of all safety index prior parameters. Assume that after converting all safety index prior parameters into numerical values (for example, levels 1 - 3 correspond to 1, 2, 3 respectively), the average data is 2.2.

[0154] The average data of the BIM model safety index estimated data is assumed to be 2.1.

[0155] The server determines the second cost parameter based on this data. A possible calculation method is to consider the difference between the prior parameter of the safety index of the fire protection level and the overall characteristics of the prior data, and at the same time consider the correlation relationship in the collaborative parameter array. For example, if the fire protection level is level 2, compared with the average data of 2.2, the difference is (vert2 - 2.2vert = 0.2). Then, this difference is weighted according to the correlation coefficient in the collaborative parameter array. Assuming a high correlation with building materials, a weight of 0.8 is given, and a second - highest correlation with evacuation routes, a weight of 0.6 is given. The calculated second cost parameter is ((0.2 * 0.8 + 0.2 * 0.6) / 2 = 0.14). This second cost parameter reflects the characteristic distance between the prior parameter of the safety index (fire protection level) of the labeled safety index and the prior data of the safety index of the BIM model.

[0156] For other safety indicators other than the labeled safety index (fire protection level), such as good electrical grounding. In the prior data of the safety index of the BIM model, the prior parameter of the safety index of good electrical grounding is a grounding resistance of 0.8 ohms.

[0157] Similarly, consider the collaborative parameter array of the prior data of the safety index of the BIM model, the average data, and the average data of the estimated data of the safety index of the BIM model. Assume that the correlation coefficients of good electrical grounding with other safety indicators in the collaborative parameter array are different. The correlation coefficient with structural stability is 0.5, and the correlation coefficient with ventilation safety is 0.4.

[0158] According to a calculation method similar to that of the second cost parameter, calculate the difference between the prior parameter of the safety index of good electrical grounding and the overall characteristics of the prior data. Assume that the ideal value of the grounding resistance corresponding to the overall characteristics of the prior data (considering the comprehensive situation of other related safety indicators) is 0.7 ohms, and the difference is (vert0.8 - 0.7vert = 0.1).

[0159] Weight this difference according to the correlation coefficient, and the calculated third cost parameter is ((0.1 * 0.5 + 0.1 * 0.4) / 2 = 0.045). This third cost parameter reflects the characteristic distance between the prior parameter of the safety index of other safety indicators (good electrical grounding) other than the labeled safety index and the prior data of the safety index of the BIM model.

[0160] Step S1324, based on the hidden - layer expression vector, and the morphological vector and mean of the distribution - model parameters followed by the hidden - layer expression vector, determine the fourth cost parameter, where the fourth cost parameter reflects the discrete parameter value of the hidden - layer expression vector compared to the distribution - model parameters, and the hidden - layer expression vector is the hidden - relationship vector of the sample BIM model generated by the hidden network layer of the AI safety optimization network.

[0161] In this embodiment, the server obtains the hidden layer expression vector generated for the sample BIM model from the hidden network layer of the AI security optimization network. This hidden layer expression vector is the result of complex calculations and transformations of the sample BIM model by the hidden layer neurons, and it contains the representation of the sample BIM model in the abstract feature space.

[0162] Assume that the hidden layer expression vector is a multi-dimensional vector (\(\vec{v}=(v_1,v_2,\cdots,v_n)\)), and the distribution model parameters it follows include the shape vector and the mean. For example, the mean \(\mu = (\mu_1,\mu_2,\cdots,\mu_n)\), and the shape vector \(\Sigma\) (which can be parameters such as the covariance matrix representing the distribution shape).

[0163] The server first determines the mean and the shape vector in the distribution model parameters followed by the hidden layer expression vector. The mean reflects the central position of the hidden layer expression vector in the current state distribution. For example, \(\mu_1\) may represent the average value in a certain abstract feature dimension. The shape vector is related to the distribution shape, and the element \(\Sigma_{ij}\) in the covariance matrix represents information such as the correlation between different dimensions.

[0164] Select a sample set of the hidden layer expression vector based on the mean and the shape vector in the distribution model parameters followed by the hidden layer expression vector. Assume that a set of samples \(\{\vec{v}_1,\vec{v}_2,\cdots,\vec{v}_m\}\) is selected from the multiple calculation results of the hidden layer expression vector.

[0165] Obtain the pre-constructed hypothesis distribution model. This hypothesis distribution model represents the distribution pattern that the hidden layer expression vector should follow under the hypothesis state. For example, it is a standard normal distribution model (with a mean of 0 and a covariance matrix of the identity matrix) or an ideal distribution model determined based on a large amount of prior data.

[0166] Use the KL divergence (Kullback-Leibler Divergence) as a measure of the difference between the distribution model parameters of the hidden layer expression vectors in the hidden layer expression vector sample set and the distribution model parameters of the hypothesis distribution model.

[0167] According to the calculation formula of the KL divergence:

[0168] \((D_{KL}(P||Q)=\sum_{i}P(i)\log\frac{P(i)}{Q(i)})\), where \(P\) is the actual distribution of the hidden layer expression vector, and \(Q\) is the hypothesis distribution model. The server determines the relevant parameters for calculating the fourth cost parameter based on the mean and the covariance matrix of the distribution model parameters of the hidden layer expression vector, and the covariance matrix related to the mean and the shape vector corresponding to the hypothesis distribution model.

[0169] For example, assume that the mean of the actual distribution of the hidden layer expression vectors is (mu_p), and the covariance matrix is (Sigma_p). Assume that the mean of the distribution model is (mu_q), and the covariance matrix is (Sigma_q). Calculate the fourth cost parameter (D_{KL}):

[0170] (D_{KL}=frac{1}{2}(logfrac{vertSigma_qvert}{vertSigma_pvert}-n+tr(Sigma_q^{-1}Sigma_p)+(mu_q-mu_p)^TSigma_q^{-1}(mu_q-mu_p)))

[0171] The calculated fourth cost parameter may be a specific value, assumed to be 0.3. Then divide this value by the reference value calculated based on the prior example BIM model (assumed to be 1) for normalization, generating the normalized fourth cost parameter as (0.3). This fourth cost parameter reflects the discrete parameter value of the hidden layer expression vector compared to the distribution model parameters.

[0172] Step S1325: Determine the probability model optimization cost parameter based on the first cost parameter, the second cost parameter, the third cost parameter, and the fourth cost parameter. The probability model optimization cost parameter reflects the feature distance between the iterative BIM model generated based on the prior data of the BIM model safety index and the example BIM model.

[0173] Step S1326: Determine the training cost parameter based on the fitness evaluation cost parameter and the probability model optimization cost parameter.

[0174] In this embodiment, the server determines the probability model optimization cost parameter based on the first cost parameter, the second cost parameter, the third cost parameter, and the fourth cost parameter. For example, assume that the first cost parameter is 2.3, the second cost parameter is 0.14, the third cost parameter is 0.045, and the fourth cost parameter is 0.3.

[0175] The probability model optimization cost parameter can be the weighted sum or some combination of these cost parameters. Assume that the simple summation method is used, and the probability model optimization cost parameter is (2.3 + 0.14 + 0.045 + 0.3 = 2.785). This probability model optimization cost parameter reflects the feature distance between the iterative BIM model generated based on the prior data of the BIM model safety index and the example BIM model.

[0176] The server determines the training cost parameter based on the fitness evaluation cost parameter and the probability model optimization cost parameter. Assume that the fitness evaluation cost parameter is 0.03 and the probability model optimization cost parameter is 2.785.

[0177] The training cost parameter can also be the weighted sum or some combination of these two cost parameters. If the simple summation method is adopted, the training cost parameter is (0.03 + 2.785 = 2.815). This training cost parameter comprehensively considers various factors such as estimation accuracy (fitness evaluation cost parameter) and model feature distance (probability model optimization cost parameter). It will be used to evaluate the performance of the AI security optimization network and guide the optimization process of network parameters.

[0178] Throughout the process, when the BIM model security index estimation data of the sample BIM model continuously updates and changes with the learning of the network parameter information of the AI security optimization network, the collaborative parameter array and the average parameter value of the prior data of the BIM model security index will follow and continuously update and change. For example, as the network learning progresses, the estimation of each security index by the estimation data becomes more accurate or changes, which will affect the relationship between each security index in the prior data (collaborative parameter array) and the overall average level (average parameter value), thereby affecting the calculation of the second cost parameter and the third cost parameter, and ultimately affecting the determination of the entire training cost parameter and the optimization process of the AI security optimization network.

[0179] In a possible implementation manner, step S1324 includes:

[0180] Step S1324-1: Obtain the hidden layer expression vector generated for the sample BIM model from the hidden network layer of the AI security optimization network. The hidden layer expression vector is the result of the sample BIM model being calculated and transformed by the hidden layer neurons.

[0181] Step S1324-2: Determine the mean and shape vector in the distribution model parameters followed by the hidden layer expression vector. The mean reflects the central position of the hidden layer expression vector in the current state distribution, and the shape vector is related to the distribution shape.

[0182] Step S1324-3: Select a hidden layer expression vector sample set based on the mean and shape vector in the distribution model parameters followed by the hidden layer expression vector, and obtain the pre-constructed hypothesis distribution model. The hypothesis distribution model represents the distribution pattern that the hidden layer expression vector should follow under the hypothesis state distribution.

[0183] Step S1324-4: Use the KL divergence as a measure of the difference between the distribution model parameters of the hidden layer expression vectors in the hidden layer expression vector sample set and the distribution model parameters of the assumed distribution model. Determine the relevant parameters for calculating the fourth cost parameter based on the mean and covariance matrix of the distribution model parameters of the hidden layer expression vectors and the covariance matrix related to the mean and shape vectors corresponding to the assumed distribution model.

[0184] Step S1324-5: Based on the calculation method of the KL divergence and the relevant parameters, calculate the fourth cost parameter by comprehensively considering the difference between the mean and covariance matrix of the distribution model parameters of the hidden layer expression vectors and the covariance matrix related to the mean and shape vectors corresponding to the assumed distribution model.

[0185] Step S1324-6: Divide the calculated fourth cost parameter by the reference value calculated based on the prior example BIM model for normalization to generate the normalized fourth cost parameter.

[0186] In this embodiment, in the BIM model security management system of a construction project, the AI security optimization network in the server has a complex structure. The hidden network layer of this network contains multiple neurons that process the data of the example BIM model. Taking the previously mentioned office area BIM model as an example, this model contains information in multiple aspects such as structure, fire protection, and electricity.

[0187] When the data of this example BIM model is input into the AI security optimization network, it first passes through the input layer and then enters the hidden network layer. In the hidden network layer, each neuron calculates the input data according to its preset weights and activation functions. For example, for data related to structural stability, the neuron may calculate based on information such as the connection method of beams and columns and material strength. Assume that the hidden network layer has 100 neurons, each neuron receives different combinations of information from the input layer and performs a linear combination calculation (such as (z = sum_{i = 1}^{n} w_{i}x_{i}+b), where (w_{i}) are the weights, (x_{i}) are the input data, and (b) is the bias), and then undergoes a non-linear transformation through an activation function (such as the ReLU function (y = max(0,z))).

[0188] After the calculations and transformations of all neurons, a hidden layer expression vector for the sample BIM model is generated. This hidden layer expression vector is a high-dimensional vector, assumed to be (\(\vec{v}=(v_1, v_2,\cdots, v_{50}\)) (assuming the dimension is 50 here), and it is the representation of the sample BIM model in the abstract feature space of the hidden network layer. The value of each dimension represents the measurement of the sample BIM model on a certain abstract feature. For example, \(v_1\) may be related to the abstract representation of the overall stability of the structure in the hidden layer, and \(v_2\) may be related to a certain hidden feature of fire safety, etc.

[0189] The server starts to determine the mean value among the distribution model parameters followed by the hidden layer expression vector. The mean value is a parameter that describes the central position of the hidden layer expression vector in the current state distribution. For the obtained hidden layer expression vector \(\vec{v}=(v_1, v_2,\cdots, v_{50}\)), the mean value \(\mu = (\mu_1, \mu_2,\cdots, \mu_{50})\) is calculated.

[0190] The method for calculating the mean value is to average the multiple calculation results of the hidden layer expression vector (assuming there are 1000 calculation results). For example, for \(\mu_1\), \(\mu_1=\frac{1}{1000}\sum_{j = 1}^{1000}v_{1j}\), where \(v_{1j}\) is the value of the first dimension of the hidden layer expression vector obtained in the \(j\)-th calculation. The same method is used to calculate \(\mu_2\) to \(\mu_{50}\). Assume that the calculated mean value \(\mu=(0.5, 0.3,\cdots, 0.4)\). Each element of this mean value vector reflects the average value of the hidden layer expression vector in the corresponding dimension, and it represents the central position of the hidden layer expression vector in this dimension.

[0191] The shape vector is related to the distribution shape of the hidden layer expression vector. Here, the covariance matrix is used as the shape vector. The element \(\Sigma_{ij}\) of the covariance matrix \(\Sigma\) represents the correlation between different dimensions of the hidden layer expression vector.

[0192] The formula for calculating the covariance matrix is (Sigma_{ij}=frac{1}{1000 - 1}sum_{k = 1}^{1000}(v_{ik}-mu_i)(v_{jk}-mu_j)), where (v_{ik}) is the value of the (i)-th dimension of the hidden layer expression vector obtained in the (k)-th calculation, and (mu_i) is the (i)-th element of the mean vector. A (50times50) covariance matrix (Sigma) is calculated through this formula. This covariance matrix describes the linear relationship between different dimensions of the hidden layer expression vector. For example, if the value of (Sigma_{12}) is large and positive, it indicates that there is a strong positive correlation between the two dimensions (v_1) and (v_2), that is, when the value of (v_1) is large, the value of (v_2) is more likely to be large.

[0193] Based on the mean and shape vector (covariance matrix) in the distribution model parameters followed by the hidden layer expression vector, the server selects a sample set of the hidden layer expression vector. A common method is to construct a multivariate normal distribution (N(mu, Sigma)) based on the mean and covariance matrix, and then randomly sample from this distribution to obtain the sample set. Suppose 100 samples are selected to obtain the sample set of the hidden layer expression vector ({vec{v}_1, vec{v}_2,cdots, vec{v}_{100}}), and each (vec{v}_i) is a vector with the same dimension as the previous hidden layer expression vector (vec{v}). These samples reflect the distribution characteristics of the hidden layer expression vector to a certain extent and will be used for subsequent comparison with the assumed distribution model.

[0194] The server obtains a pre-constructed assumed distribution model. This assumed distribution model represents the distribution pattern that the hidden layer expression vector should follow under the assumed state. In this example of building safety management, the assumed distribution model may be an ideal distribution model obtained based on a large amount of prior building model data or theoretical analysis. For example, assume it is a multivariate normal distribution (N(mu_0, Sigma_0)) with a mean of (mu_0=(0.4, 0.3,cdots, 0.3)) and a covariance matrix of (Sigma_0) (assuming (Sigma_0) is a fixed covariance matrix determined according to prior knowledge, representing the ideal relationship between dimensions). This assumed distribution model reflects the distribution characteristics that the hidden layer expression vector should have under ideal conditions, such as the distribution of the hidden layer expression vector when achieving an optimal balance in aspects such as structural safety and fire safety.

[0195] The server uses the Kullback - Leibler divergence as a measure of the difference between the distribution model parameters of the hidden layer expression vectors in the hidden layer expression vector sample set and the distribution model parameters of the assumed distribution model.

[0196] According to the calculation formula of the KL divergence:

[0197] (D_{KL}(P||Q)=\sum_{i}P(i)\log\frac{P(i)}{Q(i)}), where (P) is the actual distribution of the hidden layer expression vectors (reflected by the sample set, and its distribution parameters are the mean (\(\mu\)) and the covariance matrix (\(\Sigma\))), and (Q) is the assumed distribution model (its distribution parameters are the mean (\(\mu_0\)) and the covariance matrix (\(\Sigma_0\))).

[0198] To calculate the KL divergence, relevant parameters for calculating the fourth cost parameter need to be determined. According to the calculation formula of the KL divergence for multivariate normal distribution:

[0199] (D_{KL}=\frac{1}{2}(\log\frac{\vert\Sigma_0\vert}{\vert\Sigma\vert}-n + tr(\Sigma_0^{-1}\Sigma)+(\mu_0 - \mu)^T\Sigma_0^{-1}(\mu_0 - \mu))), and the relevant parameters here include \(\mu\), \(\Sigma\), \(\mu_0\), \(\Sigma_0\) as well as the operation rules of vectors and matrices.

[0200] For example, \(\vert\Sigma\vert\) is the determinant of the covariance matrix \(\Sigma\), \(tr(\Sigma_0^{-1}\Sigma)\) is the trace of the matrix \(\Sigma_0^{-1}\Sigma\) (i.e., the sum of the main diagonal elements), and \((\mu_0 - \mu)^T\) is the transpose of the vector \(\mu_0 - \mu\). These parameters are all obtained based on the actual distribution parameters (mean \(\mu\) and covariance matrix \(\Sigma\)) of the previously determined hidden layer expression vectors and the distribution parameters (mean \(\mu_0\) and covariance matrix \(\Sigma_0\)) of the assumed distribution model.

[0201] Based on the calculation method of the KL divergence and the previously determined relevant parameters, the server calculates the fourth cost parameter by synthesizing the differences between the mean and covariance matrix of the distribution model parameters of the hidden layer expression vectors and the mean and form vectors (covariance matrix) corresponding to the assumed distribution model.

[0202] According to the above - mentioned calculation formula of the KL divergence for multivariate normal distribution:

[0203] (D_{KL}=frac{1}{2}(logfrac{vertSigma_0vert}{vertSigmavert}-n + tr(Sigma_0^{-1}Sigma)+(mu_0 - mu)^TSigma_0^{-1}(mu_0 - mu))). Substitute the actually calculated (mu=(0.5, 0.3,cdots, 0.4)), (Sigma) (the covariance matrix calculated previously), (mu_0=(0.4, 0.3,cdots, 0.3)), and (Sigma_0) (the assumed covariance matrix) into the formula for calculation.

[0204] Suppose it is calculated that (D_{KL}=0.5) (this is just an example result here). This value reflects the degree of difference between the actual distribution of the hidden layer expression vectors and the assumed distribution model. If the value of (D_{KL}) is large, it indicates that the distribution of the hidden layer expression vectors is quite different from the ideal assumed distribution model, which may mean that there are certain biases in the representation of the AI security optimization network in the hidden layer and the network parameters need to be further adjusted.

[0205] The server normalizes the calculated fourth cost parameter by dividing it by the reference value calculated based on the prior example BIM model to generate the normalized fourth cost parameter.

[0206] The method for calculating the reference value based on the prior example BIM model can be the average KL divergence value or other statistics obtained by performing similar calculations on multiple prior example BIM models. Suppose the reference value calculated based on the prior example BIM model is 1.

[0207] Then the normalized fourth cost parameter is (0.5 / 1 = 0.5). The normalized fourth cost parameter can be more conveniently compared and combined with other cost parameters (such as the first cost parameter, the second cost parameter, the third cost parameter, etc.) to determine the probability model optimization cost parameter and the final training cost parameter, so as to evaluate the performance of the AI security optimization network and guide the optimization of the network parameters.

[0208] In a possible implementation manner, the AI security optimization network includes a first optimization branch and a second optimization branch.

[0209] The first optimization branch is used to generate a hidden layer expression vector based on the sample BIM model and the prior data of the BIM model security index of the sample BIM model. The average data and morphological vector of the distribution model parameters followed by the hidden layer expression vector are determined by the network parameter information of the first optimization branch, the sample BIM model, and the prior data of the BIM model security index of the sample BIM model.

[0210] The second optimization branch is used to generate the iterative BIM model based on the hidden layer expression vector and the prior data of the BIM model security index of the sample BIM model. The average data and morphological vector of the distribution model parameters followed by the iterative BIM model are determined by the network parameter information of the second optimization branch, the hidden layer expression vector, and the prior data of the BIM model security index of the sample BIM model.

[0211] In this embodiment, in the BIM model security management scenario of a construction project, the AI security optimization network in the server starts to work. For the previously mentioned sample BIM model of the office area, this sample BIM model of the office area contains various information related to security indicators, such as structural stability, fire protection rating, and good electrical grounding. At the same time, there is already prior data of the BIM model security index of this sample BIM model, which combines accurate labeled security index parameters (when available) and estimated security index parameters (when labeled parameters are missing).

[0212] The first optimization branch receives this sample BIM model and its prior data of the BIM model security index as inputs. These input data contain rich information. For example, in terms of structural stability, the prior parameters of the security index in the prior data may include information such as the dimensions of key components of the structure and the material strength; in terms of fire protection rating, it may include information such as the fire protection performance rating of building materials and the setting of fire partitions.

[0213] There is a series of hidden layer neurons inside the first optimization branch. These neurons start to process the input data. Taking the data related to structural stability as an example, the neurons will perform a linear combination calculation according to the information such as component dimensions and material strength in the prior data according to the preset weights. Suppose for a neuron related to structural stability, its calculation method is (z = w_1\times\text{component dimension}+ w_2\times\text{material strength}+b), where (w_1) and (w_2) are weights and (b) is the bias. This calculation result (z) then undergoes a non-linear transformation through an activation function (such as the Sigmoid function (y = \frac{1}{1 + e^{-z}})) to obtain an intermediate result.

[0214] For data related to the fire protection rating, the neurons also perform similar calculations. For example, based on information such as the fire protection performance rating of building materials and the setting of fire partitions, corresponding intermediate results are calculated through different weight combinations and activation functions. This process is carried out in parallel among multiple hidden layer neurons in the first optimization branch, and each neuron focuses on different parts or different combinations of features of the input data.

[0215] After the calculations and transformations by the hidden layer neurons, the first optimization branch generates a hidden layer expression vector. This hidden layer expression vector is a multi-dimensional vector, for example, \(\vec{v}=(v_1, v_2,\cdots, v_{n})\), where the value of each dimension represents the measurement of the sample BIM model on a certain abstract feature, and these abstract features are obtained through complex calculations and transformations of the hidden layer neurons on the original input data (sample BIM model and prior data).

[0216] The average data (mean) and the shape vector (taking the covariance matrix as an example here) of the distribution model parameters followed by the hidden layer expression vector are determined by the network parameter information of the first optimization branch, the sample BIM model, and the prior data of the BIM model safety index of the sample BIM model.

[0217] The process of calculating the mean is as follows: For the multiple calculation results of the hidden layer expression vector (assuming \(m\) calculations are performed), calculate the average value of each dimension. For example, for the \(v_1\) dimension, the mean \(\mu_1=\frac{1}{m}\sum_{i = 1}^{m}v_{1i}\), where \(v_{1i}\) is the value of the first dimension of the hidden layer expression vector obtained in the \(i\)-th calculation. The entire mean vector \(\mu = (\mu_1, \mu_2,\cdots, \mu_{n})\) is obtained through similar calculations.

[0218] When calculating the covariance matrix (shape vector), the formula \(\Sigma_{ij}=\frac{1}{m - 1}\sum_{k =1}^{m}(v_{ik}-\mu_i)(v_{jk}-\mu_j)\) is used, where \(\Sigma_{ij}\) is the element in the \(i\)-th row and \(j\)-th column of the covariance matrix, \(v_{ik}\) is the value of the \(i\)-th dimension of the hidden layer expression vector obtained in the \(k\)-th calculation, and \(\mu_i\) is the \(i\)-th element of the mean vector. Through this calculation, the covariance matrix \(\Sigma\) is obtained, which describes the correlation between different dimensions of the hidden layer expression vector. These distribution model parameters (mean and covariance matrix) reflect the distribution characteristics of the hidden layer expression vector in the first optimization branch and will affect the operations of the subsequent second optimization branch and the performance evaluation of the entire AI security optimization network.

[0219] The second optimization branch receives as input the hidden layer expression vector from the first optimization branch and the prior data of the BIM model safety indicators of the sample BIM model. This hidden layer expression vector contains the representation of the sample BIM model in the abstract feature space, while the prior data provides important information about the safety indicators of the sample BIM model.

[0220] The second optimization branch also has its own network structure internally, including multiple neuron layers. These neuron layers will generate an iterative BIM model based on the input data.

[0221] For each dimension value in the hidden layer expression vector, the neurons in the second optimization branch will calculate according to the preset weights and the corresponding safety indicator prior parameters in the prior data. For example, for the dimension of the hidden layer expression vector related to structural stability and the prior data of structural stability, the neuron may perform the following calculation: (z' = w'_1timesv_{s}+ w'_2times \text{prior parameter of structural stability}+b'), where (w'_1) and (w'_2) are weights, (v_{s}) is the dimension value related to structural stability in the hidden layer expression vector, and (b') is the bias. The calculation result (z') undergoes a non-linear transformation through an activation function (such as the ReLU function (y' = max(0,z'))) to obtain an intermediate result.

[0222] For other safety indicators, such as fire rating, good electrical grounding, etc., similar calculations will also be performed. These calculation processes are carried out step by step in multiple neuron layers of the second optimization branch, continuously transforming and integrating the input data.

[0223] After a series of neuron calculations, the second optimization branch generates an iterative BIM model. This iterative BIM model is an optimization or adjustment result of the sample BIM model. For example, in terms of structure, the iterative BIM model may adjust the dimensions or material selections of some components; in terms of fire protection, it may change the layout of fire partitions or use different fire protection materials, etc.

[0224] The average data (mean) and shape vector (covariance matrix) of the distribution model parameters followed by the iterative BIM model are determined by the network parameter information of the second optimization branch, the hidden layer expression vector, and the prior data of the BIM model safety indicators of the sample BIM model.

[0225] The process of determining the mean is similar to that of the first optimization branch, but based on the calculation results of the second optimization branch. Suppose a certain feature of the iterative BIM model is calculated (p) times, and the results are obtained as (x_1, x_2,cdots, x_{p}). For this feature, the mean is (mu'=frac{1}{p}sum_{i = 1}^{p}x_{i}). Such calculations are performed on multiple features of the iterative BIM model to obtain a mean vector.

[0226] When calculating the covariance matrix, a similar formula is also used to calculate the covariance matrix according to the relationship between the calculation results of different features of the iterative BIM model. This covariance matrix describes the correlation between the features of the iterative BIM model and reflects the distribution characteristics of the iterative BIM model. These distribution model parameters are of great significance for evaluating the relationship between the iterative BIM model and the sample BIM model and the performance of the entire AI security optimization network.

[0227] Through the collaborative work of the first optimization branch and the second optimization branch, the AI security optimization network can generate an iterative BIM model based on the sample BIM model and its prior data, and by determining the hidden layer expression vector and the distribution model parameters of the iterative BIM model, provide an important basis for subsequent network parameter learning and model optimization to improve the security performance and optimization ability of the BIM model.

[0228] In a possible implementation manner, the BIM model security index estimation data of the sample BIM model is generated by a BIM model security index estimation network, and the method further includes:

[0229] A. When the number of BIM models in the partially missing security identification data is greater than the threshold number, the BIM model security index estimation network calls a first deep learning network that has not performed network parameter pre-learning, and during the network parameter learning process of the AI security optimization network, optimizes the network parameters of the first deep learning network according to the loss function value between the BIM model security index annotation data of the sample BIM model and the BIM model security index estimation data of the sample BIM model.

[0230] Or, B. When the number of BIM models in the partially missing security identification data is not greater than the threshold number, the BIM model security index estimation network calls a second deep learning network after network parameter pre-learning. The second deep learning network is an artificial intelligence network with BIM model security index estimation performance generated after network parameter pre-learning using the complete security identification data of the sample BIM model sequence. The complete security identification data includes multiple sample BIM models and corresponding security index parameters.

[0231] In this embodiment, in the BIM model safety management system of a construction project, the server is responsible for the management and data processing of the entire process. Still taking the example of the BIM model of the office area mentioned before, which includes multiple safety indicators (such as structural stability, fire protection level, good electrical grounding, etc.).

[0232] For each sample BIM model, there is some missing safety identification data, which includes multiple BIM model safety indicator annotation data, and there are some missing data among them.

[0233] Suppose the server sets a threshold number, for example, 50. When the number of BIM models in the partially missing safety identification data is greater than this threshold number, the server adopts a specific strategy.

[0234] Suppose in the current construction project, the number of BIM models in the partially missing safety identification data is 60, which is greater than 50. At this time, the BIM model safety indicator estimation network will call the first deep learning network that has not yet performed pre-learning of network parameters.

[0235] This first deep learning network is a network in an initial state. It has not undergone pre-learning and has not been optimized for the BIM model safety indicator estimation task. It may have some default network structures, such as including multiple hidden layers, each layer having a certain number of neurons (such as 3 hidden layers, with 100, 50, and 20 neurons in each layer respectively), and using the default weight initialization method (such as random initialization).

[0236] After the first deep learning network is called, it starts to estimate the BIM model safety indicators of the sample BIM model. Taking the structural stability indicator as an example, the first deep learning network will receive relevant feature data of the sample BIM model, such as the geometric shape of the building, the type of structural material, etc. as input.

[0237] The neurons in the network will perform calculations based on the input data. In the input layer, the data will be distributed according to the number of neurons. For example, if the input data is a vector (\(\vec{x}=(x_1, x_2,\cdots, x_{n}\))), and the input layer has 100 neurons, then each neuron will receive a part of the input data or the input data after a certain transformation.

[0238] Then, the data is propagated and computed in the hidden layer. For the first hidden layer, assume the computation of the neuron is \((z_{ij}=\sum_{k = 1}^{n} w_{ijk}x_{k}+b_{ij})\), where \((w_{ijk})\) is the weight of the \(i\)-th neuron in the \(j\)-th layer, \((x_{k})\) is the \(k\)-th element of the input data, and \((b_{ij})\) is the bias of the \(i\)-th neuron in the \(j\)-th layer. Then, it undergoes a non-linear transformation through an activation function (such as the ReLU function \(y_{ij} = max(0,z_{ij})\)).

[0239] After the computation through multiple hidden layers, the estimated safety index parameters of structural stability are obtained in the output layer. This estimation process is performed for each safety index of the sample BIM model.

[0240] During the network parameter learning process of the AI security optimization network, the server optimizes the network parameters of the first deep learning network based on the loss function values between the BIM model safety index annotation data and the BIM model safety index estimated data of the sample BIM model.

[0241] For example, using the mean squared error (MSE) as the loss function, for the structural stability index, if the annotated safety index parameter is \((s)\) and the estimated safety index parameter is \((\hat{s})\), then the loss function value is \(L=(s - \hat{s})^2\). For all safety indices, the total loss function value is calculated.

[0242] Then, based on this loss function value, the weights and biases of the first deep learning network are updated using an optimization algorithm (such as the stochastic gradient descent algorithm). In the stochastic gradient descent algorithm, the update formulas for the weights and biases are \(w_{ijk}=w_{ijk}-alpha\frac{\partial L}{\partial w_{ijk}}\) and \(b_{ij}=b_{ij}-alpha\frac{\partial L}{\partial b_{ij}}\), where \((alpha)\) is the learning rate, which determines the step size of each update. By continuously iterating this process, the network parameters of the first deep learning network are optimized, thereby improving the estimation performance of the BIM model safety index.

[0243] Suppose in another construction project or different stages of the same project, the number of BIM models in the partially missing safety identification data is 40, which is not greater than 50. At this time, the BIM model safety index estimation network will call the second deep learning network after network parameter pre-learning.

[0244] This second deep learning network is an artificial intelligence network with the performance of estimating BIM model safety indicators, which is generated after pre-learning network parameters using the complete safety recognition data of a sequence of sample BIM models. This complete safety recognition data includes various sample BIM models and corresponding safety indicator parameters. For example, this complete safety recognition data may contain sample BIM models from different building areas (such as office areas, commercial areas, residential areas, etc.), and each sample BIM model has accurate annotation parameters for safety indicators such as structural stability, fire protection level, and good electrical grounding.

[0245] During the pre-learning process, the second deep learning network has learned the relationship patterns between different sample BIM models and safety indicators. When it is called to estimate the safety indicators of the current sample BIM model, it can estimate more accurately according to the characteristics of the sample BIM model. For example, for a new sample BIM model of an office area, the second deep learning network can quickly and accurately estimate the estimated safety indicator parameters of each safety indicator based on the structural characteristics and safety indicator relationships of the sample BIM models of the office area learned previously, without the need for a learning and optimization process from scratch like the first deep learning network.

[0246] Through such a mechanism, by selecting different deep learning networks for BIM model safety indicator estimation according to the relationship between the number of BIM models and the threshold number in the partially missing safety recognition data, and optimizing network parameters when necessary, the efficiency and accuracy of BIM model safety indicator estimation can be improved, thus providing better support for the operation of the entire AI safety optimization network and the safety management of BIM models.

[0247] Figure 2 The hardware structure diagram of the smart construction site service system 100 provided by the embodiment of the present application for implementing the above-mentioned AI safety recognition method based on BIM and smart construction site is shown, as Figure 2 As shown, the smart construction site service system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0248] In a possible design, the intelligent construction site service system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the intelligent construction site service system 100 can be a distributed system). In some embodiments, the intelligent construction site service system 100 can be local or remote. For example, the intelligent construction site service system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. Alternatively, for example, the intelligent construction site service system 100 can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the intelligent construction site service system 100 can be implemented on the intelligent construction site service system. By way of example only, the intelligent construction site service system can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any aggregation thereof.

[0249] The machine-readable storage medium 120 can store data and / or instructions. In some embodiments, the machine-readable storage medium 120 can store data obtained from an external terminal. In some embodiments, the machine-readable storage medium 120 can store the data and / or instructions that the intelligent construction site service system 100 uses to execute or complete the exemplary methods described in this application.

[0250] In a specific implementation process, one or more processors 110 execute computer-executable instructions stored in the machine-readable storage medium 120, so that the processors 110 can execute the AI security recognition method based on BIM and the intelligent construction site as described in the above method embodiments. The processors 110, the machine-readable storage medium 120, and the communication unit 140 are connected through a bus 130, and the processors 110 can be used to control the sending and receiving actions of the communication unit 140.

[0251] For the specific implementation process of the processors 110, reference can be made to the various method embodiments executed by the above intelligent construction site service system 100. Their implementation principles and technical effects are similar, and will not be elaborated here in this embodiment.

[0252] In addition, an embodiment of the present application also provides a readable storage medium, in which computer-executable instructions are set. When a processor executes the computer-executable instructions, the above AI security recognition method based on BIM and the intelligent construction site is implemented.

[0253] It should be noted that, in order to simplify the description of the disclosure of the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing or description thereof. Similarly, it should be noted that, in order to simplify the description of the disclosure of the present application and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, sometimes multiple features are merged into one embodiment, drawing or description thereof.

Claims

1. An AI safety identification method based on BIM and smart construction sites, characterized in that: The method comprises: Obtaining partially missing safety identification data of a sample BIM model sequence learned by an AI safety optimization network, wherein the partially missing safety identification data includes X BIM model safety indicator annotation data, each BIM model safety indicator annotation data includes annotated safety indicator parameters of a BIM model for Y safety indicators, and at least one BIM model safety indicator annotation data with missing features exists among the X BIM model safety indicator annotation data, and at least one labeled safety indicator parameter of a safety indicator is missing among the labeled safety indicator parameters of the Y safety indicators included in the BIM model safety indicator annotation data with missing features; For each sample BIM model, if the BIM model safety index annotation data corresponding to the sample BIM model in the partially missing safety identification data is feature-missing, then based on the BIM model safety index estimation data and the BIM model safety index annotation data of the sample BIM model, the BIM model safety index prior data of the sample BIM model is generated, and the BIM model safety index estimation data of the sample BIM model includes: estimated safety index parameters of the sample BIM model for at least one safety index generated by estimation; Based on the BIM model safety indicator prior data of the sample BIM model, network parameter learning is performed on the AI ​​safety optimization network to generate an AI safety optimization network that has completed network parameter learning. The AI ​​safety optimization network that has completed network parameter learning is used to optimize and generate BIM model data with target safety identification indicators, thereby providing data support for the safety management of smart construction sites.

2. The AI ​​safety identification method based on BIM and smart construction site according to claim 1 is characterized in that: The step of generating BIM model safety index prior data of the sample BIM model based on the BIM model safety index estimation data and the BIM model safety index annotation data of the sample BIM model includes: For each of the Y safety indicators, if the annotated safety indicator parameter of the safety indicator covered in the BIM model safety indicator annotated data of the sample BIM model is missing, obtaining the estimated safety indicator parameter of the safety indicator from the BIM model safety indicator estimation data of the sample BIM model as the safety indicator priori parameter of the safety indicator; If there is no missing annotated safety indicator parameter of the safety indicator covered in the BIM model safety indicator annotated data of the sample BIM model, outputting the annotated safety indicator parameter of the safety indicator as a safety indicator priori parameter of the safety indicator; Based on the safety index prior parameters of the Y safety indexes, BIM model safety index prior data of the sample BIM model is generated.

3. The AI ​​safety identification method based on BIM and smart construction site according to claim 2 is characterized in that: The method further comprises: A missing indication array corresponding to the partially missing safety identification data is generated, wherein the missing indication array includes X×Y nodes, and each node represents whether a marked safety indicator parameter of a safety indicator of a BIM model is missing.

4. The AI ​​safety identification method based on BIM and smart construction site according to claim 1 is characterized in that: The obtaining of partially missing safety identification data of the sample BIM model sequence learned by the AI ​​safety optimization network includes: Acquire complete safety identification data of multiple sample BIM model sequences, wherein the complete safety identification data of each sample BIM model sequence includes a labeled safety indicator parameter of at least one BIM model for at least one safety indicator, and the safety indicators covered by the complete safety identification data of different sample BIM model sequences are all different, and the BIM models covered by the complete safety identification data of different sample BIM model sequences are all different; Taking each BIM model as a unit, the complete safety identification data of the multiple sample BIM model sequences are merged to generate the partially missing safety identification data.

5. The AI ​​safety identification method based on BIM and smart construction site according to claim 1 is characterized in that: The method of performing network parameter learning on the AI ​​safety optimization network based on the BIM model safety index prior data of the sample BIM model to generate an AI safety optimization network that has completed network parameter learning includes: Generate an iterative BIM model corresponding to the sample BIM model based on the AI ​​safety optimization network, based on the sample BIM model and the BIM model safety index prior data of the sample BIM model; Based on the sample BIM model and the iterative BIM model, determining a training cost parameter of the AI ​​safety optimization network, wherein the training cost parameter reflects a characteristic distance between the sample BIM model and the iterative BIM model; Based on the training cost parameters, the network parameters of the AI ​​security optimization network are optimized to generate the AI ​​security optimization network that has completed network parameter learning.

6. The AI ​​safety identification method based on BIM and smart construction site according to claim 5 is characterized in that: The training cost parameters include a fitness evaluation cost parameter and a probability model optimization cost parameter, and the probability model optimization cost parameter includes a first cost parameter, a second cost parameter, a third cost parameter and a fourth cost parameter; The determining, based on the sample BIM model and the iterative BIM model, a training cost parameter of the AI ​​safety optimization network includes: Determine the goodness of fit evaluation cost parameter based on the labeled safety indicator parameters of the labeled safety indicators in the BIM model safety indicator labeled data and the estimated safety indicator parameters of the labeled safety indicators in the BIM model safety indicator estimation data, wherein the goodness of fit evaluation cost parameter reflects the estimated performance indicator of the BIM model safety indicator estimation data, and the labeled safety indicator is a safety indicator corresponding to the labeled safety indicator parameters that are not missing in the BIM model safety indicator labeled data; Determine the first cost parameter based on the iterative BIM model and the sample BIM model, where the first cost parameter reflects the explicit feature distance between the iterative BIM model and the sample BIM model; Based on the BIM model safety indicator prior data, the collaborative parameter array of the BIM model safety indicator prior data, the average data of the BIM model safety indicator prior data, and the average data of the BIM model safety indicator estimation data, the second cost parameter and the third cost parameter are determined, the second cost parameter reflecting the characteristic distance between the safety indicator prior parameters of the annotated safety indicator in the BIM model safety indicator prior data and the BIM model safety indicator prior data, and the third cost parameter reflecting the characteristic distance between the safety indicator prior parameters of other safety indicators except the annotated safety indicator among the Y safety indicators covered by the BIM model safety indicator prior data and the BIM model safety indicator prior data; Determine the fourth cost parameter based on the hidden layer expression vector, and the morphological vector and the mean of the distribution model parameter followed by the hidden layer expression vector, wherein the fourth cost parameter reflects the discrete parameter value of the hidden layer expression vector compared to the distribution model parameter, and the hidden layer expression vector is a hidden relationship vector of the sample BIM model generated by the hidden network layer of the AI ​​safety optimization network; Based on the first cost parameter, the second cost parameter, the third cost parameter and the fourth cost parameter, determining the probability model optimization cost parameter, wherein the probability model optimization cost parameter reflects the characteristic distance between the iterative BIM model generated based on the BIM model safety index prior data and the sample BIM model; Determining the training cost parameter based on the fitness evaluation cost parameter and the probability model optimization cost parameter; Among them, when the BIM model safety index estimation data of the sample BIM model is continuously updated and changed as the network parameter information of the AI ​​safety optimization network is learned, the collaborative parameter array and average parameter value of the BIM model safety index prior data are continuously updated and changed accordingly.

7. The AI ​​safety identification method based on BIM and smart construction site according to claim 6 is characterized in that: The step of determining the fourth cost parameter based on the hidden layer expression vector, and the morphological vector and mean value of the distribution model parameter followed by the hidden layer expression vector comprises: Obtaining a hidden layer expression vector generated for the sample BIM model from the hidden network layer of the AI ​​safety optimization network, wherein the hidden layer expression vector is a result of hidden layer neuron calculation and conversion of the sample BIM model; Determine a mean value and a morphological vector in the distribution model parameters followed by the hidden layer expression vector, wherein the mean value reflects the central position of the hidden layer expression vector in the current state distribution, and the morphological vector is related to the distribution shape; Selecting a hidden layer expression vector sample set based on the mean and morphological vector in the distribution model parameters followed by the hidden layer expression vector, and obtaining a pre-constructed hypothetical distribution model, wherein the hypothetical distribution model represents the distribution pattern that the hidden layer expression vector should follow under the hypothetical state distribution; Using KL divergence as a measure of the difference between the distribution model parameters of the hidden layer expression vectors in the hidden layer expression vector sample set and the distribution model parameters of the assumed distribution model, determining the relevant parameters for calculating the fourth cost parameter according to the mean and covariance matrix of the distribution model parameters of the hidden layer expression vectors and the covariance matrix associated with the mean and morphological vector corresponding to the assumed distribution model; Based on the calculation method of the KL divergence and the related parameters, the fourth cost parameter is calculated by combining the mean and covariance matrix of the distribution model parameters of the hidden layer expression vector and the difference between the mean and covariance matrix related to the morphological vector corresponding to the assumed distribution model; The calculated fourth cost parameter is divided by a reference value calculated based on the prior sample BIM model for normalization to generate a normalized fourth cost parameter.

8. The AI ​​safety identification method based on BIM and smart construction site according to claim 5 is characterized in that: The AI ​​security optimization network includes a first optimization branch and a second optimization branch; The first optimization branch is used to generate a hidden layer expression vector based on the sample BIM model and the BIM model safety index prior data of the sample BIM model, wherein the average data and the morphological vector of the distribution model parameter followed by the hidden layer expression vector are determined by the network parameter information of the first optimization branch, the sample BIM model and the BIM model safety index prior data of the sample BIM model; The second optimization branch is used to generate the iterative BIM model based on the hidden layer expression vector and the BIM model safety index prior data of the sample BIM model. The average data and morphological vector of the distribution model parameters followed by the iterative BIM model are determined by the network parameter information of the second optimization branch, the hidden layer expression vector and the BIM model safety index prior data of the sample BIM model.

9. The AI ​​safety identification method based on BIM and smart construction site according to claim 1 is characterized in that: The BIM model safety index estimation data of the sample BIM model is generated by a BIM model safety index estimation network, and the method further includes: When the number of BIM models in the partially missing safety identification data is greater than a threshold number, the BIM model safety index estimation network calls a first deep learning network that has not yet performed network parameter pre-learning, and in the network parameter learning process of the AI ​​safety optimization network, the network parameters of the first deep learning network are optimized according to the loss function value of the BIM model safety index annotation data of the sample BIM model and the BIM model safety index estimation data of the sample BIM model; Alternatively, when the number of BIM models in the partially missing safety identification data is not greater than a threshold number, the BIM model safety index estimation network calls a second deep learning network after network parameter pre-learning, and the second deep learning network is an artificial intelligence network with BIM model safety index estimation performance generated after network parameter pre-learning using the complete safety identification data of a sample BIM model sequence, and the complete safety identification data includes a variety of sample BIM models and corresponding safety index parameters.

10. A smart construction site service system, characterized in that: The smart construction site service system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the AI ​​safety identification method based on BIM and smart construction site as described in any one of claims 1-9.

Images