Electromechanical engineering construction information management method and system and medium

Abstract

The invention discloses an electromechanical engineering construction information management method and system and a medium, and relates to the field of electromechanical engineering, and the method comprises the steps: collecting an electromechanical engineering construction site image set; performing feature recognition on the image set; according to an identification result, dividing construction personnel behavior risk levels and marking material attributes; the method comprises the following steps: dividing a construction process into a plurality of stages, constructing a Bayesian network model of a time slice in each stage, adding a time sequence dependency relationship between the Bayesian network models of adjacent time slices, and establishing dynamic Bayesian network risk evolution models of different stages of the construction process; predicting a risk distribution thermodynamic diagram of each construction stage in a future period of time by using a dynamic Bayesian network risk evolution model; and performing construction management according to the risk distribution thermodynamic diagram. For static risk assessment by adopting traditional Bayesian reasoning in the prior art, the accuracy and real-time performance of electromechanical engineering construction risk prediction are improved.

Description

Technical Field

[0001] The present application relates to the field of electromechanical engineering, and in particular to a method, system and medium for managing electromechanical engineering construction information. Background Art

[0002] With the rapid development of modern science and technology and the increasing complexity of engineering projects, electromechanical engineering construction is facing more and more risks and uncertainties. Timely and accurate identification and prediction of potential risks in the construction process is of great significance for ensuring project quality, controlling costs and construction period, and ensuring construction safety.

[0003] Traditional risk management methods, such as fault tree analysis and event tree analysis, mainly use static risk assessment models, set model parameters based on expert experience, and conduct qualitative or semi-quantitative analysis of construction risks. These methods are difficult to capture and predict the dynamic changes of risk factors during the construction process, and the subjectivity and uncertainty of the assessment results are relatively large, which cannot meet the growing demand for risk management in electromechanical engineering construction.

[0004] In recent years, Bayesian networks have been widely used in the field of risk assessment due to their advantages in uncertainty reasoning and causal modeling. By constructing a Bayesian network containing risk factor nodes and obtaining conditional probability parameters based on expert knowledge or data learning methods, the probability of risk events can be inferred and predicted. However, traditional Bayesian networks are mostly static models that assume that risk factors are independent of each other, making it difficult to characterize complex causal relationships and time dynamics, resulting in less than ideal accuracy in risk assessment and prediction. Summary of the invention

[0005] In view of the low assessment accuracy caused by the use of traditional Bayesian reasoning for static risk assessment in the prior art, the present application provides a method, system and medium for mechanical and electrical engineering construction information management, which improves the accuracy and real-time performance of mechanical and electrical engineering construction risk prediction by constructing a dynamic risk assessment and prediction model suitable for the mechanical and electrical engineering construction process.

[0006] The purpose of this application is achieved through the following technical solutions.

[0007] One aspect of the present application provides a method for managing mechanical and electrical engineering construction information, including: S1, collecting a set of images of mechanical and electrical engineering construction sites; S2, performing feature recognition on the image set, including recognizing the behavior of construction personnel and materials, and obtaining recognition results; S3, dividing the risk level of construction personnel behavior and marking material attributes according to the recognition results; S4, dividing the construction process into multiple stages, constructing a Bayesian network model of a time slice in each stage, adding a temporal dependency relationship between the Bayesian network models of adjacent time slices, and establishing a dynamic Bayesian network risk evolution model for different stages of the construction process; wherein, the risk level of construction personnel behavior and material attributes are used as Bayesian network nodes, conditional probability distributions between nodes are set, and the risk probability distribution under various combinations of factors is calculated by a Bayesian network inference algorithm; S5, using a dynamic Bayesian network risk evolution model to predict a risk distribution heat map of each stage of construction in the future; S6, performing construction management according to the risk distribution heat map. Traditional risk assessment methods usually only give an overall risk level, lacking an explanation of the causes and distribution of risks. This application generates a risk distribution heat map to intuitively display the risk level distribution in different time stages and different regions, and can trace the status changes of risk nodes, so that managers can quickly identify risk hotspots and formulate targeted management and control measures based on the causes of risks, thereby improving the interpretability and operability of risk assessment results.

[0008] Furthermore, S2 performs feature recognition on the image set, including recognition of construction worker behavior and materials, and obtains recognition results, including: S21, using a convolutional neural network to extract deep features of the image as a first feature; S22, extracting texture features, edge features, and color features of the image as a second feature; extracting only the texture, edge, and color features of the image may be difficult to fully characterize the semantic information of the construction worker behavior and materials; S23, fusing the first feature and the second feature to obtain a fused feature vector; S24, using the fused feature vector as input, using a pre-trained convolutional neural network to perform forward propagation calculations, and judging the construction worker behavior type and material type in the image according to the category probability of the output layer.

[0009] On the one hand, images of electromechanical construction sites usually contain rich semantic information, such as the behavior of construction workers, the type and state of materials, etc. It may be difficult to fully express these high-level semantic contents by only extracting low-level features of images (such as texture, edge, and color). Through multi-level feature extraction and abstraction, convolutional neural networks can automatically learn deep features in images and capture key information and intrinsic patterns of construction scenes. The fusion of deep features and low-level features can more comprehensively characterize the semantic features of electromechanical construction images.

[0010] On the other hand, the electromechanical construction environment is complex and changeable, and the behaviors of construction workers and the state of materials are diverse and changeable. Traditional feature extraction methods may have difficulty in dealing with this complexity, resulting in low recognition accuracy. Through end-to-end training, convolutional neural networks can adaptively learn and optimize feature representations to capture key patterns and discriminant information in construction scenes. The introduction of deep features can significantly improve the accuracy of behavior and material recognition, especially for small target detection and fine-grained recognition in complex backgrounds.

[0011] Furthermore, S3 divides the construction personnel's behavior risk levels and marks material attributes based on the identification results, including: dividing the behavior risks into three levels of low risk, medium risk and high risk according to the construction personnel's behavior types, wherein the behavior risks include not wearing safety equipment and improper operation; marking the types and quantities of materials according to the material types.

[0012] This application divides the behavioral risks of construction workers into three levels: low risk, medium risk, and high risk, and marks material properties as types and quantities. In essence, it converts continuous behavior and material characteristics into discrete variable representations. On the one hand, the Bayesian network is a probabilistic graphical model used to describe the conditional dependencies and joint probability distributions between variables. In a Bayesian network, each node represents a random variable, and the edges represent the causal relationship or correlation between variables. Bayesian networks usually assume that variables are discrete, and each variable has a finite number of possible states or values. Converting continuous behavioral risks and material properties into discrete levels and categories can meet the modeling requirements of the Bayesian network and facilitate subsequent network construction and parameter learning.

[0013] On the other hand, if continuous behavior and material characteristics are directly used as nodes of the Bayesian network, it is necessary to process continuous probability distributions, such as Gaussian distribution, exponential distribution, etc. In this case, the computational complexity of the inference algorithm is high, and it is difficult to achieve real-time risk assessment and prediction. Converting continuous features into discrete variables can significantly reduce the computational complexity of Bayesian reasoning. The conditional probability distribution of discrete variables can be represented by a finite-dimensional probability table or factor, and the reasoning process can be completed by table lookup and simple arithmetic operations, with low time complexity. Discretization can also reduce sensitivity to data quality and noise, and improve the stability and robustness of the reasoning results.

[0014] Finally, in the field of electromechanical construction, the assessment of personnel behavior risks and material properties usually relies on domain knowledge and expert experience. Dividing continuous features into discrete levels and categories can better utilize the prior knowledge and empirical rules of domain experts. For example, experts can summarize the risk level classification standards under different behavior modes based on experience, such as "not wearing a safety helmet" corresponds to high risk, "not wearing a seat belt" corresponds to medium risk, etc. Similarly, the classification of material types and quantities can also be based on professional standards and conventions. Introducing expert knowledge into the construction process of Bayesian networks can improve the accuracy and interpretability of risk assessment, and also facilitate the initialization and constraints of network parameters.

[0015] Furthermore, S4 divides the construction process into multiple stages, constructs a Bayesian network model of a time slice in each stage, adds a temporal dependency relationship between the Bayesian network models of adjacent time slices, and establishes a dynamic Bayesian network risk evolution model for different stages of the construction process, including: S41, divides the entire construction process into multiple continuous time slices according to the construction plan and progress; each time slice represents a construction stage; S42, in each time slice, divides the behavior risk of construction personnel into three levels of low risk, medium risk and high risk, as the behavior risk nodes of the Bayesian network; labels the type and quantity of materials according to the material type, as the material attribute nodes of the Bayesian network; constructs a Bayesian network based on the behavior risk nodes and material attribute nodes; S43, adds temporal dependency edges between the Bayesian networks of adjacent time slices to obtain a dynamic Bayesian network; wherein the temporal dependency edges represent the impact of the construction risk status of the previous moment on the current moment; S44, based on historical data, learns the conditional probability distribution parameters of the dynamic Bayesian network to obtain a dynamic Bayesian network model.

[0016] Further, S44, based on historical data, the conditional probability distribution parameters of the dynamic Bayesian network are learned to obtain a dynamic Bayesian network model, including: organizing and labeling the historical data according to time slices, counting the state distribution of the behavior risk nodes and material attribute nodes in each time slice in the historical data, and obtaining the empirical probability of the node state; for the time-dependent edges between adjacent time slices, counting the joint distribution of the node state at the previous moment and the node state at the current moment, and obtaining the empirical probability of state transition; based on the state distribution and the joint distribution, estimating the initial state transition probability between nodes in the dynamic Bayesian network as the initial value of the conditional probability distribution parameter; using the expectation-maximization EM algorithm and the random restart strategy to iteratively optimize the conditional probability distribution parameters of the dynamic Bayesian network; using the initial state transition probability to correct the iteratively optimized conditional probability distribution parameters to obtain the corrected conditional probability distribution parameters; using the final conditional probability distribution parameters to construct a dynamic Bayesian network model.

[0017] Furthermore, the expectation-maximization EM algorithm and the random restart strategy are used to iteratively optimize the conditional probability distribution parameters of the dynamic Bayesian network, including: setting the number of random restarts N, where random restarts represent reinitializing the parameters during the iterative optimization process; for each random restart, the initial state transition probability is randomly perturbed to generate the initial parameter value of the EM algorithm; for each random restart, the initial state transition probability is randomly perturbed to generate the initial parameter value of the EM algorithm. This step introduces random perturbations to the initial parameters so that the starting point of each restart is different, which increases the diversity of parameter searches and helps to escape from the local optimum and find a better solution. In step E, the expected log-likelihood function of the observed data is calculated using the current conditional probability distribution parameters; the observed data includes state distribution and state transition; the observed data of the present application includes state distribution and state transition, which makes full use of the timing information in the dynamic Bayesian network. In the M step, the expected log-likelihood function is maximized and the estimate of the conditional probability distribution parameters is updated; the E step and the M step are repeated until the conditional probability distribution parameters converge or the maximum number of iterations is reached; after completing N random restarts, the result with the largest expected log-likelihood function value is selected as the conditional probability distribution parameters after iterative optimization.

[0018] The traditional expectation maximization (EM) algorithm has defects such as local optimality and slow convergence speed in parameter estimation of dynamic Bayesian networks in electromechanical engineering, resulting in the learned conditional probability distribution parameters may not be the global optimal solution, affecting the accuracy of risk prediction. In this application, the number of random restarts N is set, which represents the number of times the parameters are reinitialized during the iterative optimization process. This step introduces a random restart mechanism. By reinitializing the parameters multiple times and running the EM algorithm independently, the parameter search range is expanded, the probability of finding the global optimal solution is increased, and the defect of the traditional EM algorithm that is prone to falling into the local optimal solution is overcome.

[0019] Furthermore, the initial state transition probability is used to correct the iteratively optimized conditional probability distribution parameters to obtain corrected conditional probability distribution parameters, including: using the initial state transition probability as a parameter of the Dirichlet distribution to construct a prior distribution of the conditional probability distribution parameters; using the iteratively optimized conditional probability distribution parameters as observation data; calculating the posterior distribution of the conditional probability distribution parameters according to the prior distribution and the observation data based on Bayes' theorem; and using the mean or mode of the posterior distribution as the corrected conditional probability distribution parameters.

[0020] In the construction information management of electromechanical engineering, due to the complexity and variability of the construction environment and the uncertainty of risk factors, the conditional probability distribution parameters obtained by relying solely on data-driven parameter learning methods (such as the EM algorithm) may deviate from the actual situation, affecting the risk prediction performance of the dynamic Bayesian network model. This application introduces the prior information of the conditional probability distribution parameters by taking the initial state transition probability as the parameter of the Dirichlet distribution, which makes up for the shortcomings of the pure data-driven method. The iteratively optimized conditional probability distribution parameters are used as observation data. The conditional probability distribution parameters obtained by learning the EM algorithm and the random restart strategy are fully utilized and used as observation data to reflect the risk evolution law mined from historical construction data. Using the basic principles of Bayesian inference, the prior distribution (expert experience) and the observed data (data-driven results) are integrated, and by calculating the posterior distribution of the conditional probability distribution parameters, the parameter estimation results that take into account both subjective knowledge and objective data are obtained, and the optimization and correction of the initial parameters are achieved.

[0021] Furthermore, S5 uses a dynamic Bayesian network risk evolution model to predict a risk distribution heat map of each construction stage in the future, including: S51, according to the construction plan and progress, determine the future time period that needs to be predicted, and divide the future time period into continuous time slices, each time slice representing a construction stage; S52, input the construction personnel behavior type and material type in each future time slice into the dynamic Bayesian network model as an evidence node; S53, according to the state of the evidence node, use the conditional probability distribution parameters to calculate the posterior probability distribution of other risk nodes; S54, according to the posterior probability distribution of the risk nodes in each future time slice, generate a time series probability distribution map of the construction risk level; S55, visualize the time series probability distribution map of the construction risk level as a risk distribution heat map.

[0022] Further, S53, according to the state of the evidence node, using the conditional probability distribution parameters, calculates the posterior probability distribution of other risk nodes, including: taking the state of the evidence node as a known condition, and constructing a chain graph of the dynamic Bayesian network according to the modified conditional probability distribution parameters; the chain graph is composed of multiple time slices, and the nodes in each time slice represent construction risk factors and material attribute factors; from the end time slice of the chain graph, according to the state of the evidence node and the conditional probability distribution parameters, calculate the posterior probability of other risk nodes; for the previous time slice, take the posterior probability of the risk node of the next time slice as the observed evidence, and combine the state of the evidence node and the conditional probability distribution parameters of the current time slice to calculate the posterior probability of other risk nodes in the current time slice; recursively calculate until the starting time slice is reached to obtain the posterior probability distribution of the risk nodes in all time slices.

[0023] This application adopts an interface tree algorithm to calculate the posterior probability distribution of other risk nodes based on the status of the evidence node. Different from the static method of assessing risk once, this method constructs a chain graph of a dynamic Bayesian network and uses the modified conditional probability distribution parameters to recursively calculate the posterior probability distribution of risk nodes in each time slice starting from the terminal time slice. It fully considers the causal relationship and transmission effect of risk in the time dimension, and obtains more accurate and comprehensive risk prediction results.

[0024] In the process of calculating the posterior probability of risk nodes, the status information of evidence nodes and the corrected conditional probability distribution parameters are fully utilized. Compared with the existing methods, when executing the interface tree algorithm, this method not only considers the status of evidence nodes in the current time slice, but also uses the posterior probability of risk nodes at the previous moment as observation evidence. Combined with the status of evidence nodes and conditional probability distribution parameters of the current time slice, the posterior probability of other risk nodes in the current time slice is recursively calculated, which fully utilizes the temporal dependency of the dynamic Bayesian network and improves the accuracy and reliability of risk prediction.

[0025] Another aspect of the present application also provides a mechanical and electrical engineering construction information management system for executing a mechanical and electrical engineering construction information management method of the present application.

[0026] Compared with the prior art, the advantages of this application are:

[0027] The traditional Bayesian network model has defects such as static description and independent and identically distributed assumptions in dealing with mechanical and electrical engineering construction management. It is difficult to characterize the dynamic evolution characteristics and temporal dependencies of risk factors in the construction process, resulting in low accuracy of risk assessment and prediction. In this application, the construction process is divided into multiple stages, and a Bayesian network model of a time slice is constructed for each stage. This breaks through the limitation that the traditional Bayesian network can only describe the state of the system at a single moment. By discretizing the construction process in the time dimension to form a series of time slices, a static Bayesian network is constructed in each time slice, realizing a dynamic description of the construction process. In each time slice, a Bayesian network is constructed according to the risk level of construction personnel behavior and material properties, comprehensively considering the two major risk sources of unsafe behavior of people and unsafe state of things. By dividing the behavioral risk into levels and marking the material properties, a Bayesian network containing behavioral risk nodes and material attribute nodes is constructed, realizing a comprehensive characterization of risk factors in a single time slice.

[0028] By adding time-dependent edges between the Bayesian network models of adjacent time slices, a dynamic Bayesian network is obtained. This overcomes the assumption of independent and identical distribution between time slices in the traditional Bayesian network. By introducing time-dependent edges that represent the impact of the construction risk status at the previous moment on the current moment, the causal relationship and transmission law of risks in the time dimension are characterized, thus realizing the modeling of the dynamic evolution of risks.

[0029] Based on historical data, the conditional probability distribution parameters of the dynamic Bayesian network are learned to obtain a complete dynamic Bayesian network model. Using the historical data accumulated during engineering construction, the conditional probability distribution between nodes in the network is estimated through a parameter learning algorithm. This breaks through the subjectivity of traditional Bayesian networks that rely on expert experience to set parameters, making the model closer to reality and with better generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present application will be further described in the form of exemplary embodiments, which will be described in detail by the accompanying drawings. These embodiments are not restrictive, and in these embodiments, the same number represents the same structure, wherein:

[0031] Figure 1 is an exemplary flow chart of a method for managing mechanical and electrical engineering construction information according to some embodiments of the present application;

[0032] Figure 2 is an exemplary flow chart for dividing construction personnel behaviors and material types according to some embodiments of the present application;

[0033] Figure 3 This is an exemplary flowchart for constructing a dynamic Bayesian network model according to some embodiments of the present application. DETAILED DESCRIPTION

[0034] The method and system provided in the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0035] like Figure 1 As shown, a set of images of mechanical and electrical engineering construction sites is collected; feature recognition is performed on the image set, including recognition of construction personnel behavior and materials, to obtain recognition results; based on the recognition results, the construction personnel behavior risk levels are divided and material attributes are annotated; the construction process is divided into multiple stages, and a Bayesian network model of a time slice is constructed for each stage. A temporal dependency relationship is added between the Bayesian network models of adjacent time slices, and a dynamic Bayesian network risk evolution model for different stages of the construction process is established; among them, the construction personnel behavior risk level and material attributes are used as Bayesian network nodes, and the conditional probability distribution between nodes is set. The risk probability distribution under each factor combination is calculated through the Bayesian network inference algorithm; the dynamic Bayesian network risk evolution model is used to predict the risk distribution heat map of each construction stage in the future.

[0036] S1, deploy multiple high-definition digital cameras at the mechanical and electrical engineering construction site to collect real-time images of the construction process from all directions and angles. The location and number of cameras are reasonably set according to the scale, layout and key risk points of the construction site to ensure that the main areas and links of the construction can be covered. The collected images are automatically stored in the local or cloud image database at a certain frame rate (such as 10 frames per second) to form a set of construction site images. At the same time, the metadata information such as the acquisition time and location of each image is recorded to provide a spatiotemporal background for subsequent image recognition and risk analysis.

[0037] like Figure 2 As shown, S2 performs feature recognition on the image set, including recognizing the construction worker behavior and materials, and obtains the recognition results, including: S21, selects a pre-trained convolutional neural network model (such as VGG, ResNet, etc.), and inputs the image into the network for forward propagation calculation. The output of the middle layer of the network (such as the pooling layer before the fully connected layer) is extracted as the deep feature of the image, which represents the high-level semantic information of the image. The deep feature is represented in the form of a high-dimensional vector and has strong discrimination and generalization capabilities.

[0038] S22, perform traditional feature engineering on the image to extract manually designed low-level visual features. Use algorithms such as local binary pattern (LBP) and histogram of oriented gradients (HOG) to extract the texture features of the image and characterize the texture pattern and structural information of the local area of ​​the image. Use operators such as Canny and Sobel to calculate the edge features of the image and highlight the outline and boundary information of the object. Use methods such as color histogram and color moment to extract the color features of the image to reflect the color distribution and main color information of the image.

[0039] S23, the depth features extracted in step S21 and the texture, edge, and color features extracted in step S22 are fused to form a more comprehensive and robust feature representation. Different types of feature vectors are connected or linearly combined using strategies such as feature concatenation and feature weighting to generate a high-dimensional fused feature vector. The fused feature vector takes into account both high-level semantic information and low-level visual information, and can better characterize the content and attributes of the image.

[0040] S24, select a convolutional neural network model pre-trained on a large-scale image dataset (such as ImageNet) as the basic architecture for image recognition, and fine-tune and professionally customize it for construction scenarios. Input the fused feature vector generated in step S23 into the network, and obtain the category probability distribution of the output layer through forward propagation calculation. According to the level of category probability, determine whether the image contains construction personnel, as well as the specific behavior type of the construction personnel (such as normal operation, illegal operation, lack of protection, etc.). At the same time, identify the main material types appearing in the image (such as steel bars, concrete, scaffolding, etc.). Output the image recognition results and store them in association with the metadata information of the image to form a structured recognition record.

[0041] S3, after completing the feature recognition of the construction site image in step S2, the recognition results including the construction worker behavior type and material type are obtained. Based on these recognition information, the behavior risk level is further divided and the material attributes are annotated, as follows: a risk level classification rule is set, and the behavior risk is divided into three levels: low risk, medium risk and high risk according to the construction worker's behavior type.

[0042] Low-risk behavior: Normal operation behavior, such as wearing a safety helmet in a standardized manner, using tools and equipment correctly, and following the operating procedures, indicates that the construction workers' behavior is safe and compliant, and the risk level is low. Medium-risk behavior: Not wearing some safety equipment, such as not wearing a safety rope, safety shoes, etc., or not operating in a standardized manner, such as not fully following the operating steps, not using protective equipment correctly, etc., indicates that the construction workers' behavior has certain safety hazards and the risk level is medium. High-risk behavior: Serious illegal operations, such as not using any safety protection equipment when working at high altitudes, smoking in dangerous areas, working under the influence of alcohol, etc., indicate that the construction workers' behavior is very likely to cause safety accidents and the risk level is high. According to the above classification rules, the risk level of each identified construction worker behavior is determined, and a behavior risk label (such as "low risk", "medium risk", "high risk") is generated, and stored in association with the original image and recognition record.

[0043] Attribute annotation is performed on the identified material types, mainly including the type and quantity information of the materials. Material type: Based on the recognition results, the main material types appearing in the image are determined, such as steel bars, concrete, wood boards, sandstone, etc., and material type labels (such as "steel bars", "concrete", etc.) are generated. Material quantity: Combining image segmentation and target detection technology, different instances of the same material are counted, the quantity of materials is estimated, and quantity labels are generated (such as "steel bars: 20 pieces", "concrete: 5 cubic meters", etc.). The material type and quantity labels are associated with the original image and recognition records and stored to form complete material attribute annotation information.

[0044] like Figure 3As shown, S4, the construction process is divided into multiple stages, a Bayesian network model of a time slice is constructed for each stage, a temporal dependency is added between the Bayesian network models of adjacent time slices, and a dynamic Bayesian network risk evolution model of different stages of the construction process is established, including: S41, according to the construction plan and progress, the entire construction process is divided into multiple continuous time slices, each time slice represents a construction stage. Specifically, it can be divided according to the work breakdown structure (WBS) and critical path method (CPM) of the construction project. WBS decomposes the construction project into multiple manageable work packages, and CPM determines the logical relationship and time constraints between each work package. On this basis, the construction process is divided into several construction stages with continuous time and relatively independent content, and each stage corresponds to a time slice. The granularity of the time slice can be adjusted according to the needs of construction management, such as dividing by day, week, month, etc.

[0045] S42, in each time slice, construct a Bayesian network containing construction personnel behavior risk nodes and material attribute nodes. According to the risk level classification rules in step S3, the construction personnel behavior risk nodes are divided into three states: low risk, medium risk and high risk. The low-risk state indicates that the construction personnel's behavior complies with safety regulations and the risk level is low; the medium-risk state indicates that the construction personnel's behavior has certain unsafe factors and the risk level is medium; the high-risk state indicates that the construction personnel's behavior seriously violates safety regulations and the risk level is high. In each time slice, the state of the behavior risk node is determined based on the construction personnel behavior data and risk level classification rules of the corresponding stage.

[0046] According to the material management situation at the construction site, set material attribute nodes to represent the attribute information of different materials. The status of the material attribute node can be determined according to the material type and quantity annotation results. For example, for steel bar materials, two attribute nodes "Specification" and "Quantity" can be set. The status of the specification node can be "qualified" or "unqualified", and the status of the quantity node can be "sufficient" or "insufficient". According to the data of material on-site inspection, inventory count, etc., the specific status of the material attribute node in each time slice is determined.

[0047] Between the behavior risk node and the material attribute node, a directed edge is added to indicate the causal dependency relationship between them. The direction of the edge is determined based on domain knowledge and data analysis results, usually from the influencing factor to the affected object. For example, if the unqualified specifications of the steel bar material may lead to an increase in the behavior risk of the construction workers, the direction is from the "steel bar specification" node to the "behavioral risk" node. The causal dependency relationship can be identified and verified through methods such as expert experience, accident case analysis, and statistical data mining.

[0048] The behavioral risk nodes, material attribute nodes, and the causal dependency edges between them are combined to form a complete Bayesian network structure. The Bayesian network uses a directed acyclic graph (DAG) to represent the dependency between nodes. The nodes represent random variables, and the edges represent conditional dependencies. An independent Bayesian network is constructed in each time slice to reflect the causal relationship between the construction personnel's behavioral risks and material attributes in that stage.

[0049] A conditional probability distribution (CPD) is assigned to each node in the Bayesian network, which represents the probability distribution of the node under different state combinations of its parent node. The conditional probability distribution can be preliminarily determined by expert knowledge estimation, historical data statistics and other methods, and can be subsequently optimized by parameter learning algorithms (as described in step S44). For example, for behavioral risk nodes, the following conditional probability distribution can be set: when the steel bar specifications are qualified and the quantity is sufficient, the low risk probability is 0.8, the medium risk probability is 0.15, and the high risk probability is 0.05; when the steel bar specifications are unqualified or the quantity is insufficient, the low risk probability is 0.1, the medium risk probability is 0.3, and the high risk probability is 0.6.

[0050] For each time slice, the Bayesian network of the corresponding stage is constructed repeatedly. The Bayesian network structure and parameters in different time slices can be adjusted appropriately according to the actual situation to reflect the dynamic changes in the construction process. A Bayesian network containing construction personnel behavior risk nodes and material attribute nodes is constructed in each time slice. The network links behavioral risks with material attributes through causal dependencies, providing a basis for a probabilistic graphical model for analyzing and reasoning about risk evolution in the construction process.

[0051] S43, add time-dependent edges between the Bayesian networks of adjacent time slices to form a dynamic Bayesian network structure. Analyze the causal relationship and influence mechanism of various risk factors and material attribute factors in the time dimension during the construction process, and determine which factors have dependencies between the previous and next moments. Consider the dynamic characteristics and persistence of the factors, and evaluate the impact of the state at the previous moment on the state at the current moment. Based on expert knowledge and historical data, summarize significant time-dependent relationships, such as "the risk status of personnel behavior in the previous stage will affect the risk status of personnel behavior in the current stage".

[0052] In a dynamic Bayesian network, each time slice represents a construction stage, and there is a temporal dependency relationship between adjacent time slices. For nodes with temporal dependencies, directed edges are added between the instances of the previous and next moments, with the arrow pointing from the previous moment to the next moment, indicating that the state of the previous moment will affect the state of the next moment. The direction of the temporal dependency edge reflects the temporal order of the causal relationship. The node state of the previous moment is the influencing factor, and the node state of the next moment is the affected object. Each node can have multiple temporal dependency edges, which are connected to different influencing factor nodes of the previous moment, indicating that the current state is affected by the combined influence of multiple previous states.

[0053] The temporal dependency edge represents the influence relationship between the states at the previous and next moments, and the strength of this influence needs to be quantified. The temporal dependency strength is characterized by conditional probability distribution parameters, that is, the probability of the node being in various states at the current moment given the state of the node at the previous moment. The conditional probability distribution parameters can be initially set through expert experience, and then learned and optimized using historical data (as described in step S44). A higher conditional probability value indicates that the state at the previous moment has a stronger influence on the current moment, while a lower conditional probability value indicates a weaker influence.

[0054] By adding time-dependent edges between the Bayesian networks of adjacent time slices, a chain-like dynamic Bayesian network structure is formed. The dynamic Bayesian network structure consists of multiple time slices, each of which contains multiple nodes, representing the risk factors and material attribute factors in the time slice. Time slices are connected by time-dependent edges, which reflects the dynamic evolution of risks, that is, the state of the previous stage will affect the state of the current stage.

[0055] S44, based on historical data, learn the conditional probability distribution parameters of the dynamic Bayesian network to obtain a dynamic Bayesian network model; including: organizing and labeling historical construction data according to time slices. For each time slice, extract the behavior risk state (low, medium, high risk) and material attribute state (type, quantity) therein to form a time slice-state data set. Statistically calculate the state distribution of the behavior risk node and the material attribute node in each time slice. For example, in a certain time slice, low-risk behavior accounts for 80%, medium risk 15%, and high risk 5%; the number of steel bars is 50 tons, the number of concrete is 200 cubic meters, and so on. These statistical results are used as the empirical probability of the node state. For the time-dependent edges between adjacent time slices, the joint distribution of the node state at the previous moment and the node state at the current moment is calculated. For example, the probability of low-risk behavior from the previous moment to the current moment is 90% to maintain low risk, 8% to medium risk, and 2% to high risk; the law of change of the number of steel bars before and after the moment, etc. These statistical results are used as the empirical probability of state transition.

[0056] According to the empirical probability of node state distribution and state transition distribution, the initial state transition probability between nodes in the dynamic Bayesian network is estimated as the initial value of the conditional probability distribution parameter. Commonly used parameter estimation methods include maximum likelihood estimation and Bayesian estimation.

[0057] Set Node The state at time t is , its parent node The state at time t-1 is . State transition probability The maximum likelihood estimate of is: ,in, Indicates that a node is observed in the data set At time t, the state is And the state of its parent node at time t-1 is The number of times; Indicates that the parent node state is observed to be The total number of times.

[0058] Use the EM algorithm to iteratively optimize the initial parameters to improve the accuracy of parameter estimation. At the same time, introduce a random restart strategy to avoid parameter estimation from falling into local optimality through multiple iterations. Set the number of random restarts N, such as N=10. For each random restart: randomly perturb the initial state transition probability to generate the initial parameter value of the EM algorithm The perturbation can be achieved by superimposing a small Gaussian noise on the original estimate. Step E: Using the current parameter estimate , calculate the expected log-likelihood function of the observed data . Assume that the log-likelihood function of the complete data (including observed data and latent variables) is , then the expected log-likelihood function is: , the expected log-likelihood function is expanded as: , where obs represents observed data (state distribution and transition), and z represents latent variables (unobserved states); Indicates that the current parameters Under this condition, the conditional probability of the latent variable z given the observed data obs;

[0059] M step: maximize the expected log-likelihood function to obtain the updated parameter estimates: , repeat the EM steps until the parameters converge or the maximum number of iterations is reached. After all random restarts are completed, the result with the largest expected log-likelihood function value is selected as the optimized parameter estimate obtained by the EM algorithm .

[0060] The initial state transition probability is used to modify the parameters after EM optimization to obtain the final conditional probability distribution parameter estimate. As a prior distribution Parameters. Dirichlet distribution is usually chosen as the prior, and its probability density function is: ,in, and are the parameter vectors of the Dirichlet distribution, K is the dimension of the parameters, Represents the Gamma function. The parameter estimate after EM optimization As the likelihood of observed data .

[0061] According to Bayes' theorem, calculate the posterior distribution of the parameters: , ,in, Represents the number of occurrences of the i-th parameter in the observed data.

[0062] Extract point estimates from the posterior distribution as final parameter estimates , commonly used point estimates include the posterior mean and the posterior mode.

[0063] Posterior mean: ; Posterior mode: . Using the final conditional probability distribution parameters , construct a dynamic Bayesian network model. The node relationship and temporal dependency of the model have been determined in steps S42 and S43. Now fill in the learned numerical parameters for the conditional probability distribution of each node to form a complete dynamic Bayesian network model.

[0064] S5, using the dynamic Bayesian network risk evolution model to predict the risk distribution heat map of each stage of construction in the future, including: S51, according to the construction plan and schedule, determine the future time period that needs to be predicted, and divide the future time period into continuous time slices, each time slice represents a construction stage. Analyze the construction plan and schedule to determine the future time period that needs to be predicted, such as the next 1 month, 3 months or the entire construction cycle. According to the work breakdown structure (WBS) and critical path method (CPM) of the construction project, divide the future time period into several continuous time slices. Each time slice represents a construction stage, and the length of the time slice can be days, weeks, months, etc., which is set according to the granularity of construction management and the needs of prediction.

[0065] S52, input the construction personnel behavior type and material type in each future time slice into the dynamic Bayesian network model as evidence nodes. For each future time slice, according to the construction plan and safety management requirements, determine the personnel behavior type (such as normal operation, illegal operation, etc.) and material type (such as steel bar, concrete, etc.) in this stage. In the dynamic Bayesian network model, these personnel behavior types and material types are input as evidence nodes. Evidence nodes represent known or observed information, and their status is considered to be certain when predicting risks.

[0066] S53, according to the state of the evidence node, using the conditional probability distribution parameters, calculate the posterior probability distribution of other risk nodes; including: taking the state of the evidence node as a known condition, and constructing a chain graph of the dynamic Bayesian network according to the modified conditional probability distribution parameters. According to the time slice division of the construction process, a chain dynamic Bayesian network graph is constructed. Each time slice corresponds to a subgraph, which represents the static Bayesian network in the time slice. Each time slice contains multiple nodes, which represent construction risk factors (such as personnel behavior risk, equipment risk, environmental risk, etc.) and material attribute factors (such as material type, quantity, quality, etc.). In the same time slice, a directed edge is added between the risk node and the material attribute node according to the causal relationship to represent the dependency relationship between them. Between adjacent time slices, for each node, a time-dependent edge is added from the previous time slice to the next time slice to represent the influence relationship between the node before and after. Using the modified conditional probability distribution parameters learned in step S44, the conditional probability table of each node is assigned to quantify the causal strength and time-dependent strength between nodes.

[0067] Choose an accurate algorithm suitable for dynamic Bayesian network reasoning, such as the Junction Tree Algorithm. This algorithm achieves efficient probability propagation and updating by converting the network into a tree structure. Instantiate the state of the evidence node in the terminal time slice as a known condition. For example, if the risk of personnel behavior in a certain time slice is observed to be "high risk", the state of the corresponding node is set to "high risk". According to the instantiated evidence node state, use the conditional probability distribution parameters to calculate the posterior probability distribution of other risk nodes. The specific steps are as follows: For each non-evidence node, calculate its posterior probability under different states. The posterior probability is calculated according to Bayes' theorem: , where X is a non-evidence node and E is an evidence node. is the prior probability of X, is the conditional probability of E given X, is the marginal probability of E. Calculated using the conditional probability distribution parameters , prior probability Obtained according to the posterior probability or initial probability of the previous time slice, For each possible state combination, the posterior probability is calculated to form the posterior probability distribution of the non-evidence node. The posterior probability distribution of all non-evidence risk nodes in the terminal time slice is obtained, which describes the probability of different risk factors being in various states under known evidence conditions.

[0068] Starting from the penultimate time slice of the chain graph, the posterior probability of the risk node in the penultimate time slice (i.e., the terminal time slice) is used as the observation evidence. For each risk node in the penultimate time slice, the posterior probability of the corresponding node at the next moment is used as the observation value and added to the evidence node set of the current time slice. Combining the original evidence node status in the current time slice and the newly added observation evidence, the reasoning calculation is repeated to obtain the posterior probability distribution of other risk nodes in the current time slice. Move forward recursively and calculate the posterior probability distribution of the risk node in each time slice in turn until the starting time slice is reached.

[0069] Recursive calculations are performed to gradually transfer the risk prediction results from the back to the front, and the risk prediction information of the latter moment is used to update the risk assessment of the previous moment. This recursive transfer process makes full use of the temporal dependency of the dynamic Bayesian network, so that the risk prediction can be continuously adjusted and optimized according to new observational evidence. Finally, the posterior probability distribution of risk nodes in all time slices from the starting time slice to the end time slice is obtained, which comprehensively describes the risk status and evolution trend of each future construction stage.

[0070] S54, for each future time slice, summarize the posterior probability distribution of all risk nodes, and calculate the comprehensive probability values ​​of the three risk levels of low risk, medium risk and high risk in the time slice; arrange the comprehensive probability values ​​of the three risk levels of each time slice in chronological order to form a probability sequence of risk levels changing over time; smooth the probability sequence of risk levels to eliminate the interference of short-term fluctuations on long-term trends; map the smoothed risk level probability sequence to the time axis to generate a time series probability distribution diagram of construction risk levels; mark the risk threshold points in the time series probability distribution diagram, and the risk threshold points indicate the time points when the risk level changes significantly.

[0071] S55, visualize the time series probability distribution map of construction risk levels as a risk distribution heat map; select a suitable color scheme to map different risk levels to different colors. For example, use green, yellow, and red to represent low risk, medium risk, and high risk, respectively. Convert the risk level probability of each time slice in the time series probability distribution map into corresponding colors to generate a two-dimensional risk distribution heat map. The horizontal axis of the heat map represents time, the vertical axis represents the risk level, and the size of the area of ​​different colors represents the probability of the corresponding risk level. Add necessary legends, titles, and descriptions to the heat map to clearly show the distribution of future construction risks. The heat map intuitively and vividly presents the evolution trend of construction risks in the time dimension, which is convenient for managers to quickly identify high-risk intervals and key risk factors, and take targeted risk control measures in a timely manner.

[0072] S6. Carry out construction management based on the risk distribution heat map. According to the risk distribution heat map, focus on areas with darker colors and higher risk levels, which represent high-risk parts of the construction process. Pay attention to identifying time periods where risks are concentrated or last for a long time. These time periods usually require key monitoring and control. Combined with the construction plan and progress, determine the reasons for the emergence of high-risk areas and time periods, and whether they are related to key processes, major changes and other factors. Use the dynamic Bayesian network model to track the causal relationship and temporal dependency between risk nodes, and reveal the mechanism of risk transmission and evolution.

Claims

1. A method for managing mechanical and electrical engineering construction information, characterized in that: include: S1, collects the image set of the mechanical and electrical engineering construction site; S2, performing feature recognition on the image set, including identifying the construction worker behavior and materials, and obtaining recognition results; S3, based on the recognition results, classify the risk level of construction personnel behavior and mark material attributes; S4, divide the construction process into multiple stages, build a Bayesian network model of time slices in each stage, add temporal dependencies between the Bayesian network models of adjacent time slices, and establish a dynamic Bayesian network risk evolution model for different stages of the construction process; in which, the risk level of construction personnel behavior and material properties are used as Bayesian network nodes, and the conditional probability distribution between nodes is set. The risk probability distribution under each factor combination is calculated through the Bayesian network inference algorithm; S5, using the dynamic Bayesian network risk evolution model to predict the risk distribution heat map of each construction stage in the future; S6, construction management based on risk distribution heat map; Among them, S5 uses the dynamic Bayesian network risk evolution model to predict the risk distribution heat map of each stage of construction in the future, including: S51, determining a future time period to be predicted according to the construction plan and progress, and dividing the future time period into continuous time slices, each time slice representing a construction stage; S52, inputting the construction worker behavior type and material type in each future time slice into the dynamic Bayesian network model as evidence nodes; S53, according to the state of the evidence node, using the conditional probability distribution parameters, calculate the posterior probability distribution of other risk nodes; S54, generating a time series probability distribution diagram of the construction risk level according to the posterior probability distribution of the risk nodes in each future time slice; S55, visualize the time series probability distribution map of the construction risk level as a risk distribution heat map.

2. The method for managing electromechanical engineering construction information according to claim 1, characterized in that: S2, perform feature recognition on the image set, including identifying the construction worker behavior and materials, and obtain recognition results, including: S21, using a convolutional neural network to extract deep features of the image as the first feature; S22, extracting texture features, edge features and color features of the image as second features; S23, fusing the first feature and the second feature to obtain a fused feature vector; S24, taking the fused feature vector as input, using the pre-trained convolutional neural network to perform forward propagation calculation, and judging the construction worker behavior type and material type in the image according to the category probability of the output layer.

3. The method for managing electromechanical engineering construction information according to claim 2, characterized in that: S3, based on the identification results, divide the construction personnel behavior risk level and mark the material attributes, including: The behavior risks are divided into three levels: low risk, medium risk and high risk according to the behavior types of construction workers. The behavior risks include not wearing safety equipment and improper operation. Mark the type and quantity of materials according to the material type.

4. The method for managing electromechanical engineering construction information according to claim 3, characterized in that: S4, establish a dynamic Bayesian network risk evolution model for different stages of the construction process, including: S41, according to the construction plan and progress, the entire construction process is divided into a plurality of continuous time slices; each time slice represents a construction stage; S42, in each time slice, the construction personnel behavior risk is divided into three levels: low risk, medium risk and high risk, which are used as the behavior risk nodes of the Bayesian network; the type and quantity of materials are marked according to the material type, which are used as the material attribute nodes of the Bayesian network; the Bayesian network is constructed based on the behavior risk nodes and material attribute nodes; S43, adding time-dependent edges between the Bayesian networks of adjacent time slices to obtain a dynamic Bayesian network; wherein the time-dependent edges represent the impact of the construction risk status at the previous moment on the current moment; S44, based on historical data, learning the conditional probability distribution parameters of the dynamic Bayesian network to obtain a dynamic Bayesian network model.

5. The method for managing electromechanical engineering construction information according to claim 4, characterized in that: S44, based on historical data, learning the conditional probability distribution parameters of the dynamic Bayesian network to obtain a dynamic Bayesian network model, including: Organize and annotate historical data according to time slices, count the state distribution of behavior risk nodes and material attribute nodes in historical data within each time slice, and obtain the empirical probability of node states; For the time-dependent edges between adjacent time slices, the joint distribution of the node state at the previous moment and the node state at the current moment is statistically analyzed to obtain the empirical probability of state transition; According to the state distribution and joint distribution, the initial state transition probability between nodes in the dynamic Bayesian network is estimated as the initial value of the conditional probability distribution parameter; The expectation maximization EM algorithm and random restart strategy are used to iteratively optimize the conditional probability distribution parameters of the dynamic Bayesian network. The conditional probability distribution parameters after iterative optimization are corrected using the initial state transition probability to obtain the corrected conditional probability distribution parameters; The dynamic Bayesian network model is constructed using the final conditional probability distribution parameters.

6. The method for managing mechanical and electrical engineering construction information according to claim 5, characterized in that: The expectation maximization EM algorithm and random restart strategy are used to iteratively optimize the conditional probability distribution parameters of the dynamic Bayesian network, including: Set the number of random restarts N. Random restarts mean reinitializing parameters during the iterative optimization process. For each random restart, the initial state transition probability is randomly perturbed to generate the initial parameter values ​​of the EM algorithm; In step E, the expected log-likelihood function of the observed data is calculated using the current conditional probability distribution parameters; the observed data includes state distribution and state transition; In the M step, the expected log-likelihood function is maximized and the estimate of the conditional probability distribution parameters is updated; Repeat steps E and M until the conditional probability distribution parameters converge or the maximum number of iterations is reached; After completing N random restarts, the result with the largest expected log-likelihood function value is selected as the conditional probability distribution parameter after iterative optimization.

7. The method for managing electromechanical engineering construction information according to claim 6, characterized in that: The initial state transition probability is used to correct the iteratively optimized conditional probability distribution parameters to obtain the corrected conditional probability distribution parameters, including: The initial state transition probability is used as the parameter of the Dirichlet distribution to construct the prior distribution of the conditional probability distribution parameters; The conditional probability distribution parameters after iterative optimization are used as observation data; According to Bayes' theorem, the posterior distribution of the conditional probability distribution parameters is calculated based on the prior distribution and the observed data; The mean or mode of the posterior distribution is used as the modified conditional probability distribution parameter.

8. The method for managing electromechanical engineering construction information according to claim 1, characterized in that: S53, according to the state of the evidence node, using the conditional probability distribution parameters, calculate the posterior probability distribution of other risk nodes, including: Taking the state of the evidence node as a known condition, a chain graph of a dynamic Bayesian network is constructed according to the modified conditional probability distribution parameters; the chain graph is composed of multiple time slices, and the nodes in each time slice represent construction risk factors and material attribute factors; From the end time slice of the chain graph, the posterior probability of other risk nodes is calculated according to the status of the evidence node and the conditional probability distribution parameters; For the previous time slice, the posterior probability of the risk node of the next time slice is used as the observation evidence, and the posterior probability of other risk nodes in the current time slice is calculated by combining the evidence node status and conditional probability distribution parameters of the current time slice; The calculation is performed recursively until the starting time slice is reached, and the posterior probability distribution of risk nodes in all time slices is obtained.

9. A mechanical and electrical engineering construction information management system, characterized in that: include: At least one processing unit; used to execute instructions to implement the electromechanical engineering construction information management method described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions, which implement the method according to any one of claims 1 to 8 when executed by a processor.

Images