A method, system, and medium for managing construction data during non-stop flight operations in the flight zone.

Abstract

This invention discloses a method, system, and medium for managing construction data during non-stop operation in the airfield, relating to the field of airport construction management. This invention introduces a dynamic Bayesian network model into airfield construction risk management. By dividing construction time slices and establishing temporal dependencies between risk factor nodes and target risk nodes, it achieves accurate modeling of the dynamic evolution of risks throughout the entire construction process. Utilizing diagnostic reasoning technology, it predicts the probability of occurrence of different risk factors at each construction stage and generates an intuitive risk distribution heatmap. This transforms risk management from traditional manual inspections, static assessments, and post-event responses into a new model of intelligent, dynamic prediction, and precise intervention, significantly improving the accuracy of risk identification, the timeliness of early warnings, and the effectiveness of management measures, fundamentally ensuring the safety and efficiency of airport operations during non-stop construction.

Description

Technical Field

[0001] This invention relates to the field of airport construction management, and more specifically, to a method, system, and medium for managing construction data during non-stop flight operations in the flight zone. Background Technology

[0002] Most construction work in the flight area of ​​civil aviation transport airports is carried out without airport closure. This involves strict time constraints, a large impact area, complex procedures, and is highly prone to safety accidents. Construction without airport closure refers to engineering work being conducted within the flight area while the transport airport remains open or partially closed for certain periods, and aircraft are received and released according to flight schedules.

[0003] Currently, risk management for non-stop construction in the flight zone mainly relies on regular, fixed-point manual inspections and recording by safety officers. This method has significant drawbacks: manual inspections are arduous, leading to low regulatory efficiency and delayed data feedback, making it difficult to promptly identify and address dynamic risks; manual recording is highly subjective, resulting in inaccurate descriptions of hazard locations, leading to a large workload and insufficient focus in subsequent reviews.

[0004] At the level of risk assessment methodologies, the current construction field often employs static models such as fuzzy comprehensive evaluation, grey relational analysis, and neural networks. While these methods are helpful for risk analysis to some extent, they generally fail to characterize the intrinsic relationships between risk factors, let alone reflect the dynamic characteristics of risk evolution as construction progresses. Summary of the Invention

[0005] The purpose of this invention is to improve the management efficiency of construction without stopping navigation and to achieve dynamic risk assessment.

[0006] To achieve the above objectives, the present invention provides a method for managing construction data during non-stop flight operations in an airfield, the method comprising the following steps:

[0007] Acquire on-site images of the construction area, identify unsafe behaviors of construction workers based on the on-site images, and obtain the results of unsafe behavior identification;

[0008] Personnel intrusion monitoring was conducted in key activity areas adjacent to the construction area to obtain intrusion monitoring results;

[0009] The final warning result is obtained based on the results of unsafe behavior identification and intrusion detection, and the threat level is classified based on the final warning result;

[0010] The construction process is divided into several consecutive time slices, where any one time slice represents a construction stage.

[0011] Define target risk nodes, take the threat level corresponding to each time slice as the risk factor node of that time slice, establish the temporal dependency relationship of the same risk factor nodes between consecutive time slices, and construct a dynamic Bayesian network risk evolution model based on the temporal dependency relationship. The target risk nodes are used to characterize the overall safety status of construction.

[0012] Diagnostic reasoning is performed based on the dynamic Bayesian network risk evolution model, and the posterior probability distribution of the risk factor nodes in each time slice is output.

[0013] A risk distribution heatmap is generated based on the posterior probability distribution of the risk factor nodes within each time slice.

[0014] Construction management is based on risk distribution heatmaps.

[0015] This solution first identifies unsafe behaviors of construction workers and monitors intrusion into key activity areas adjacent to the construction zone to obtain early warning results, achieving automated and real-time perception of risk sources. Next, the construction process is divided into several consecutive time slices, and a dynamic Bayesian network risk evolution model is constructed. The model's principle is to use the threat level within each time slice as a risk factor node and the overall construction safety state as the target risk node. It characterizes the evolution of risk factors with each construction stage by establishing temporal dependencies between identical risk factor nodes in adjacent time slices. Diagnostic reasoning based on the dynamic Bayesian network risk evolution model is mathematically an application of Bayes' theorem. By using the high-risk state of the target risk node as evidence, the posterior probability distribution of each risk factor node is calculated in reverse. That is, in a simulated accident analysis, the contribution of each risk factor to the overall high risk is quantified; the higher the posterior probability, the greater the likelihood that the risk factor corresponding to that node will cause a safety accident. Ultimately, a risk distribution heatmap is generated based on the posterior probability distribution of each risk factor node. This heatmap is a spatial and visual representation of future risk probabilities. Therefore, construction management based on the risk distribution heatmap allows for the early identification and rectification of high-risk factors, enabling preventative measures. This solution, through modeling and reverse diagnosis of the risk evolution process, achieves a shift from passive response to proactive prediction and precise source tracing, solving the fundamental problems of low efficiency in manual inspections and the inability of static risk assessment methods to reflect the dynamic evolution of risks in non-stop construction risk management in flight zones.

[0016] Furthermore, the node parameters of the dynamic Bayesian network risk evolution model are obtained in the following way:

[0017] The prior probability of the root node in the risk factor nodes is determined by an expert survey method based on trapezoidal fuzzy decision-making.

[0018] For non-root nodes, the initial conditional probability is determined based on the Leaky Noisy-or Gate extended model;

[0019] Based on the expectation-maximization algorithm, the initial conditional probabilities of non-root nodes are iteratively corrected.

[0020] In the dynamic Bayesian network risk evolution model, the node parameters directly determine the reliability of risk prediction. This scheme employs a hybrid strategy combining prior knowledge and data-driven approaches to ensure parameter reliability. First, for risk factor nodes without parent nodes (i.e., root nodes), their prior probabilities are determined using an expert survey method based on trapezoidal fuzzy decision-making. This method transforms the qualitative experience of experts into precise probability values, ensuring that the model is built based on domain knowledge. Then, for non-root nodes (including intermediate nodes and target risk nodes), their conditional probability tables are enormous, making direct evaluation difficult. Therefore, this scheme first determines their initial conditional probabilities based on the Leaky Noisy-or Gate extended model. The principle of this model is to assume that each parent node independently triggers the state of its child nodes with a certain probability, and to introduce leakage probabilities to cover unknown causes, thereby generating reasonable initial values ​​for the entire conditional probability table with a small number of parameters, greatly reducing the workload of experts. However, this is still an estimation based on simplification assumptions and has limitations. Therefore, this invention further employs the expectation-maximization algorithm to iteratively correct the initial conditional probabilities of non-root nodes. The expectation-maximization algorithm can find the maximum likelihood estimate of model parameters under incomplete data. Through iterative E-step (expectation calculation) and M-step (parameter maximization), it optimizes the initial conditional probability using actual sample data, so that the model parameters can reflect both expert priors and the statistical regularity of historical data.

[0021] Furthermore, based on the aforementioned dynamic Bayesian network risk evolution model, diagnostic inference is performed, outputting the posterior probability distribution of the risk factor nodes within each time slice, including:

[0022] The high-risk state of the target risk node is taken as known evidence and input into the dynamic Bayesian network risk evolution model;

[0023] The dynamic Bayesian network risk evolution model performs reverse probabilistic inference to calculate the posterior probability distribution of the risk factor nodes.

[0024] In this model, the high-risk state of the target risk node is used as known evidence input into the model, which treats the overall high security risk as an established fact. Then, the model performs reverse probabilistic reasoning, calculating the posterior probability of the state of each risk factor node in the network given the evidence of the target node. Specifically, the model network uses the inter-node dependencies encoded by a conditional probability table to propagate evidence information from child nodes to parent nodes, updating the probability distribution of all nodes. This process quantifies the probability of each specific unsafe factor (such as not wearing a helmet, intrusion into a critical area, etc.) existing under the assumption of overall high risk. This diagnostic reasoning capability enables the management system not only to know the risk level but also to pinpoint the source of risk, providing a direct and effective quantitative basis for precise control measures.

[0025] Furthermore, based on on-site images of the construction area, unsafe behaviors of construction workers can be identified, including:

[0026] Build a knowledge base of unsafe behaviors;

[0027] Historical images are labeled based on the aforementioned unsafe behavior knowledge base to obtain a training set;

[0028] Construct a support vector machine model, train the support vector machine model based on the training set, and obtain the recognition model;

[0029] Input the on-site images of the construction area into the recognition model, and output the results of unsafe behavior recognition.

[0030] This solution first clarifies the target categories to be identified by constructing a knowledge base of unsafe behaviors. Then, historical images are labeled based on this knowledge base to obtain a training set, providing learning samples for the machine learning model. The principle of the Support Vector Machine (SVM) model is to find an optimal classification hyperplane in a high-dimensional feature space that maximizes the margin between samples of different categories, thereby achieving strong generalization ability. The training process is the process of solving for this optimal hyperplane. Finally, the images of the construction area are classified and judged using the trained model. This method solves the problems of low efficiency and high subjectivity of manual identification, and achieves automatic identification of specific unsafe behaviors through machine learning models, providing reliable basic data input for the entire risk management process.

[0031] Furthermore, the construction of the unsafe behavior knowledge base includes:

[0032] Collect historical construction accident reports for the airfield;

[0033] The accident report text is cleaned and segmented to obtain the text processing results;

[0034] A custom thesaurus is established based on the construction standards of the flight zone. The text processing results are then merged into phrases based on the custom thesaurus to obtain key causal words.

[0035] Human causative factors were extracted from the key causative words using the BERT model.

[0036] The feature value weights of each of the aforementioned anthropogenic causative factors were calculated using the term frequency-inverse document frequency algorithm.

[0037] The human-causing factors are arranged in descending order of their feature value weights, and a predetermined number of the top-ranked human-causing factors are selected to construct the unsafe behavior knowledge base.

[0038] The principle of this solution is to extract core human risk factors from historical data using text mining and information retrieval techniques, ensuring the authority and comprehensiveness of the knowledge base. Specifically, the accident report text undergoes text cleaning and word segmentation, transforming unstructured text into structured data suitable for analysis. A custom thesaurus is established based on flight zone construction standards, and the text processing results are then grouped using this thesaurus, ensuring the accuracy of technical terminology. The BERT model, due to its powerful semantic understanding capabilities, can more accurately identify causal factors such as human error or violations expressed in the text. Finally, the term frequency-inverse document frequency (TNF) algorithm is used to calculate the feature value weights of each human risk factor. The principle of the TNF algorithm is that a word that appears frequently in a specific document but infrequently in the entire document set has high category discrimination and therefore a high weight. By selecting the human risk factors with high weight rankings, it is ensured that the behaviors included in the database are key factors highly relevant to flight zone construction safety.

[0039] Furthermore, intrusion monitoring of key activity areas adjacent to the construction area includes:

[0040] Based on the three-dimensional coordinates of the key points of the temporary boundary water-filled barriers in the key activity area, the planar equation of the temporary boundary is constructed.

[0041] Establish a transformation model between the world coordinate system, camera coordinate system, image coordinate system, and pixel coordinate system;

[0042] The actual position of the intruder in the world coordinate system is determined by the transformation model, and the actual position is compared with the plane equation of the temporary boundary to generate the intrusion monitoring result.

[0043] This solution implements intrusion monitoring for key activity areas adjacent to construction zones, such as running tracks and ski slopes, to prevent intrusion security risks caused by construction activities. The planar equation of the electronic fence defines a virtual security boundary in three-dimensional space. The geometric basis of the transformation model is perspective projection transformation, which establishes a mapping relationship from two-dimensional image pixels to real-world three-dimensional coordinates through intrinsic and extrinsic parameters obtained by camera calibration. The pixel position of a person in the image is used to calculate their ground coordinates in the real world through the coordinate transformation model, and then the spatial relationship (such as distance) between these coordinates and the planar equation of the electronic fence is calculated to determine whether an intrusion has occurred.

[0044] Furthermore, obtaining the final early warning result based on the results of unsafe behavior identification and intrusion monitoring includes:

[0045] Calculate the first warning confidence level corresponding to the unsafe behavior identification result, and calculate the second warning confidence level corresponding to the intrusion detection result;

[0046] The final warning confidence level is calculated using a weighted product fusion strategy based on the first warning confidence level and the second warning confidence level.

[0047] The final warning result is output based on the preset range in which the final warning confidence level falls.

[0048] The mathematical principle of the weighted product fusion strategy is that the weighting part takes into account the importance of the two information sources, while the product part will significantly reduce the credibility of the final result when the confidence of either information source is very low, which can effectively suppress false alarms caused by the distortion of a single monitoring result.

[0049] Furthermore, the formula for calculating the second warning confidence level corresponding to the intrusion detection result is as follows:

[0050] ;

[0051] in, This represents the second warning confidence level. The distance between the actual location of the intruders and the temporary perimeter fence. This refers to the width of the entry threshold zone.

[0052] The present invention also provides a data management system for non-stop construction in flight zones, the system comprising:

[0053] The intelligent identification and early warning unit is used to acquire on-site images of the construction area, identify unsafe behaviors of construction personnel based on the on-site images of the construction area, obtain unsafe behavior identification results, implement personnel intrusion monitoring in key activity areas adjacent to the construction area, obtain intrusion monitoring results, obtain a final early warning result based on the unsafe behavior identification result and the intrusion monitoring result, and classify the threat level based on the final early warning result;

[0054] Dynamic risk assessment units are used to divide the construction process into several consecutive time slices, where any one time slice represents a construction stage.

[0055] The dynamic risk assessment unit is also used to define target risk nodes, take the threat level corresponding to each time slice as the risk factor node of that time slice, establish the temporal dependency relationship of the same risk factor nodes between consecutive time slices, and construct a dynamic Bayesian network risk evolution model based on the temporal dependency relationship. The target risk node is used to characterize the overall safety status of construction.

[0056] The dynamic risk assessment unit is also used to perform diagnostic reasoning based on the dynamic Bayesian network risk evolution model, output the posterior probability distribution of the risk factor nodes in each time slice, and form a risk distribution heatmap based on the posterior probability distribution of the risk factor nodes in each time slice.

[0057] The decision-making unit is used for construction management based on the risk distribution heat map.

[0058] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any one of the above-described methods for managing construction data during non-stop flight operations in a flight zone.

[0059] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0060] This invention introduces a dynamic Bayesian network model into airfield construction risk management. By dividing construction time slices and establishing temporal dependencies between risk factor nodes, it achieves modeling of the dynamic evolution of risks throughout the construction process. Utilizing diagnostic reasoning technology, it predicts the probability of occurrence of different risk factors at each construction stage and generates an intuitive risk distribution heatmap. This transforms risk management from traditional manual inspections, static assessments, and post-event responses into a new model of intelligent, dynamic prediction, and early intervention, improving the accuracy of risk identification and the effectiveness of management measures, and ensuring the safety of airport operations during non-stop construction. Attached Figure Description

[0061] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0062] Figure 1 This is a flowchart illustrating a method for managing construction data in an airfield without stopping operations, as described in this invention.

[0063] Figure 2 This is a schematic diagram of the composition of a non-stop construction data management system for flight zones in this invention. Detailed Implementation

[0064] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0065] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0066] Example 1

[0067] Please refer to Figure 1 Embodiment 1 of the present invention provides a method for managing construction data during non-stop flight operations in an airfield, characterized in that the method includes the following steps:

[0068] Acquire on-site images of the construction area, identify unsafe behaviors of construction workers based on the on-site images, and obtain the results of unsafe behavior identification;

[0069] Personnel intrusion monitoring was conducted in key activity areas adjacent to the construction area to obtain intrusion monitoring results;

[0070] The final warning result is obtained based on the results of unsafe behavior identification and intrusion detection, and the threat level is classified based on the final warning result;

[0071] The construction process is divided into several consecutive time slices, where any one time slice represents a construction stage.

[0072] Define target risk nodes, take the threat level corresponding to each time slice as the risk factor node of that time slice, establish the temporal dependency relationship of the same risk factor nodes between consecutive time slices, and construct a dynamic Bayesian network risk evolution model based on the temporal dependency relationship. The target risk nodes are used to characterize the overall safety status of construction.

[0073] Diagnostic reasoning is performed based on the dynamic Bayesian network risk evolution model, and the posterior probability distribution of the risk factor nodes in each time slice is output.

[0074] A risk distribution heatmap is generated based on the posterior probability distribution of the risk factor nodes within each time slice.

[0075] Construction management is based on risk distribution heatmaps.

[0076] This method is applied to non-stop construction projects in the flight area of ​​civil airports, such as runway maintenance or taxiway expansion projects, aiming to achieve real-time monitoring and dynamic prediction of construction risks through automation. The following details each step in a specific scenario:

[0077] Acquiring on-site images of the construction area is achieved by deploying a high-definition video surveillance system around the area. This includes installing cameras on high-mast lights, temporary fencing, and the exterior of the standby area to cover the entire construction zone. The video surveillance system acquires real-time video streams at a rate of 20 frames per second and breaks the stream down into continuous, independent video frames in RGB format.

[0078] Based on on-site images of construction sites, unsafe behaviors of construction workers are identified. The acquired images undergo preprocessing, including Gaussian filtering for noise reduction and histogram equalization to enhance contrast. Then, a pre-trained recognition model is used for analysis. The model can detect typical unsafe behaviors, such as not wearing a safety helmet, not wearing a reflective vest, using a mobile phone, and not wearing a safety harness while working at heights. The output of the recognition results is the unsafe behavior identification result; for example, if a person is detected not wearing a safety helmet in the image, the label "not wearing a safety helmet" is output.

[0079] Personnel intrusion monitoring is implemented in key activity areas adjacent to the construction area. Electronic fence technology is used to detect whether construction personnel have intruded into critical activity areas, such as runway protection zones or taxiways. The electronic fence constructs a virtual boundary based on the temporary water-filled barriers surrounding the construction area. When personnel intrusion is detected, intrusion monitoring results are generated, including the intrusion location and distance information.

[0080] The final early warning result is obtained based on the results of unsafe behavior identification and intrusion monitoring. The two information are integrated and the threat level is divided according to preset rules.

[0081] The construction process is divided into several consecutive time slices, each representing a construction stage. A time slice is a continuous, equal-length unit of time used to discretize the construction progress; consecutive time slices mean that these time units are connected end-to-end on the timeline, covering the entire construction cycle. For example, a 30-day construction project might be divided into 10 consecutive time slices, each lasting 3 days. The division of time slices is determined based on the characteristics of project progress and risk changes, such as stages like earthwork excavation, foundation construction, and equipment installation.

[0082] Define target risk nodes to characterize the overall safety status of construction, such as "comprehensive safety risk". Use the threat level corresponding to each time slice as the risk factor node for that time slice; for example, the threat level of not wearing a safety helmet is used as a risk factor node. Set the conditional probability distribution between nodes to construct a static Bayesian network. Further establish the temporal dependencies of identical risk factor nodes across consecutive time slices to construct a dynamic Bayesian network risk evolution model (DBN model). The DBN model captures the evolution of risk over time through the conditional probability distribution between nodes; for example, the state of the "not wearing a safety helmet" node in the previous time slice affects the state of the "not wearing a safety helmet" node in the current time slice.

[0083] In practical implementation, considering the actual engineering situation and the aircraft operation support status outside the unclosed area, temporal dependencies between adjacent time slices are established, and temporal dependency edges are added. The causal relationships and impact mechanisms of each identified unsafe behavior during construction are analyzed in the time dimension to determine the dependencies between consecutive time slices. Considering the dynamic characteristics and persistence of factors, the strength of the influence of the previous time slice's state on the current time slice's state is assessed. For nodes with temporal dependencies, directed edges are added between instances at consecutive time slices, with arrows pointing from the previous time slice to the next time slice, indicating that the state at the previous time slice will affect the state at the next time slice. The strength of the influence is characterized by conditional probability distribution parameters.

[0084] When performing diagnostic reasoning based on the DBN model, the target risk node is set to a high-risk state as evidence, and the posterior probability of each risk factor node is calculated through reverse probabilistic reasoning. For example, in time slice T+1, assuming the overall safety risk is high, the posterior probability of the node not wearing a safety helmet is calculated to be 0.6, indicating that this behavior is the main risk source.

[0085] Based on the posterior probability distribution, a risk distribution heatmap is generated. The heatmap uses color coding to display areas with different risk levels. For example, areas with a posterior probability between 0.7 and 1.0 are red, indicating high risk requiring urgent control; those between 0.3 and 0.7 are yellow, indicating medium risk requiring enhanced monitoring; and those between 0 and 0.3 are blue, indicating low risk that can be managed routinely. The heatmap can be overlaid on the construction site plan to visually display risk hotspots.

[0086] Finally, construction management is based on a risk distribution heatmap. For example, for red areas, work is immediately suspended for rectification; for yellow areas, preventative measures are strengthened; and for blue areas, routine management is maintained.

[0087] The node parameters of the dynamic Bayesian network risk evolution model are obtained in the following way:

[0088] The prior probability of the root node in the risk factor nodes is determined by an expert survey method based on trapezoidal fuzzy decision-making.

[0089] For non-root nodes, the initial conditional probability is determined based on the Leaky Noisy-or Gate extended model;

[0090] Based on the expectation-maximization algorithm, the initial conditional probabilities of non-root nodes are iteratively corrected.

[0091] In this context, the root node refers to a risk factor node without a parent node. The trapezoidal fuzzy decision-making expert survey method is a mathematical approach to handle the fuzziness and uncertainty in expert judgment. It involves inviting airport engineering professionals responsible for flight area renovation projects or senior engineers with extensive experience in airport flight area construction. Weights are assigned based on the experts' education, work experience, and professional titles to conduct fuzzy evaluations. Experts score the probability of risk events occurring, and trapezoidal fuzzy numbers are used to handle uncertainty. Finally, the prior probability value is determined by synthesizing the evaluation results from all experts.

[0092] For non-root nodes (such as intermediate nodes or target risk nodes), initial conditional probabilities are determined based on the Leaky Noisy-or Gate extended model. The Leaky Noisy-or Gate model simplifies the construction of the conditional probability table (CPT) and reduces the workload for experts. This model assumes that parent nodes independently influence child nodes and covers unknown factors through leaked probabilities. For example, child nodes in the network... exist and Two parent nodes, For leakage rate, and Let each be a connection probability, then:

[0093] ;

[0094] ;

[0095] compute nodes The connection probability corresponding to their respective parent nodes, i.e.:

[0096] ;

[0097] Ultimately, the nodes can be obtained. The conditional probability is:

[0098] ;

[0099] This involves the risk state classification of leaf nodes and intermediate nodes. The state of intermediate nodes and leaf nodes is determined by the activation probability and a preset risk threshold. After comparison, determine: if ,but (node No risk occurs; if ,but (node Risk occurs. Risk threshold. It is a core indicator for distinguishing the states of leaf nodes and intermediate nodes, and the scenario of construction in the airport flight area without stopping operations is quite complex, with high risk thresholds. The criteria need to be jointly set by experts and airport management agencies based on the actual situation of uninterrupted construction, the operational efficiency and status of surrounding aircraft, support capability assessments, and relevant historical research data to improve the sensitivity of identification.

[0100] Based on the Expectation-Maximization (EM) algorithm, the initial conditional probabilities of non-root nodes are iteratively corrected. The EM algorithm is an iterative optimization algorithm used to estimate the parameters of a probabilistic model containing latent variables. It achieves a local optimum by alternately executing the expectation (E-step) and maximization (M-step) steps and is widely used in machine learning, statistics, and data clustering. Specifically, for... Observational data of one sample middle( Representing the (One sample of observed data), containing unobserved latent data. , ( Representing the (hidden data), the model parameters for the sample are Joint distribution Conditional distribution The maximum number of iterations is At this point, the likelihood function of the maximized model distribution is... for:

[0101] ,in, To observe the data, This is implicit data;

[0102] The model parameters are randomly initialized through EM algorithm iteration. initial value ,according to Begin iterating until convergence:

[0103] Step E: Calculate the posterior probability of the latent variables based on the initial values ​​or the parameters from the previous iteration. This probability is used as the current estimate of the latent variables. The calculation formula is as follows:

[0104] ;

[0105] In the formula, For the first Conditional probability distribution of observed data for The posterior probability distribution.

[0106] M-step: The likelihood function is maximized as follows:

[0107]

[0108] In the formula, For the first The joint probability distribution of the observed data.

[0109] If parameter If convergence is achieved, the iteration ends; otherwise, the iteration continues.

[0110] The final output model maximizes the parameters. .

[0111] Specifically, diagnostic inference is performed based on the dynamic Bayesian network risk evolution model, outputting the posterior probability distribution of the risk factor nodes within each time slice, including:

[0112] The high-risk state of the target risk node is taken as known evidence and input into the dynamic Bayesian network risk evolution model;

[0113] The dynamic Bayesian network risk evolution model performs reverse probabilistic inference to calculate the posterior probability distribution of the risk factor nodes.

[0114] The purpose of diagnostic reasoning is to trace the main factors leading to a known high-risk node. First, the target risk node (e.g., "comprehensive security risk") is set to a high-risk state as known evidence and input into the DBN model. For example, at time slice T, assuming the comprehensive security risk is high, this evidence is encoded as node state value 1 (if it is medium risk, the node state value is 2; and if it is low risk, the node state value is 3).

[0115] Then, the DBN model performs backward probabilistic inference. Backward inference, based on Bayes' theorem, calculates the posterior probability distribution of each risk factor node given evidence of the target node. For example, for For a network with multiple time slices, assuming that the various unsafe behavior factors affecting the safety of uninterrupted construction are variables... ,by For example, in the observed values (in, For the first The 30th observation corresponding to each time slice. For the first Under the first observation corresponding to each time slice, In time slice The posterior probability of is:

[0116] .

[0117] By assuming conditional independence, the joint probability distribution is:

[0118] ;

[0119] in, For nodes The set of parent nodes.

[0120] By repeating the reasoning calculations and moving forward, the posterior probability distribution of risk nodes within each time slice can be obtained sequentially until the starting time slice is reached, thus describing the construction risks and temporal evolution characteristics of each stage in the entire non-stop construction process.

[0121] Among them, unsafe behaviors of construction workers identified based on on-site images of the construction area include:

[0122] Build a knowledge base of unsafe behaviors;

[0123] Historical images are labeled based on the aforementioned unsafe behavior knowledge base to obtain a training set;

[0124] Construct a support vector machine model, train the support vector machine model based on the training set, and obtain the recognition model;

[0125] Input the on-site images of the construction area into the recognition model, and output the results of unsafe behavior recognition.

[0126] The labeling of the specific training set and the training of the support vector machine model are existing technologies in this field, which can be obtained by those skilled in the art through consulting relevant materials, and will not be elaborated on here.

[0127] The construction of the unsafe behavior knowledge base includes:

[0128] Collect historical construction accident reports for the airfield;

[0129] The accident report text is cleaned and segmented to obtain the text processing results;

[0130] Build a custom word library based on the construction standards of the flight area, and merge phrases in the text processing result based on the custom word library to obtain key cause words;

[0131] Use the Bert model to extract human cause factors from the key cause words;

[0132] Calculate the eigenvalue weights of each human cause factor using the term frequency-inverse document frequency algorithm;

[0133] Arrange the human cause factors in descending order according to the eigenvalue weights, and select the top preset number of human cause factors to construct the unsafe behavior knowledge base.

[0134] Specifically, during implementation, collect incident investigation reports related to flight area construction accidents from channels such as government departments, safety information networks, civil aviation resource networks, and airport management agencies, and perform text cleaning and word segmentation on the accident report text. Use the yaha toolkit for word segmentation in the Python environment, remove stop words (such as "de", "he", etc.) and irrelevant characters, and extract key words.

[0135] When building a custom word library based on the construction standards of the flight area, refer to documents such as the "Administrative Measures for Non-stop Construction of Transport Airports" and the "Technical Standards for Civil Airport Flight Areas", list professional terms, such as "ground protection area", "taxiway", "safety helmet", etc., and merge synonyms (such as "safety helmet" and "helmet" unified as "safety helmet").

[0136] Use the Bert model to extract human cause factors from the word-segmented text. For example, input the sentence "The accident was caused by construction workers not wearing safety helmets", and output "not wearing safety helmets" as the human cause.

[0137] Calculate the eigenvalue weights of each human cause factor using the term frequency-inverse document frequency (TF-IDF) algorithm. The formula is:

[0138] ;

[0139] ;

[0140] ;

[0141] Among them, is the term frequency, is the number of times the word appears in , is the sum of all word occurrences in the file , is the inverse document frequency, is the total number of reports in the dataset, represents the file Contains words The number of reports; The feature values ​​are weighted by the words. The top 30 words with the highest feature weights are selected as typical unsafe construction behaviors in the unsafe behavior knowledge base for subsequent identification steps.

[0142] Among these measures, personnel intrusion monitoring of key activity areas adjacent to the construction area includes:

[0143] Based on the three-dimensional coordinates of the key points of the temporary boundary water-filled barriers in the key activity area, the planar equation of the temporary boundary is constructed.

[0144] Establish a transformation model between the world coordinate system, camera coordinate system, image coordinate system, and pixel coordinate system;

[0145] The actual position of the intruder in the world coordinate system is determined by the transformation model, and the actual position is compared with the plane equation of the temporary boundary to generate the intrusion monitoring result.

[0146] When constructing the planar equations for the temporary boundary, the coordinates of the two three-dimensional key points of the single row of barriers constituting the temporary boundary are considered as the reference points for dividing the plane. For example, assuming the three-dimensional coordinates of the near and far ends of the single row of barriers are respectively... and .based on , The direction vector of the line containing the single-row barricade, determined by two points, can be represented as:

[0147] ;

[0148] The direction vector in the above formula This describes the spatial extension direction of a single row of water-filled barriers. Since the barriers are all the same size and nearly rectangular (the thickness is negligible), the direction determined by the near and far endpoints can be considered as the extension direction of the electronic fence on the horizontal plane. The normal vector is obtained by taking the cross product with the vertical unit vector. The spatial relationship between the electronic fence and the ground is defined as perpendicular to the direction of the single-row barricade, ensuring that the fence is a vertical plane. In the camera coordinate system of this invention, the direction perpendicular to the ground is used as the vertical direction vector. Then at this time... Based on this, the plane equation of the electronic fence is obtained as follows:

[0149] ;

[0150] Simplifying, we get .

[0151] A transformation model is established between the world coordinate system, camera coordinate system, image coordinate system, and pixel coordinate system. This model is based on the camera imaging principle, and the transformation process involves the following three steps:

[0152] A. Transformation from world coordinate system to camera coordinate system: This transformation is a rigid body transformation, determined by the rotation matrix. Translation vector Description. Let the homogeneous coordinates of a point P in the world coordinate system be... The formula for transforming it to the camera coordinate system is:

[0153] ;

[0154] In the formula, the rotation matrix An orthogonal matrix describing the camera's orbit around the world coordinate system. Rotation angle of the axis; translation vector Describe the coordinates of the camera's optical center in the world coordinate system; This is the camera extrinsic parameter matrix.

[0155] B. Transformation from camera coordinate system to image coordinate system: This transformation is based on the pinhole imaging principle, and the formula is:

[0156] ;

[0157] C. Transformation from image coordinate system to pixel coordinate system: The transformation relationship is as follows:

[0158] ;

[0159] Finally, by combining the above three steps, we obtain the direct mapping relationship from world coordinates to pixel coordinates: ;

[0160] in, Using world coordinates to describe the camera position, in meters (m). The coordinate system is the camera coordinate system, with the optical center at the origin, and the unit is meters (m). The coordinate system is the image coordinate system, with the optical center at the midpoint of the image, and the unit is mm. This is a pixel coordinate system, with the origin at the top left corner of the image, and the unit is pixels. The imaging point in the image has the following coordinates in the image coordinate system: The coordinates in the pixel coordinate system are , The camera focal length, in mm, equals and The distance, i.e. , and This indicates how many mm each column and each row represent. and The origin of the image coordinate system is respectively Coordinates in pixel coordinate system.

[0161] The actual location of the intruder in the world coordinate system is determined by a transformation model. The outline of the person is detected from the image, the pixel coordinates of its bottom midpoint are calculated, and then the world coordinates are obtained through coordinate transformation.

[0162] The actual location is compared with the plane equation of the temporary boundary, and the distance from the point to the plane is calculated. If the distance is less than or equal to 0, it indicates that personnel have intruded into the critical activity area, and an intrusion monitoring result is generated.

[0163] The process of obtaining the final early warning result based on the results of unsafe behavior identification and intrusion monitoring includes:

[0164] Calculate the first warning confidence level corresponding to the unsafe behavior identification result, and calculate the second warning confidence level corresponding to the intrusion detection result;

[0165] The final warning confidence level is calculated using a weighted product fusion strategy based on the first warning confidence level and the second warning confidence level.

[0166] The final warning result is output based on the preset range in which the final warning confidence level falls.

[0167] The formula for calculating the second warning confidence level corresponding to the intrusion detection result is as follows:

[0168] ;

[0169] in, This represents the second warning confidence level. The distance between the actual location of the intruders and the temporary perimeter fence. The width of the entry threshold zone can be determined by those skilled in the art based on the actual construction situation.

[0170] As for the confidence level of the first warning The confidence scores of the target category, the detection confidence scores of the protective equipment, and the tracking stability confidence scores are calculated using a weighted summation method, with the weights determined based on the construction scenario and regulatory focus.

[0171] Specifically, the confidence score of the target category within the slice at any given time is obtained using the following method:

[0172] Count the total number of unsafe behaviors identified in each category within the time slice. For example, in a certain time slice, a total of 150 times of not wearing a safety helmet, 80 times of not wearing a reflective vest, and 30 times of not wearing safety goggles while operating cutting equipment were identified.

[0173] The category of unsafe behavior that occurs most frequently is identified as the dominant unsafe behavior category for that time slot. For example, in the time slot mentioned above, not wearing a helmet is the dominant unsafe behavior category.

[0174] Calculate the frequency percentage of the dominant unsafe behavior category within the time slice. For example, the frequency percentage of the dominant unsafe behavior category in the above time slice is 150 / (150+80+30)=57.7%.

[0175] Within the time slice, acquire the dominant frame image that identifies the dominant unsafe behavior category, and calculate the average of the original confidence scores of all the dominant frame images;

[0176] The confidence level of the target category is obtained by multiplying the frequency percentage of the dominant unsafe behavior category by the average of the original confidence levels of all the dominant frame images.

[0177] The original confidence level is obtained by converting the decision function value corresponding to the dominant frame image output by the recognition model using the Platt Scaling method. The specific conversion method is a prior art in this field.

[0178] The detection confidence level of the protective equipment at any given time slice is obtained using the following method:

[0179] The total duration T of the image frames corresponding to unsafe behaviors related to protective equipment identified within this time slice is calculated. v And the longest duration T of unsafe acts related to protective equipment. m Among these, unsafe behaviors related to protective equipment include: not wearing a safety helmet, not wearing a reflective vest, and not wearing safety goggles when operating cutting equipment; assuming that in a certain time frame, it is determined from image frames that a construction worker did not wear a reflective vest for a maximum of 4 hours, then T m It lasts for 4 hours.

[0180] Calculate T v Relative to the total duration T of the time slice t The percentage of total violation time R v ;

[0181] Calculate T m Relative to T t The percentage of violations that obtained the longest violation percentage R m ;

[0182] Acquire equipment violation frames within the time slice that identify unsafe behaviors related to protective equipment, and calculate the average original confidence level P of all equipment violation frames. v ;

[0183] Based on Rm R v and P v Calculate the detection confidence level of the protective equipment. Those skilled in the art can choose an appropriate method for calculation as needed. As a specific implementation method, R... m R v and P v After normalizing each product separately, they are multiplied directly.

[0184] The tracking stability confidence score for any given time slice is obtained using the following method:

[0185] The cumulative duration T during which the tracked target was successfully and continuously tracked within this time slice is calculated. s And the total duration T of the target's appearance within that time slice. T Calculate T s Relative to T T The proportion of the trajectory completeness R is obtained. c For example, if a construction worker's tracking is interrupted several times within a 4-hour time slice due to obstruction or other reasons, and the total successful tracking time is 3.5 hours, then the trajectory integrity Rc is 3.5 / 4.0 = 87.5%.

[0186] If a tracking interruption occurs within this time slice followed by re-association with the same target, the re-identification confidence of all re-identification events is obtained, and their average value is calculated to obtain the average re-identification confidence R. i The re-identification confidence score is obtained by comparing the similarity between the appearance features of the target in the last frame before the trajectory interruption and the appearance features of the target in the first frame after the trajectory recovery. For example, the cosine similarity of the features of the two frames can be used as the re-identification confidence score. If no tracking interruption occurs, then R... i Set it to 1.0.

[0187] Analyze the target's complete trajectory throughout the entire time slice and calculate its trajectory rationality R. p Specifically, the instantaneous velocity generated by the displacement between consecutive frames in the trajectory is calculated, and the proportion of frames whose instantaneous velocity falls within a preset reasonable velocity range is counted out of the total number of trajectory frames. This proportion is used as the trajectory reasonableness score R. p For example, if the reasonable walking speed for construction workers is preset to [0, 2] meters per second, and if the instantaneous speed of 950 out of 1000 frames in the total trajectory falls within this range, then R p It is 95%.

[0188] Based on R c R i and R p Calculate the tracking stability confidence level C. t As a specific implementation method, R...c R i and R p After normalizing each product separately, they are multiplied directly.

[0189] Finally, a weighted product fusion strategy is used to determine the final warning confidence level. :

[0190]

[0191] in, , The second confidence level and first confidence level The corresponding module weights are determined based on the information provided about the target being identified. Used to penalize situations where the confidence level of a module is too low;

[0192] Based on the final warning confidence level The system classifies the threat levels of construction personnel behavior into three levels, triggering different response actions: Level 1: Highly likely to result in ground protection zone intrusion and personnel casualties; Level 2: Likely to result in rapid departure from the taxiway and intrusion into the runway-parallel taxiway, leading to personnel casualties; Level 3: Likely to result in apron taxiway intrusion and personnel injury. For example, a final warning confidence level ≥ 0.8 is classified as Level 1, 0.5-0.8 as Level 2, and 0.3-0.5 as Level 3.

[0193] Example 2

[0194] Please refer to Figure 2 Embodiment 2 of the present invention provides a data management system for non-stop construction in flight zones, the system comprising:

[0195] The intelligent identification and early warning unit is used to acquire on-site images of the construction area, identify unsafe behaviors of construction personnel based on the on-site images of the construction area, obtain unsafe behavior identification results, implement personnel intrusion monitoring in key activity areas adjacent to the construction area, obtain intrusion monitoring results, obtain a final early warning result based on the unsafe behavior identification result and the intrusion monitoring result, and classify the threat level based on the final early warning result;

[0196] Dynamic risk assessment units are used to divide the construction process into several consecutive time slices, where any one time slice represents a construction stage.

[0197] The dynamic risk assessment unit is also used to define target risk nodes, take the threat level corresponding to each time slice as the risk factor node of that time slice, establish the temporal dependency relationship of the same risk factor nodes between consecutive time slices, and construct a dynamic Bayesian network risk evolution model based on the temporal dependency relationship. The target risk node is used to characterize the overall safety status of construction.

[0198] The dynamic risk assessment unit is also used to perform diagnostic reasoning based on the dynamic Bayesian network risk evolution model, output the posterior probability distribution of the risk factor nodes in each time slice, and form a risk distribution heatmap based on the posterior probability distribution of the risk factor nodes in each time slice.

[0199] The decision-making unit is used for construction management based on the risk distribution heat map.

[0200] Example 3

[0201] Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of any one of the flight area non-stop construction data management methods described in Embodiment 1.

[0202] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0203] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for managing data of non-stop construction in a flight zone, characterized by, The method comprises the following steps: Obtain a construction area field image, identify unsafe behaviors of construction personnel based on the construction area field image, and obtain an unsafe behavior identification result; Implement personnel intrusion monitoring on a key activity area adjacent to the construction area, and obtain an intrusion monitoring result; Obtain a final warning result based on the unsafe behavior identification result and the intrusion monitoring result, and divide a threat level based on the final warning result; Divide a construction process into a plurality of continuous time slices, wherein any time slice represents a construction stage; Define a target risk node, take the threat level corresponding to each time slice as a risk factor node of the time slice, establish a time sequence dependency relationship of the same risk factor node between the continuous time slices, construct a dynamic Bayesian network risk evolution model based on the time sequence dependency relationship, and the target risk node is used to represent a construction overall safety state; Perform diagnostic reasoning based on the dynamic Bayesian network risk evolution model, and output a posterior probability distribution of the risk factor node in each time slice; Form a risk distribution heat map based on the posterior probability distribution of the risk factor node in each time slice; Perform construction management based on the risk distribution heat map; The personnel intrusion monitoring on the key activity area adjacent to the construction area comprises: Based on the three-dimensional coordinates of the key points of the temporary fence water horse of the key activity area, a plane equation of the temporary fence is constructed; A conversion model between a world coordinate system, a camera coordinate system, an image coordinate system and a pixel coordinate system is established; The actual position of the intruding personnel in the world coordinate system is determined through the conversion model, and the actual position is compared with the plane equation of the temporary fence to generate the intrusion monitoring result; The final warning result based on the unsafe behavior identification result and the intrusion monitoring result comprises: A first warning confidence corresponding to the unsafe behavior identification result is calculated, and a second warning confidence corresponding to the intrusion monitoring result is calculated; The final warning confidence is calculated based on the first warning confidence and the second warning confidence by using a weighted product fusion strategy; The final warning result is output according to a preset range in which the final warning confidence is located; The first warning confidence is calculated by using a weighted summation method, which comprehensively considers a target category confidence, a protective equipment detection confidence and a tracking stability confidence; The tracking stability confidence of any time slice is obtained by the following method: counting the total time T that the tracked target is successfully continuously tracked within the time slice s and the total time T that the target appears within the time slice T ; Compute T s The ratio of T T , obtain the trajectory integrity R c ; If the same target is re-associated after a tracking interruption in the time slice, the re-identification confidence of all re-identification association events is obtained, and the average value is calculated to obtain the average re-identification confidence R i ; the re-identification confidence R i The similarity between the appearance features of the last frame of the target before the track interruption and the appearance features of the first frame of the target after the track recovery is compared to obtain analyze the complete motion trajectory of the target in the whole time slice, and calculate the motion trajectory rationality R p , specifically, calculate the instantaneous speed generated by the displacement between the continuous frames in the trajectory, and count the proportion of the frame number in the preset reasonable speed interval to the total trajectory frame number, and take the proportion as the trajectory rationality R p ; based on R c , R i , and R p to calculate the tracking stability confidence C t .

2. The method of claim 1, wherein, The node parameters of the dynamic Bayesian network risk evolution model are obtained by the following method: The prior probability of the root node in the risk factor node is determined by using an expert investigation method based on a trapezoidal fuzzy decision; For non-root nodes, the initial condition probability is determined based on a Leaky Noisy-or Gate extension model; The initial condition probability of the non-root node is iteratively corrected based on an expectation maximization algorithm.

3. The method of claim 1, wherein the method further comprises: The diagnostic reasoning based on the dynamic Bayesian network risk evolution model and the output of the posterior probability distribution of the risk factor node in each time slice comprise: The target risk node in a high risk state is taken as known evidence, and is input into the dynamic Bayesian network risk evolution model; The dynamic Bayesian network risk evolution model performs reverse probability reasoning to calculate a posterior probability distribution of the risk factor node.

4. The method of claim 1, wherein the method further comprises: The unsafe behavior of the construction personnel is identified based on the on-site image of the construction area, including: An unsafe behavior knowledge base is constructed; The historical images are labeled based on the unsafe behavior knowledge base to obtain a training set; A support vector machine model is constructed, and the support vector machine model is trained based on the training set to obtain an identification model; The on-site image of the construction area is input into the identification model, and an unsafe behavior identification result is output.

5. The method of claim 4, wherein the method further comprises: The construction of the unsafe behavior knowledge base includes: Flight area historical construction accident report texts are collected; Text cleaning and word segmentation processing are performed on the accident report texts to obtain text processing results; A self-defined word library is established based on the flight area construction standards, and key cause words are obtained by merging word groups based on the self-defined word library; Human cause factors are extracted from the key cause words by using a Bert model; The characteristic value weights of the human cause factors are calculated by using a term frequency-inverse document frequency algorithm; The human cause factors are arranged in descending order of the characteristic value weights, and a preset number of human cause factors with high rankings are selected to construct the unsafe behavior knowledge base.

6. The method of claim 1, wherein, The formula for calculating the second early warning confidence corresponding to the intrusion monitoring result is: ； wherein, is the second early warning confidence, is the distance between the actual position of the intruder and the temporary perimeter, is the ingress threshold zone width.

7. A flight zone non-stop construction data management system, characterized by, The system includes: An intelligent identification and early warning unit is configured to acquire on-site images of a construction area, identify unsafe behaviors of construction personnel based on the on-site images of the construction area, obtain an unsafe behavior identification result, perform personnel intrusion monitoring on a key activity area of a neighboring construction area to obtain an intrusion monitoring result, obtain a final early warning result based on the unsafe behavior identification result and the intrusion monitoring result, and divide a threat level based on the final early warning result; A dynamic risk assessment unit is configured to divide a construction process into a plurality of continuous time slices, wherein any time slice represents a construction stage. The dynamic risk assessment unit is further configured to define a target risk node, take the threat level corresponding to each time slice as a risk factor node of the time slice, establish a time sequence dependency relationship of the risk factor nodes between the continuous time slices, construct a dynamic Bayesian network risk evolution model based on the time sequence dependency relationship, and use the target risk node to represent a construction overall safety state. The dynamic risk assessment unit is further configured to perform diagnostic reasoning based on the dynamic Bayesian network risk evolution model, output a posterior probability distribution of the risk factor node in each time slice, and form a risk distribution heat map based on the posterior probability distribution of the risk factor node in each time slice. A decision unit is configured to perform construction management based on the risk distribution heat map. The personnel intrusion monitoring on the key activity area of the neighboring construction area includes: A plane equation of the temporary enclosure is constructed based on the three-dimensional coordinates of the key points of the temporary enclosure water mark of the key activity area; A conversion model between a world coordinate system, a camera coordinate system, an image coordinate system, and a pixel coordinate system is established; and The actual position of the intruder in a world coordinate system is determined through the conversion model, and the actual position is compared with a plane equation of the temporary enclosure to generate the intrusion monitoring result; The final warning result is obtained based on the unsafe behavior identification result and the intrusion monitoring result, including: A first warning confidence corresponding to the unsafe behavior identification result is calculated, and a second warning confidence corresponding to the intrusion monitoring result is calculated; A final warning confidence is calculated based on the first warning confidence and the second warning confidence using a weighted product fusion strategy; The final warning result is output according to a preset range in which the final warning confidence is located; The first warning confidence is calculated using a weighted summation method, which comprehensively considers a target category confidence, a protective equipment detection confidence, and a tracking stability confidence; The tracking stability confidence of any time slice is obtained through the following method: counting the accumulated time length T that the tracked target is successfully continuously tracked within the time slice s and the total time length T that the target appears within the time slice T ; Compute T s The ratio of T T , obtain the trajectory integrity R c ; If the same target is re-associated after a tracking interruption in the time slice, the re-identification confidence of all re-identification association events is obtained, and the average value is calculated to obtain the average re-identification confidence R i ; the re-identification confidence R i The similarity between the appearance features of the last frame of the target before the track interruption and the appearance features of the first frame of the target after the track recovery is compared to obtain analyze the complete motion trajectory of the target in the whole time slice to calculate a motion trajectory rationality R p Specifically, the instantaneous speed generated by the displacement between the continuous frames in the trajectory is calculated, the proportion of the frame number in which the instantaneous speed is in a preset reasonable speed interval in the total trajectory frame number is counted, and the proportion is taken as the trajectory rationality R p ; Based on R c , R i , and R p , the tracking stability confidence C t is calculated.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program, when executed by a processor, implements the steps of the flight zone non-stop construction data management method according to any one of claims 1-6.

Images