Construction site safety situation awareness method and system based on deep learning network

Abstract

The application discloses a construction site safety situation awareness method and system based on a deep learning network, relates to the technical field of building construction safety management, and comprises the following steps: acquiring multi-source sensing data of a construction site, extracting a multi-dimensional feature vector through a deep learning network; identifying a work stage based on a feature component; extracting a hazard source feature component and constructing an interaction matrix, and obtaining a coupled risk factor through coupling operation; calculating a current risk quantitative value according to a work stage identifier and a risk propagation rule; determining a stage evolution sequence and calculating a subsequent stage risk quantitative value to generate a risk evolution sequence; and generating an early warning signal when the risk quantitative value exceeds a safety threshold. The application realizes dynamic perception and prediction of the safety situation of a construction site, improves the accuracy and foresight of risk early warning, and provides an intelligent solution for construction site safety management.

Description

Technical Field

[0001] This invention relates to the field of construction safety management technology, specifically to a method and system for perceiving safety situations at construction sites based on deep learning networks. Background Technology

[0002] Currently, safety monitoring at construction sites mainly relies on single data sources such as video surveillance and sensor monitoring, which cannot comprehensively and accurately reflect the safety situation at the construction site. Although some studies have attempted to acquire multi-source data using various sensing devices, the lack of effective data fusion and analysis methods makes it difficult to deeply explore the inherent correlations between multi-source data, resulting in low accuracy in safety risk identification and poor timeliness of early warnings.

[0003] In terms of risk assessment, most existing technologies employ static assessment models, which cannot adapt to the dynamically changing risk conditions during construction. Construction is a continuous process with close connections between each stage; risks in previous stages often affect the safety status of subsequent stages. Traditional risk assessment methods neglect the dynamic propagation characteristics of risks, making it difficult to accurately predict risk evolution trends and thus failing to provide early warnings of potential risks.

[0004] Most existing construction site safety management systems lack consideration for the interactions between various hazards. Construction sites contain a wide variety of hazards, including environmental, equipment, and personnel factors, which interact and couple with each other in complex ways. Traditional methods often consider each hazard in isolation, neglecting the impact of their interactions on the overall safety situation. This leads to significant discrepancies between risk assessment results and actual conditions, hindering a comprehensive understanding and accurate prediction of the construction site's safety status.

[0005] In summary, there is an urgent need for a construction site safety situation awareness method that can effectively integrate multi-source sensing data, dynamically assess risk status, and accurately predict risk evolution trends, in order to improve the intelligence level and early warning capabilities of construction site safety management. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for safety situation awareness at construction sites based on deep learning networks, aiming to solve at least one of the technical problems existing in the prior art.

[0007] The technical solution of this invention is: a construction site safety situation awareness method based on deep learning networks, comprising the following steps: Acquire multi-source sensing data from the construction site, and extract features from the multi-source sensing data using a pre-constructed deep learning network to obtain multi-dimensional feature vectors; The current construction operation stage is identified based on the feature components in the multidimensional feature vector, thus obtaining the operation stage identifier. Extract the hazard source feature components corresponding to multiple hazard sources within a preset spatial range from the multidimensional feature vector, construct an interaction matrix based on the hazard source feature components, and perform coupling operations on the hazard source feature components and the interaction matrix to obtain the coupled risk factor; Based on the work phase identifier, the corresponding risk propagation rule is retrieved, and the coupled risk factors are substituted into the risk propagation rule for calculation to obtain the risk quantification value of the current work phase. The corresponding stage evolution sequence is determined based on the work stage identifier. Based on the risk quantification value of the current work stage and the risk propagation rule of the next work stage in the stage evolution sequence, the risk quantification value of the next work stage is calculated, and the risk evolution sequence is generated. Traverse the risk evolution sequence. When the risk quantification value in the risk evolution sequence exceeds the safety threshold of the corresponding operation stage, extract the moment when the safety threshold is exceeded and generate an early warning signal.

[0008] Multi-source sensing data from the construction site is acquired, and features are extracted from the multi-source sensing data using a pre-constructed deep learning network to obtain a multi-dimensional feature vector, including: Construction operation video data is collected by video acquisition equipment deployed at the construction site, and environmental monitoring data and equipment status data are collected by sensor equipment deployed at the construction site. The construction operation video data, environmental monitoring data and equipment status data are time-stamped and aligned to obtain multi-source sensing data. Construction operation video data from multi-source sensing data is input into a pre-constructed deep learning network, which includes a video coding branch and a numerical coding branch. Video feature vectors are obtained by feature extraction through the video coding branch, and numerical feature vectors are obtained by feature encoding through the numerical coding branch. The spatial correlation between the spatial location information in the video feature vector and the sensor location information in the numerical feature vector is calculated. Based on the spatial correlation, the video feature vector and the numerical feature vector are weighted and fused to obtain a multidimensional feature vector.

[0009] Based on the feature components in the multidimensional feature vector, the current construction operation stage is identified, resulting in the following operation stage identifiers: Feature components are extracted from the multidimensional feature vector, and the feature components are divided into progress feature components and state feature components. Calculate the feature similarity between the progress feature components and each work stage in the preset work stage mapping table, and select the work stage with the highest feature similarity as the candidate work stage; Extract the duration of the candidate operation stage, calculate the duration of the current construction operation based on the state feature components, and compare the duration of the current operation with the duration of the candidate operation stage to obtain the duration deviation. When the time deviation is less than the preset deviation threshold, the candidate operation stage is determined as the current operation stage. When the time deviation is greater than the preset deviation threshold and has lasted for a duration exceeding the stage duration, the next operation stage of the candidate operation stage is determined as the current operation stage. Based on the current work stage, the corresponding work stage identifier is obtained from the work stage mapping table.

[0010] Extract hazard source feature components corresponding to multiple hazard sources within a preset spatial range from the multidimensional feature vector. Construct an interaction matrix based on the hazard source feature components. Perform coupling operations on the hazard source feature components and the interaction matrix to obtain coupled risk factors, including: Extract the hazard source feature components corresponding to multiple hazard sources located within a preset spatial range from the multidimensional feature vector; Extract the type identifier and location information of each hazard source from the hazard source feature components, and filter out hazard source pairs with triggering relationships based on the type identifier and location information; For a pair of hazard sources, the hazard source that initiates the risk transmission is marked as the transmission source and the feature components of the transmission source are extracted. The hazard source that receives the risk transmission is marked as the transmission target and the feature components of the transmission target are extracted. A directed association graph is constructed based on the connection relationship between the transmission source and the transmission target. The risk transmission coefficient is calculated based on the feature components of the transmission source and the feature components of the transmission target. The risk transmission coefficient is then filled into the connection position between the transmission source and the transmission target in the directed association graph to generate the interaction matrix. Multiple hazard source feature components are arranged in the order of hazard sources in the directed association graph to form an initial risk vector. The initial risk vector is then multiplied by the interaction matrix in an iterative matrix operation. The operation stops when the difference between two adjacent iterations is less than a preset difference threshold or the number of iterations reaches a preset upper limit, thus obtaining a steady-state risk vector. The coupled risk factor is obtained by weighted summation of the risk components corresponding to each hazard source in the steady-state risk vector.

[0011] Based on the work phase identifier, the corresponding risk propagation rule is retrieved. The coupled risk factors are substituted into the risk propagation rule for calculation, and the risk quantification value of the current work phase is obtained, including: Based on the operation stage identifier, retrieve the corresponding risk propagation rule from the preset risk propagation rule base, and extract the risk propagation path from the risk propagation rule; Collect the personnel density distribution and equipment distribution at each work location within the construction work area, construct a risk carrying capacity field along the risk propagation path based on the personnel density distribution and equipment distribution, and extract the local carrying capacity value corresponding to each work location from the risk carrying capacity field; By substituting the coupled risk factors into the risk propagation rule, the risk allocation weight of each work location is calculated based on the local carrying capacity value corresponding to each work location on the risk propagation path. The coupled risk factors are then allocated to each work location according to the risk allocation weight to obtain the local risk component of each work location. For the local risk components of each work location, it is determined whether the local risk components exceed the local carrying capacity value of the corresponding work location. If they do, the excess portion is extracted as the overflow risk quantity and transmitted to the adjacent work location along the risk propagation path. The local risk components of the work location receiving the overflow risk quantity are recalculated and iteratively determined until no overflow risk quantity is generated. The local risk components of each work location at the end of the iteration are accumulated to obtain the risk quantification value of the current work stage.

[0012] Based on the work stage identifier, the corresponding stage evolution sequence is determined. Based on the risk quantification value of the current work stage and the risk propagation rule for the next work stage in the stage evolution sequence, the risk quantification value for the next work stage is calculated, generating the risk evolution sequence, which includes: Based on the work stage identifier, query the corresponding stage evolution sequence and extract the subsequent work stages and corresponding risk propagation rules in the stage evolution sequence. Based on the risk propagation rules of the current operation phase, the risk quantification value of the current operation phase is decomposed into a current risk value and an inherent risk value. Calculate the environmental difference between the current operation stage and the next operation stage, obtain the stage transition decay rate based on the environmental difference, and perform a decay calculation on the flow risk value and the stage transition decay rate to obtain the transmission risk value. Extract the hazard trigger threshold and risk amplification coefficient from the risk propagation rules of the next operation phase. When the transmitted risk value reaches the hazard trigger threshold, the amplified transmitted risk value is calculated based on the transmitted risk value and the risk amplification coefficient. Otherwise, the transmitted risk value is used as the amplified transmitted risk value. Extract the basic risk value of the next operation stage from the risk propagation rules of the next operation stage, and then combine the amplified transmission risk value, the basic risk value of the stage, and the inherent risk value to obtain the risk quantification value of the next operation stage. Repeat the iteration until all subsequent operation stages have been traversed, and arrange the risk quantification values ​​of each operation stage in order to generate a risk evolution sequence.

[0013] Traverse the risk evolution sequence. When the risk quantification value in the risk evolution sequence exceeds the safety threshold of the corresponding operation stage, extract the moment when the safety threshold is exceeded and generate an early warning signal, including: The risk quantification value of each operation stage in the risk evolution sequence is traversed in chronological order to obtain the safety threshold of each operation stage. When the risk quantification value exceeds the safety threshold, the corresponding operation stage is marked as the over-limit operation stage, and the time corresponding to the over-limit operation stage is extracted as the time when the safety threshold is exceeded. Extract the preceding and subsequent operation stages in the risk evolution sequence of the over-limit operation stage, calculate the preceding difference between the risk quantification value of the over-limit operation stage and the risk quantification value of the preceding operation stage, calculate the subsequent difference between the risk quantification value of the subsequent operation stage and the risk quantification value of the over-limit operation stage, and determine the risk evolution trend of the over-limit operation stage based on the preceding and subsequent differences. Based on the risk evolution trend, determine whether the risk quantification value of the over-limit operation stage is in a continuous upward state. When it is in a continuous upward state, the moment when the safety threshold is exceeded and the moment of the subsequent operation stage will be used as the warning time range and a warning signal will be generated. When it is not in a continuous upward state, the moment when the safety threshold is exceeded will be used as the warning time range and a warning signal will be generated.

[0014] This invention provides a construction site safety situation awareness system based on deep learning networks, the system comprising: The data processing module is used to acquire multi-source sensing data from the construction site and extract features from the multi-source sensing data through a pre-built deep learning network to obtain multi-dimensional feature vectors. The stage identification module is used to identify the current construction stage based on the feature components in the multi-dimensional feature vector, and obtain the work stage identifier. The coupling calculation module is used to extract the hazard source feature components corresponding to multiple hazard sources within a preset spatial range in the multidimensional feature vector, construct an interaction matrix based on the hazard source feature components, and perform coupling calculations on the hazard source feature components and the interaction matrix to obtain the coupling risk factor. The risk quantification module is used to retrieve the corresponding risk propagation rule based on the work stage identifier, substitute the coupled risk factors into the risk propagation rule for calculation, and obtain the risk quantification value of the current work stage. The evolution prediction module is used to determine the corresponding stage evolution sequence based on the operation stage identifier, calculate the risk quantification value of the next operation stage based on the risk quantification value of the current operation stage and the risk propagation rule of the next operation stage in the stage evolution sequence, and generate the risk evolution sequence. The early warning generation module is used to traverse the risk evolution sequence. When the risk quantification value in the risk evolution sequence exceeds the safety threshold of the corresponding operation stage, the module extracts the moment when the safety threshold is exceeded and generates an early warning signal.

[0015] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0016] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps in any of the aforementioned methods.

[0017] This invention utilizes deep learning networks to extract features and fuse data from multiple sources, achieving comprehensive awareness of the safety situation at construction sites and overcoming monitoring blind spots caused by single data sources. By identifying construction operation stages and constructing a hazard interaction matrix, it effectively assesses the coupling effects between multiple hazard sources, improving the accuracy of risk assessment. Based on risk propagation rules, it calculates the quantified risk value of the current operation stage and predicts the risk evolution trend of subsequent stages, enabling safety management to shift from static assessment to dynamic prediction and achieving proactive risk identification. Through traversal analysis of risk evolution sequences, it can identify potential risks in advance and generate early warning signals, effectively extending the early warning response time. Attached Figure Description

[0018] Figure 1 A flowchart of a construction site safety situation perception method based on deep learning networks provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the construction site safety situation awareness system based on deep learning network according to an embodiment of the present invention. Detailed Implementation

[0019] like Figure 1 As shown, Figure 1 A flowchart of a construction site safety situation awareness method based on deep learning networks provided in an embodiment of the present invention, the method comprising the following steps: Step 101: Obtain multi-source sensing data from the construction site, and extract features from the multi-source sensing data using a pre-built deep learning network to obtain a multi-dimensional feature vector.

[0020] In some embodiments of the present invention, step 101 may specifically include the following sub-steps: Sub-step 1011: Collect construction operation video data through video acquisition equipment deployed at the construction site, collect environmental monitoring data and equipment status data through sensor equipment deployed at the construction site, and align the construction operation video data, environmental monitoring data and equipment status data with timestamps to obtain multi-source sensing data; Sub-step 1012: Input the construction operation video data from the multi-source sensing data into a pre-constructed deep learning network, which includes a video coding branch and a numerical coding branch; Sub-step 1013: Extract video feature vectors through video coding branch, and encode numerical feature vectors through numerical coding branch. Sub-step 1014: Calculate the spatial correlation between the spatial location information in the video feature vector and the sensor location information in the numerical feature vector. Based on the spatial correlation, perform weighted fusion of the video feature vector and the numerical feature vector to obtain a multi-dimensional feature vector.

[0021] Video capture equipment is deployed at the construction site to collect operational video data. Simultaneously, multiple sensor devices are deployed to collect environmental monitoring data and equipment status data. The video capture equipment includes high-definition cameras, which can be installed in key locations on the construction site, such as around tower cranes, material storage areas, and construction elevators, covering the main areas of the site. Environmental monitoring data is collected by temperature and humidity sensors, dust concentration sensors, and noise sensors, reflecting the conditions of the construction environment. Equipment status data is collected by vibration sensors, tilt sensors, and pressure sensors, monitoring the operating status of the construction equipment.

[0022] The collected data undergoes timestamp alignment to ensure temporal consistency across multiple data sources. Timestamp alignment employs nearest-neighbor interpolation to unify data sampled at different frequencies onto the same timeline. For video data, the timestamp of each frame is extracted; for sensor data, the timestamp of each sample is recorded. Based on a reference timeline, data from different sources are mapped to the closest available time point, forming aligned multi-source sensing data.

[0023] The pre-built deep learning network includes a video coding branch and a numerical coding branch. The video coding branch uses a 3D convolutional neural network structure to process construction operation video data. This branch first segments the input video into fixed-length segments, each containing 16 frames with a resolution of 224×224 pixels. Spatiotemporal features are extracted through 3D convolutional layers with a kernel size of 3×3×3 and a stride of 1×1×1. Each layer is followed by a batch normalization layer and a ReLU activation function. After processing through 5 convolutional blocks, the feature map is converted into a 1024-dimensional video feature vector through a global average pooling layer.

[0024] The numerical coding branch processes environmental monitoring data and equipment status data. This branch employs a multilayer perceptron architecture. Input data includes sensor-collected data such as temperature and humidity values, dust concentration values, noise levels, equipment vibration frequencies, and tilt angle values. The number of nodes in the input layer is related to the number of sensors; for example, when using 30 sensors, the input dimension is 30. The intermediate layers consist of two fully connected layers with 256 and 128 nodes respectively, each followed by a batch normalization layer and a ReLU activation function. The output layer generates a 256-dimensional numerical feature vector.

[0025] The fusion of video feature vectors and numerical feature vectors is based on spatial correlation calculation. The video feature vectors contain spatial location information, which can be used by object detection algorithms to identify the coordinates of workers, equipment, and materials in the video. The sensor location information in the numerical feature vectors includes the sensor's installation coordinates. The formula for calculating spatial correlation is: , where d ij Let σ represent the Euclidean distance between the i-th identified target and the j-th sensor in the video, where σ is the distance attenuation parameter with a value of 5 meters.

[0026] Based on the calculated spatial correlation matrix, the video feature vector and the numerical feature vector are weighted and fused. The fusion formula is: F = αV + βN, where F is the fused multidimensional feature vector, V is the video feature vector, N is the numerical feature vector, and α and β are the weight coefficients of the two feature vectors, determined by the spatial correlation matrix. When the target is close to the sensor in spatial location, the weight of the corresponding sensor data increases; conversely, it decreases. The final multidimensional feature vector has a dimension of 1280, containing comprehensive information from both video and sensor data.

[0027] Each element in the multidimensional feature vector has a specific physical meaning. The first 1024 dimensions reflect the spatiotemporal features of the video, including information such as worker behavior, equipment movement, and construction progress; the last 256 dimensions reflect environmental conditions and equipment operating parameters, including temperature, humidity, dust concentration, and equipment vibration frequency. The multidimensional feature vector will serve as input data for subsequent safety situation assessments.

[0028] This invention utilizes deep learning networks to extract and fuse features from multi-source sensor data, effectively combining visual information and sensor data from construction sites and overcoming the shortcomings of incomplete information from single data sources. The weighted fusion mechanism based on spatial correlation considers the spatial relationships between data sources, enhancing the accuracy of feature representation. The multi-dimensional feature vectors comprehensively reflect the state information of the construction site, providing a reliable data foundation for subsequent safety situation assessment and effectively improving the accuracy and real-time performance of safety situation awareness at construction sites.

[0029] Step 102: Identify the current construction stage based on the feature components in the multidimensional feature vector to obtain the work stage identifier.

[0030] In some embodiments of the present invention, step 102 may specifically include the following sub-steps: Sub-step 1021: Extract feature components from the multidimensional feature vector and divide the feature components into progress feature components and state feature components. Sub-step 1022: Calculate the feature similarity between the progress feature components and each work stage in the preset work stage mapping table, and select the work stage with the highest feature similarity as the candidate work stage. Sub-step 1023: Extract the duration of the candidate operation stage, calculate the duration of the current construction operation based on the state feature components, and compare the duration of the current operation with the duration of the candidate operation stage to obtain the duration deviation. Sub-step 1024: When the time deviation is less than the preset deviation threshold, the candidate operation stage is determined as the operation stage of the current construction operation; when the time deviation is greater than the preset deviation threshold and has lasted for a longer period than the stage duration, the next operation stage of the candidate operation stage is determined as the operation stage of the current construction operation. Sub-step 1025: Based on the determined current construction operation stage, obtain the corresponding operation stage identifier from the operation stage mapping table.

[0031] Multidimensional feature vectors contain rich information about the construction site. Feature components need to be extracted and classified from them. These feature components can be divided into two categories: progress feature components and status feature components. Progress feature components reflect the degree of completion of construction operations, including structural morphology features, material stacking features, and process completion features. Status feature components reflect the current state of construction operations, including the distribution of construction personnel, equipment operation features, and environmental change features.

[0032] When extracting feature components from a multidimensional feature vector, an eigenvalue decomposition method is employed. For a 1280-dimensional feature vector, the first 800 dimensions mainly contain progress-related information, while the last 480 dimensions mainly contain state-related information. Principal component analysis is used to extract the progress feature components, compressing the 800 dimensions of progress-related information into 100 main progress feature components, retaining more than 95% of the original information. An autoencoder structure is used to extract the state feature components, encoding the 480 dimensions of state-related information into 80 main state feature components. The extracted feature components are easier to calculate similarity and estimate duration.

[0033] The pre-defined work phase mapping table is based on construction specifications and experience, and contains standard feature templates for each work phase of a construction project. Taking concrete pouring as an example, work phases may include multiple stages such as formwork installation, rebar tying, concrete pouring, and curing. Each work phase corresponds to a set of standard feature components, describing the typical characteristics of that phase. The work phase mapping table also records the standard duration and phase identifier code for each phase.

[0034] The cosine similarity method is used to calculate the similarity between the progress feature components and the features of each work stage in the preset work stage mapping table. The formula for calculating cosine similarity is: Let A represent the currently extracted progress feature component, and B represent the standard feature component of the preset task stage. Calculate the feature similarity between the current progress feature component and each preset task stage; the closer the similarity value is to 1, the more similar the components. Select the task stage with the highest similarity as the candidate task stage.

[0035] After the candidate operation stages are determined, their correctness is further verified by extracting the stage duration corresponding to the candidate operation stages from the preset operation stage mapping table. Taking the concrete pouring stage as an example, the standard duration might be 6 hours. The duration of the current construction operation is calculated based on the state feature components. The calculation of the duration utilizes the temporal change characteristics in the state feature components. By detecting the change points of the state features, the start time of the current operation is determined, and then the duration is calculated.

[0036] The duration of the current work phase is compared with the duration of the candidate work phase to calculate the duration deviation, which is equal to the difference between the current duration and the stage duration. A preset deviation threshold is set at 20% of the stage duration. When the duration deviation is less than the preset deviation threshold, it indicates that the current work progress is in line with expectations, and the candidate work phase is determined as the current work phase. When the duration deviation is greater than the preset deviation threshold and the current duration exceeds the stage duration, it indicates that the current work has entered the next stage, and the next work phase of the candidate work phase is determined as the current work phase.

[0037] After determining the current construction stage, obtain the corresponding stage identifier. This stage identifier is a unique code used to represent different construction stages. Look up the identifier for the determined stage in the stage mapping table; for example, the identifier for the concrete pouring stage could be "S111". The stage identifier will be used for subsequent safety risk assessments and early warning generation.

[0038] The acquisition of job phase identifiers also takes into account special cases. When the calculated similarity scores are all below the set similarity threshold of 0.6, it indicates that the current job may be in a transitional or non-standard job state. In this case, the identifier is set to "Special Job State" with the identifier code "S000". For cases of abnormal duration, such as when the duration significantly exceeds 1.5 times the standard duration, an abnormal duration flag "T" is added after the identifier, indicating that it requires special attention.

[0039] This invention achieves accurate identification of construction operation stages through a triple mechanism of feature component extraction, similarity calculation, and duration verification. It not only considers construction progress characteristics but also incorporates time-dimensional information for verification, overcoming the misjudgment problems that may arise from relying solely on visual features. Accurate identification of operation stages provides a foundation for subsequent safety risk assessments, enabling targeted safety precautions to be taken based on the characteristics of different operation stages, thereby improving the accuracy and timeliness of safety situation awareness at the construction site.

[0040] Step 103: Extract the hazard source feature components corresponding to multiple hazard sources within a preset spatial range from the multidimensional feature vector, construct an interaction matrix based on the hazard source feature components, and perform coupling operation on the hazard source feature components and the interaction matrix to obtain the coupled risk factor.

[0041] In some embodiments of the present invention, step 103 may specifically include the following sub-steps: Sub-step 1031: Extract the hazard source feature components corresponding to multiple hazard sources located within a preset spatial range from the multi-dimensional feature vector; Sub-step 1032: Extract the type identifier and location information of each hazard source from the hazard source feature components, and filter out hazard source pairs with triggering relationships based on the type identifier and location information; Sub-step 1033: For hazard source pairs, mark the hazard source that initiates risk transmission as the transmission source and extract the feature components of the transmission source; mark the hazard source that receives risk transmission as the transmission target and extract the feature components of the transmission target; and construct a directed association graph based on the connection relationship between the transmission source and the transmission target. Sub-step 1034: Calculate the risk transmission coefficient based on the feature components of the transmission source and the feature components of the transmission target, fill the risk transmission coefficient into the connection position between the transmission source and the transmission target corresponding to the directed association graph, and generate the interaction matrix. Sub-step 1035: Combine multiple hazard source feature components into an initial risk vector according to the hazard source order in the directed association graph. Perform matrix multiplication iterative operation on the initial risk vector and the interaction matrix. Stop when the difference between two adjacent iterations is less than a preset difference threshold or the number of iterations reaches a preset upper limit, and obtain the steady-state risk vector. Sub-step 1036 involves weighted summation of the risk components corresponding to each hazard source in the steady-state risk vector to obtain the coupled risk factor.

[0042] The multidimensional feature vector contains information about various hazards at the construction site. It is necessary to extract hazard feature components located within a preset spatial range. This preset spatial range is determined based on the construction area and can be set as a circular area with a radius of 50 meters or a rectangular area with a side length of 80 meters, centered on the current point of interest. The extraction of hazard feature components employs a feature selection algorithm, extracting hazard-related feature subsets from the corresponding dimensions of the multidimensional feature vector.

[0043] The hazard source characteristic component includes the hazard source type identifier and location information. The type identifier uses codes to represent different types of hazard sources, such as "D01" for hoisting equipment, "D02" for working at heights, "D03" for openings near edges, "D04" for hot work, and "D05" for chemical storage. The location information includes the three-dimensional coordinates of the hazard source, measured with the origin set at the construction site as the reference point. Hazard source pairs with triggering relationships are screened based on the type identifier and location information. Screening conditions include: distance threshold condition (the distance between two hazard sources is less than the safe distance); type mutual exclusion condition (a certain type of hazard source should not appear in the adjacent area at the same time); and type association condition (certain types of hazard sources have potential triggering relationships).

[0044] For the selected pairs of hazards, risk transmission relationships are determined. In each pair of associated hazards, the hazard initiating risk transmission is marked as the transmission source, and the hazard receiving risk transmission is marked as the transmission target. Transmission relationship determination is based on hazard type and relative location; for example, a high-altitude work area constitutes a transmission source to the area below. When extracting feature components of the transmission source, attention is paid to features related to its risk release capability; when extracting feature components of the transmission target, attention is paid to features related to its risk reception sensitivity. Based on the connection relationships of all transmission sources and transmission targets, a directed association graph is constructed, where nodes represent hazard sources and directed edges represent the direction of risk transmission.

[0045] The risk transmission coefficient is calculated based on the characteristic components of the transmission source and the transmission target. The calculation formula is: R ij =α·E i ·S j ·D ij , where R ij E represents the risk transmission coefficient from hazard source i to hazard source j. i S represents the risk release intensity of hazard source i. j D represents the risk acceptance sensitivity of hazard source j. ij This represents the distance attenuation coefficient, where α is the adjustment coefficient. The formula for calculating the distance attenuation coefficient is: D ij =exp(-dij / d0), where d ij d0 is the distance between the two hazard sources.

[0046] The calculated risk transmission coefficients are then filled into the connection positions between the transmission sources and transmission targets in the directed association graph to generate an interaction matrix. This interaction matrix is ​​a square matrix with a dimension equal to the number of hazard sources, and matrix elements M. ij This represents the risk transmission coefficient from hazard i to hazard j. When there is no direct risk transmission between two hazard sources, the corresponding matrix element value is 0. The interaction matrix describes the complex mutual influence relationships between hazard sources and is the basis for subsequent risk coupling calculations.

[0047] Multiple hazard source feature components are arranged according to the hazard source order in the directed association graph to form an initial risk vector. Each element of the initial risk vector represents the inherent risk value of the corresponding hazard source, calculated using the formula: V i =β·P i ·C i V i P represents the inherent risk value of hazard source i. i C represents the probability of occurrence of hazard source i. i This indicates the severity of the risk consequences of hazard source i, and β is the normalization coefficient with a value of 0.5.

[0048] The initial risk vector and the interaction matrix are multiplied iteratively to simulate the risk transmission process in the hazard source network. The iterative calculation formula is: V (t+1) =V (0) +M·V (t) V (t) V represents the risk vector after the t-th iteration. (0) Let M represent the initial risk vector, and let M represent the interaction matrix. Iteration stops when the difference between two consecutive iterations is less than a preset difference threshold of 0.01 or the number of iterations reaches a preset upper limit of 20, yielding the steady-state risk vector. The difference is calculated using Euclidean distance. .

[0049] The coupled risk factor is obtained by weighted summation of the risk components corresponding to each hazard source in the steady-state risk vector. The weighted summation formula is as follows: Where RF represents the coupling risk factor, w represents the risk component of the i-th hazard source in the steady-state risk vector. i Let represent the weighting coefficient of the i-th hazard, and n represent the total number of hazard sources. The weighting coefficient is related to the type and importance of the hazard, with critical equipment and high-risk work areas having higher weights. The coupled risk factor reflects the comprehensive risk level after considering the interaction of hazard sources and is an important indicator for subsequent safety situation assessment.

[0050] This invention presents a hazard risk coupling method based on an interaction matrix, overcoming the limitations of traditional independent risk assessment of individual hazards. By constructing a directed correlation graph and performing matrix iterative operations, it systematically simulates the transmission and amplification of risks on construction sites. It fully considers the spatial relationships and type associations between hazards, achieving dynamic transmission and cumulative calculation of risks, making risk assessment more comprehensive and accurate. Coupled risk factors, as comprehensive indicators, can more realistically reflect the overall safety situation on construction sites, providing a more reliable basis for safety management decisions and improving the accuracy and timeliness of risk warnings at construction sites.

[0051] Step 104: Retrieve the corresponding risk propagation rule based on the work stage identifier, substitute the coupled risk factors into the risk propagation rule for calculation, and obtain the risk quantification value of the current work stage.

[0052] In some embodiments of the present invention, step 104 may specifically include the following sub-steps: Sub-step 1041: Based on the operation stage identifier, retrieve the corresponding risk propagation rule from the preset risk propagation rule base, and extract the risk propagation path from the risk propagation rule; Sub-step 1042: Collect the personnel density distribution and equipment distribution at each work location within the construction work area; construct a risk carrying capacity field along the risk propagation path based on the personnel density distribution and equipment distribution; and extract the local carrying capacity value corresponding to each work location from the risk carrying capacity field. Sub-step 1043: Substitute the coupled risk factor into the risk propagation rule, calculate the risk allocation weight of each work location based on the local carrying capacity value corresponding to each work location on the risk propagation path, and allocate the coupled risk factor to each work location according to the risk allocation weight to obtain the local risk component of each work location. Sub-step 1044: For the local risk components of each work location, determine whether the local risk components exceed the local carrying capacity value of the corresponding work location. If they exceed, extract the excess part as the overflow risk amount and pass it along the risk propagation path to the adjacent work location. Recalculate the local risk components of the work location that receives the overflow risk amount and iterate until no overflow risk amount is generated. When the iteration ends, accumulate the local risk components of each work location to obtain the risk quantification value of the current work stage.

[0053] The work phase identifier is a unique identifier for the stage of a construction operation. Different work phases face different risk characteristics and propagation patterns. A pre-set risk propagation rule base stores the risk propagation rules corresponding to each work phase. These rules describe how risks spread and affect the construction site.

[0054] Based on the acquired work stage identifier, the corresponding risk propagation rule is retrieved from a pre-defined risk propagation rule base. This rule base uses a key-value pair structure, with the work stage identifier as the key and the corresponding risk propagation rule as the value. Each risk propagation rule includes parameters such as the risk propagation path, propagation attenuation coefficient, and propagation threshold. The risk propagation path represents the direction and channel of risk propagation in the construction site space, expressed as a set of directed connections. For the concrete pouring stage, the risk propagation path may include multiple directions, such as propagation downwards from high-altitude work areas, propagation from machinery and equipment to surrounding areas, and propagation from chemical storage areas to usage areas. The extraction of the risk propagation path is achieved by parsing the path description data in the risk propagation rules, and the resulting path information is represented in a graph structure.

[0055] A risk-bearing capacity field is constructed by collecting personnel density distribution and equipment distribution data at various work locations within the construction area. Personnel density distribution is obtained by processing construction site monitoring images using a deep learning network, expressed as the number of personnel per square meter. Equipment distribution is obtained through an equipment positioning system and image recognition technology, recording the location and model of various equipment. Based on the personnel density distribution and equipment distribution, a risk-bearing capacity field is constructed along the risk propagation path. The risk-bearing capacity field is a spatial distribution function describing the risk-bearing capacity of each location on the construction site, calculated using the following formula: Where C(x,y) represents the risk carrying capacity at location (x,y), and D 人员 (x,y) represents the normalized value of the population density at location (x,y), S 设备 (x,y) represents the safety factor of the equipment at location (x,y), and γ1, γ2, and γ3 are the weight parameters for the corresponding items, with values ​​of 0.5, 0.3, and 0.2, respectively. The local bearing capacity value corresponding to each work location is extracted from the risk bearing capacity field. The construction site is divided into a 10m × 10m grid, with each grid corresponding to one work location. The average risk bearing capacity within that grid area is extracted as the local bearing capacity value.

[0056] By substituting the coupled risk factors into the risk propagation rule, the risk allocation weight for each work location is calculated based on its local carrying capacity value along the risk propagation path. This risk allocation weight represents the proportion of each work location in the overall risk allocation, taking into account the relative magnitude of the local carrying capacity value. The coupled risk factors are then allocated to each work location according to their risk allocation weights, yielding the local risk component for each work location. This local risk component represents the amount of risk allocated to a specific work location and serves as the basis for subsequent risk spillover assessments.

[0057] For each work location's local risk component, determine whether it exceeds the corresponding local carrying capacity value. When the local risk component exceeds the local carrying capacity value, it indicates that the risk level at that location has exceeded its tolerance, resulting in spillover risk that will be transmitted to adjacent locations. The spillover risk amount is equal to the difference between the local risk component and the local carrying capacity value. The spillover risk amount is transmitted to adjacent work locations along the risk propagation path, taking into account risk attenuation and path weights during the transmission process. The risk attenuation coefficient represents the degree of risk attenuation during propagation, is related to the propagation distance and propagation medium, and ranges from 0 to 1. The path weight represents the relative importance of different propagation paths and is affected by spatial layout and obstacles. For work locations receiving the spillover risk amount, the local risk component is recalculated. The updated local risk component is the original local risk component plus the sum of the risk amount transmitted from adjacent locations.

[0058] The iterative judgment process continues until no overflow risk occurs, meaning that the local risk components at all work locations do not exceed their corresponding local carrying capacity values, or the preset maximum number of iterations is reached. In practical applications, the maximum number of iterations is usually set to 10 to 15 to balance computational accuracy and efficiency. At the end of the iteration, the local risk components at each work location represent the spatial distribution of risk. The local risk components at each work location at the end of the iteration are summed to obtain the quantified risk value for the current work stage. The quantified risk value is a dimensionless numerical value representing the overall risk level of the current work stage; a larger value indicates a higher risk.

[0059] During the calculation process, the risk change trends and spillover risk transmission paths at each work location can be recorded. This information helps identify key risk areas and risk hotspots. For work phases with particularly high risk quantification values, the risk composition and main sources can be further analyzed to provide targeted recommendations for risk control. When construction conditions change, such as adjustments to personnel allocation, changes to equipment layout, or modifications to work content, the risk quantification values ​​need to be recalculated to reflect the latest safety situation. The calculation results of the risk quantification values ​​can be compared with historical data to assess the relative level of the current safety situation and predict risk development trends.

[0060] Risk quantification calculations also consider the impact of time. For long-duration operation phases, they are divided into multiple time windows, and the risk quantification value for each time window is calculated separately. The maximum value is then taken as the final risk quantification value for that operation phase. The time window division is based on changes in the operation content and fluctuations in risk factors, typically 4 or 8 hours. This time window division method can capture risk fluctuations during the operation process and avoid the risk being averaged out, which would mask high-risk periods.

[0061] This invention utilizes a risk quantification method based on risk propagation rules and risk-bearing capacity fields to accurately simulate the dynamic propagation and distribution of risks at construction sites. The quantified risk value, as a comprehensive indicator, not only reflects the risk level of the hazard itself but also considers the spatial propagation and cumulative effects of risk, as well as the differences in risk-bearing capacity among different areas of the site. This dynamic risk assessment method improves the accuracy and timeliness of safety situation awareness at construction sites, providing more comprehensive and reliable data support for safety management decisions.

[0062] Step 105: Determine the corresponding stage evolution sequence based on the work stage identifier, calculate the risk quantification value of the next work stage based on the risk quantification value of the current work stage and the risk propagation rule of the next work stage in the stage evolution sequence, and generate the risk evolution sequence.

[0063] In some embodiments of the present invention, step 105 may specifically include the following sub-steps: Sub-step 1051: Query the corresponding stage evolution sequence based on the work stage identifier, and extract the subsequent work stages and corresponding risk propagation rules from the stage evolution sequence. Sub-step 1052: Based on the risk propagation rules of the current operation stage, decompose the risk quantification value of the current operation stage into the flow risk value and the inherent risk value; Sub-step 1053: Calculate the environmental difference between the current operation stage and the next operation stage, obtain the stage transition decay rate based on the environmental difference, and perform a decay calculation on the flow risk value and the stage transition decay rate to obtain the transmission risk value. Sub-step 1054: Extract the hazard trigger threshold and risk amplification coefficient from the risk propagation rules of the next operation stage. When the transmitted risk value reaches the hazard trigger threshold, calculate the amplified transmitted risk value based on the transmitted risk value and the risk amplification coefficient; otherwise, use the transmitted risk value as the amplified transmitted risk value. Sub-step 1055: Extract the basic risk value of the stage from the risk propagation rules of the next operation stage, and superimpose the amplified transmission risk value, the basic risk value of the stage, and the inherent risk value to obtain the risk quantification value of the next operation stage. Sub-step 1056: Repeat the iteration until all subsequent operation stages have been traversed, and arrange the risk quantification values ​​of each operation stage in order to generate a risk evolution sequence.

[0064] Querying the stage evolution sequence based on the work stage identifier requires extracting relevant information from the construction plan database. The stage evolution sequence is stored in the form of a directed linked list, where each node contains the work stage identifier and the corresponding risk propagation rule. Extracting subsequent work stages and their corresponding risk propagation rules from the stage evolution sequence prepares for subsequent risk quantification calculations. For concrete structure construction, the stage evolution sequence might be: rebar tying → formwork installation → concrete pouring → curing → formwork removal → quality inspection.

[0065] Based on the risk propagation rules of the current operational phase, the risk quantification value of the current operational phase is decomposed into a flowing risk value and an inherent risk value. The flowing risk value represents the portion of risk that may be transmitted as the operational phase changes, while the inherent risk value represents the portion of risk that remains unchanged as the operational phase changes. The decomposition calculation formula is as follows: RV 当前 α represents the quantified risk value of the current operational phase. 流动 This represents the liquidity risk ratio coefficient, ranging from 0 to 1, and is determined according to risk propagation rules. For operation phases primarily involving high-altitude work, α... 流动 It may take a relatively high value, such as 0.7; for the operation phase that is mainly ground-based, α 流动 It may take a low value, such as 0.3.

[0066] Calculating the environmental difference between the current work phase and the next work phase is crucial for determining the degree of risk transmission. Environmental difference considers factors such as workspace, equipment configuration, and personnel composition. The formula for calculating environmental difference is: D 环境 =k1×D 空间 +k2×D 设备 +k3×D 人员 D 环境 D represents the degree of environmental variability. 空间 D 设备 D 人员 These represent spatial variability, equipment variability, and personnel variability, respectively. k1, k2, and k3 are weighting coefficients with values ​​of 0.4, 0.35, and 0.25, respectively. The phase conversion attenuation rate is obtained based on environmental variability, calculated using the following formula: , where λ 转换 RV represents the phase transition decay rate, and β is an adjustment coefficient with a value of 0.8. The transferred risk value is obtained by performing a decay calculation on the liquidity risk value and the phase transition decay rate. The calculation formula is: RV 传递 =RV 流动 ×λ 转换 RV 传递 This indicates the transmission of risk values.

[0067] The hazard trigger threshold and risk amplification factor are extracted from the risk propagation rules of the next operational phase. The hazard trigger threshold represents the minimum risk value that can trigger a risk amplification effect, and the risk amplification factor represents the risk amplification multiple after triggering. When the transmitted risk value reaches the hazard trigger threshold, the amplified transmitted risk value is calculated based on the transmitted risk value and the risk amplification factor. The calculation formula is as follows: RV 放大传递 Indicates amplified risk value, RV 触发阈值 η represents the hazard triggering threshold, and η represents the risk amplification factor, ranging from 0.5 to 2. When the transmitted risk value does not reach the hazard triggering threshold, the transmitted risk value is used as the amplified transmitted risk value, i.e., RV. 放大传递 =RV 传递 .

[0068] The basic risk value for the next operational phase is extracted from the risk propagation rules. This basic risk value represents the inherent basic risk level of the current operational phase and is unrelated to the risk transmitted from the previous phase. The amplified risk value, the basic risk value, and the inherent risk value are added together to obtain the quantified risk value for the next operational phase.

[0069] Repeat the iterative calculation process described above, sequentially calculating the risk quantification value for each subsequent work stage in the stage evolution sequence. In each iteration, the risk quantification value calculated in the previous stage is used as the risk quantification value for the current stage. It is then decomposed into flowing risk and inherent risk values ​​using the same method. The environmental difference and stage transition attenuation rate with the next stage are calculated to obtain the transmitted risk value and amplified transmitted risk value. Finally, the risk quantification value for the next stage is calculated. This iterative process continues until all subsequent work stages have been traversed. The risk quantification values ​​for each work stage are arranged in the order of the stage evolution sequence to generate a risk evolution sequence. The risk evolution sequence can be represented as a time series, with the horizontal axis representing the work stage and the vertical axis representing the risk quantification value, visually displaying the changing trend of risk levels during construction.

[0070] Risk evolution sequences can predict the risk level of future work phases and identify risk peaks and critical risk inflection points. For predicted risk peaks in the risk evolution sequence, construction plans can be adjusted in advance, resource allocation optimized, and targeted risk control measures implemented. For critical risk inflection points, focused monitoring can be deployed to prevent risk accumulation and amplification. The risk evolution sequence is generated using a rolling prediction method; as construction progresses, information on the current work phase is continuously updated, and the risk evolution trend for subsequent phases is recalculated, ensuring the timeliness and accuracy of the prediction results.

[0071] This invention, based on a risk prediction method using a phased evolution sequence, enables proactive perception of construction safety status, breaking through the limitations of passive response in traditional safety management. The risk evolution sequence, as a crucial output of safety status perception, provides construction managers with a clear prediction of future risk trends, supporting the development of forward-looking safety management plans, optimization of resource allocation, and early deployment of risk prevention and control measures. This significantly improves the initiative and foresight of on-site safety management, effectively reducing the probability of safety accidents.

[0072] Step 106: Traverse the risk evolution sequence. When the risk quantification value in the risk evolution sequence exceeds the safety threshold of the corresponding operation stage, extract the moment when the safety threshold is exceeded and generate an early warning signal.

[0073] In some embodiments of the present invention, step 106 may specifically include the following sub-steps: Sub-step 1061: Traverse the risk quantification values ​​of each operation stage in the risk evolution sequence in chronological order, obtain the safety threshold of each operation stage, mark the corresponding operation stage as the over-limit operation stage when the risk quantification value exceeds the safety threshold, and extract the time corresponding to the over-limit operation stage as the time when the safety threshold is exceeded. Sub-step 1062: Extract the preceding and subsequent operation stages in the risk evolution sequence of the over-limit operation stage, calculate the preceding difference between the risk quantification value of the over-limit operation stage and the risk quantification value of the preceding operation stage, calculate the subsequent difference between the risk quantification value of the subsequent operation stage and the risk quantification value of the over-limit operation stage, and determine the risk evolution trend of the over-limit operation stage based on the preceding and subsequent differences. Sub-step 1063: Based on the risk evolution trend, determine whether the risk quantification value of the over-limit operation stage is in a continuous upward state. When it is in a continuous upward state, the time when the safety threshold is exceeded and the time of the subsequent operation stage are used as the warning time range and a warning signal is generated. When it is not in a continuous upward state, the time when the safety threshold is exceeded is used as the warning time range and a warning signal is generated.

[0074] When traversing the risk evolution sequence, the risk quantification value of each work stage must be checked sequentially according to time. A safety threshold refers to the maximum permissible risk level at each work stage; exceeding this threshold may lead to a safety accident. Safety thresholds can be obtained from a construction specification database or trained using a deep learning network based on historical accident data. The safety thresholds differ for different work stages, with lower thresholds for high-risk stages and higher thresholds for low-risk stages. The setting of safety thresholds considers various factors such as work type, environmental conditions, and personnel allocation.

[0075] The system iterates through the risk evolution sequence, traversing the risk quantification values ​​of each work stage. The currently traversed work stage is designated as the current work stage, and its safety threshold is retrieved from a pre-set threshold database. The risk quantification value of the current work stage is compared with the safety threshold to determine if it exceeds the threshold. If the risk quantification value exceeds the safety threshold, the current work stage is marked as an over-limit work stage, and the corresponding time is extracted as the moment the safety threshold is exceeded. This time information can be a specific date or time in the construction plan, or a time interval relative to the construction start time. For example, in the concrete pouring work stage, if the risk quantification value is 85 and the safety threshold is 80, this work stage is marked as an over-limit work stage, and the corresponding time might be the 15th day after construction begins.

[0076] Extract the preceding and subsequent operation stages from the risk evolution sequence of the over-limit operation stage. The preceding operation stage refers to the operation stage preceding the over-limit operation stage in the risk evolution sequence, and the subsequent operation stage refers to the operation stage following the over-limit operation stage. Calculate the preceding and subsequent differences: Preceding difference = Over-limit operation stage risk quantification value - Preceding operation stage risk quantification value; Subsequent difference = Subsequent operation stage risk quantification value - Over-limit operation stage risk quantification value. Determine the risk evolution trend of the over-limit operation stage based on the preceding and subsequent differences. When both the preceding and subsequent differences are greater than 0, it indicates that the risk quantification value is in a continuously rising state; when both are less than 0, it indicates that the risk quantification value is in a state of first rising and then falling; when both are less than 0, it indicates that the risk quantification value is in a state of first falling and then rising; when both are less than 0, it indicates that the risk quantification value is in a state of continuously falling.

[0077] The risk quantification value during the over-limit operation phase is determined based on the risk evolution trend. A continuous upward trend refers to a situation where the risk quantification value shows an increasing trend starting from the preceding operation phase and continues to increase after reaching its peak during the over-limit operation phase. The criteria for a continuous upward trend are: the difference between the preceding and subsequent operations is greater than 0, or the difference between the preceding and subsequent operations is greater than 0 and the absolute value of the difference between the subsequent operations is less than 20% of the difference between the preceding and subsequent operations. When the over-limit operation phase is in a continuous upward trend, the risk situation will further deteriorate, requiring a wider range of warnings. In this case, the moment exceeding the safety threshold and the moment of the subsequent operation phase are used as the warning time range, and a warning signal is generated. The warning time range is expressed as a time interval, starting at the moment exceeding the safety threshold and ending at the end of the subsequent operation phase. When the over-limit operation phase is not in a continuous upward trend, the risk situation is relatively stable or has eased somewhat, and only the moment exceeding the safety threshold is used as the warning time range, and a warning signal is generated. The warning signal includes the warning time range, the over-limit operation phase identifier, the risk quantification value, the safety threshold, and information on the risk evolution trend.

[0078] The generation of early warning signals also considers the degree of risk exceeding limits, which is defined as the ratio of the quantified risk value to the safety threshold. The calculation formula is: Degree of Exceeding Limits = Quantified Risk Value / Safety Threshold. Early warning signals are divided into three levels based on the degree of exceeding limits: mild, moderate, and severe. A mild warning is defined as an exceedance level < 1.2; a moderate warning is defined as an exceedance level ≥ 1.2 and < 1.5; and a severe warning is defined as an exceedance level ≥ 1.5. Different notification methods and processing procedures are used for different levels of early warning signals. Mild warnings are notified to relevant personnel through internal system notifications; moderate warnings are notified to project managers and safety management personnel via SMS or application push notifications; and severe warnings are notified to project managers, safety management personnel, and higher-level management departments simultaneously via telephone and SMS. The early warning signal is generated an advance notice period before the expected start time of the exceedance operation phase, and this advance notice period is related to the warning level. Mild warnings are generated 3 days in advance, moderate warnings 5 ​​days in advance, and severe warnings 7 days in advance.

[0079] For generated early warning signals, early warning tracking is implemented. Early warning tracking involves continuously monitoring changes in the actual quantified risk value within the warning timeframe and comparing it with the predicted quantified risk value. When the deviation between the actual and predicted risk values ​​exceeds 20%, the risk evolution sequence is recalculated and the early warning signal is updated. The frequency of early warning tracking is related to the early warning level: minor warnings are tracked daily, moderate warnings every 12 hours, and severe warnings every 6 hours. Early warning tracking ensures the accuracy and timeliness of early warning signals, avoiding warning failures or false alarms due to prediction deviations.

[0080] This invention presents a risk warning mechanism based on safety thresholds, enabling proactive early warning of construction safety risks and transforming traditional reactive measures into proactive prevention. By analyzing the relationship between quantified risk values ​​and safety thresholds in the risk evolution sequence, potential high-risk periods are identified. Combined with risk evolution trend analysis, the warning scope is determined, generating warning signals of different levels. This warning mechanism not only considers the static judgment of risk exceeding limits but also incorporates the dynamic analysis of risk change trends, providing construction managers with more comprehensive and accurate risk prediction information. This supports the early deployment of prevention and control measures, adjustment of construction plans, and optimization of resource allocation, thereby effectively preventing safety accidents and ensuring safe production at construction sites.

[0081] like Figure 2 As shown, Figure 2 This is a schematic diagram of a construction site safety situation awareness system based on deep learning networks provided in an embodiment of the present invention. The system includes: The data processing module 201 is used to acquire multi-source sensing data from the construction site and extract features from the multi-source sensing data through a pre-built deep learning network to obtain multi-dimensional feature vectors. Stage identification module 202 is used to identify the current construction stage based on the feature components in the multi-dimensional feature vector and obtain the work stage identifier. The coupling calculation module 203 is used to extract the hazard source feature components corresponding to multiple hazard sources located within a preset spatial range in the multidimensional feature vector, construct an interaction matrix based on the hazard source feature components, and perform coupling calculation on the hazard source feature components and the interaction matrix to obtain the coupling risk factor. The risk quantification module 204 is used to retrieve the corresponding risk propagation rule according to the work stage identifier, substitute the coupled risk factors into the risk propagation rule for calculation, and obtain the risk quantification value of the current work stage. The evolution prediction module 205 is used to determine the corresponding stage evolution sequence based on the work stage identifier, calculate the risk quantification value of the next work stage based on the risk quantification value of the current work stage and the risk propagation rule of the next work stage in the stage evolution sequence, and generate a risk evolution sequence. The early warning generation module 206 is used to traverse the risk evolution sequence. When the risk quantification value in the risk evolution sequence exceeds the safety threshold of the corresponding operation stage, the module extracts the moment when the safety threshold is exceeded and generates an early warning signal.

[0082] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0083] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0084] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for safety situation awareness at construction sites based on deep learning networks, characterized in that, Includes the following steps: Acquire multi-source sensing data from the construction site, and extract features from the multi-source sensing data using a pre-constructed deep learning network to obtain multi-dimensional feature vectors; The current construction operation stage is identified based on the feature components in the multidimensional feature vector, thus obtaining the operation stage identifier. Extract the hazard source feature components corresponding to multiple hazard sources within a preset spatial range from the multidimensional feature vector, construct an interaction matrix based on the hazard source feature components, and perform coupling operations on the hazard source feature components and the interaction matrix to obtain the coupled risk factor; Based on the work phase identifier, the corresponding risk propagation rule is retrieved, and the coupled risk factors are substituted into the risk propagation rule for calculation to obtain the risk quantification value of the current work phase. The corresponding stage evolution sequence is determined based on the work stage identifier. Based on the risk quantification value of the current work stage and the risk propagation rule of the next work stage in the stage evolution sequence, the risk quantification value of the next work stage is calculated, and the risk evolution sequence is generated. Traverse the risk evolution sequence. When the risk quantification value in the risk evolution sequence exceeds the safety threshold of the corresponding operation stage, extract the moment when the safety threshold is exceeded and generate an early warning signal.

2. The method according to claim 1, characterized in that, Multi-source sensing data from the construction site is acquired, and features are extracted from the multi-source sensing data using a pre-constructed deep learning network to obtain a multi-dimensional feature vector, including: Construction operation video data is collected by video acquisition equipment deployed at the construction site, and environmental monitoring data and equipment status data are collected by sensor equipment deployed at the construction site. The construction operation video data, environmental monitoring data and equipment status data are time-stamped and aligned to obtain multi-source sensing data. Construction operation video data from multi-source sensing data is input into a pre-constructed deep learning network, which includes a video coding branch and a numerical coding branch. Video feature vectors are obtained by feature extraction through the video coding branch, and numerical feature vectors are obtained by feature encoding through the numerical coding branch. The spatial correlation between the spatial location information in the video feature vector and the sensor location information in the numerical feature vector is calculated. Based on the spatial correlation, the video feature vector and the numerical feature vector are weighted and fused to obtain a multidimensional feature vector.

3. The method according to claim 1, characterized in that, Based on the feature components in the multidimensional feature vector, the current construction operation stage is identified, resulting in the following operation stage identifiers: Feature components are extracted from the multidimensional feature vector, and the feature components are divided into progress feature components and state feature components. Calculate the feature similarity between the progress feature components and each work stage in the preset work stage mapping table, and select the work stage with the highest feature similarity as the candidate work stage; Extract the duration of the candidate operation stage, calculate the duration of the current construction operation based on the state feature components, and compare the duration of the current operation with the duration of the candidate operation stage to obtain the duration deviation. When the time deviation is less than the preset deviation threshold, the candidate operation stage is determined as the current operation stage. When the time deviation is greater than the preset deviation threshold and has lasted for a duration exceeding the stage duration, the next operation stage of the candidate operation stage is determined as the current operation stage. Based on the current work stage, the corresponding work stage identifier is obtained from the work stage mapping table.

4. The method according to claim 1, characterized in that, Extract hazard source feature components corresponding to multiple hazard sources within a preset spatial range from the multidimensional feature vector. Construct an interaction matrix based on the hazard source feature components. Perform coupling operations on the hazard source feature components and the interaction matrix to obtain coupled risk factors, including: Extract the hazard source feature components corresponding to multiple hazard sources located within a preset spatial range from the multidimensional feature vector; Extract the type identifier and location information of each hazard source from the hazard source feature components, and filter out hazard source pairs with triggering relationships based on the type identifier and location information; For a pair of hazard sources, the hazard source that initiates the risk transmission is marked as the transmission source and the feature components of the transmission source are extracted. The hazard source that receives the risk transmission is marked as the transmission target and the feature components of the transmission target are extracted. A directed association graph is constructed based on the connection relationship between the transmission source and the transmission target. The risk transmission coefficient is calculated based on the feature components of the transmission source and the feature components of the transmission target. The risk transmission coefficient is then filled into the connection position between the transmission source and the transmission target in the directed association graph to generate the interaction matrix. Multiple hazard source feature components are arranged in the order of hazard sources in the directed association graph to form an initial risk vector. The initial risk vector is then multiplied by the interaction matrix in an iterative matrix operation. The operation stops when the difference between two adjacent iterations is less than a preset difference threshold or the number of iterations reaches a preset upper limit, thus obtaining a steady-state risk vector. The coupled risk factor is obtained by weighted summation of the risk components corresponding to each hazard source in the steady-state risk vector.

5. The method according to claim 1, characterized in that, Based on the work phase identifier, the corresponding risk propagation rule is retrieved. The coupled risk factors are substituted into the risk propagation rule for calculation, and the risk quantification value of the current work phase is obtained, including: Based on the operation stage identifier, retrieve the corresponding risk propagation rule from the preset risk propagation rule base, and extract the risk propagation path from the risk propagation rule; Collect the personnel density distribution and equipment distribution at each work location within the construction work area, construct a risk carrying capacity field along the risk propagation path based on the personnel density distribution and equipment distribution, and extract the local carrying capacity value corresponding to each work location from the risk carrying capacity field; By substituting the coupled risk factors into the risk propagation rule, the risk allocation weight of each work location is calculated based on the local carrying capacity value corresponding to each work location on the risk propagation path. The coupled risk factors are then allocated to each work location according to the risk allocation weight to obtain the local risk component of each work location. For the local risk components of each work location, it is determined whether the local risk components exceed the local carrying capacity value of the corresponding work location. If they do, the excess portion is extracted as the overflow risk quantity and transmitted to the adjacent work location along the risk propagation path. The local risk components of the work location receiving the overflow risk quantity are recalculated and iteratively determined until no overflow risk quantity is generated. The local risk components of each work location at the end of the iteration are accumulated to obtain the risk quantification value of the current work stage.

6. The method according to claim 1, characterized in that, Based on the work stage identifier, the corresponding stage evolution sequence is determined. Based on the risk quantification value of the current work stage and the risk propagation rule for the next work stage in the stage evolution sequence, the risk quantification value for the next work stage is calculated, generating the risk evolution sequence, which includes: Based on the work stage identifier, query the corresponding stage evolution sequence and extract the subsequent work stages and corresponding risk propagation rules in the stage evolution sequence. Based on the risk propagation rules of the current operation phase, the risk quantification value of the current operation phase is decomposed into a current risk value and an inherent risk value. Calculate the environmental difference between the current operation stage and the next operation stage, obtain the stage transition decay rate based on the environmental difference, and perform a decay calculation on the flow risk value and the stage transition decay rate to obtain the transmission risk value. Extract the hazard trigger threshold and risk amplification coefficient from the risk propagation rules of the next operation phase. When the transmitted risk value reaches the hazard trigger threshold, the amplified transmitted risk value is calculated based on the transmitted risk value and the risk amplification coefficient. Otherwise, the transmitted risk value is used as the amplified transmitted risk value. Extract the basic risk value of the next operation stage from the risk propagation rules of the next operation stage, and then combine the amplified transmission risk value, the basic risk value of the stage, and the inherent risk value to obtain the risk quantification value of the next operation stage. Repeat the iteration until all subsequent operation stages have been traversed, and arrange the risk quantification values ​​of each operation stage in order to generate a risk evolution sequence.

7. The method according to claim 1, characterized in that, Traverse the risk evolution sequence. When the risk quantification value in the risk evolution sequence exceeds the safety threshold of the corresponding operation stage, extract the moment when the safety threshold is exceeded and generate an early warning signal, including: The risk quantification value of each operation stage in the risk evolution sequence is traversed in chronological order to obtain the safety threshold of each operation stage. When the risk quantification value exceeds the safety threshold, the corresponding operation stage is marked as the over-limit operation stage, and the time corresponding to the over-limit operation stage is extracted as the time when the safety threshold is exceeded. Extract the preceding and subsequent operation stages in the risk evolution sequence of the over-limit operation stage, calculate the preceding difference between the risk quantification value of the over-limit operation stage and the risk quantification value of the preceding operation stage, calculate the subsequent difference between the risk quantification value of the subsequent operation stage and the risk quantification value of the over-limit operation stage, and determine the risk evolution trend of the over-limit operation stage based on the preceding and subsequent differences. Based on the risk evolution trend, determine whether the risk quantification value of the over-limit operation stage is in a continuous upward state. When it is in a continuous upward state, the moment when the safety threshold is exceeded and the moment of the subsequent operation stage will be used as the warning time range and a warning signal will be generated. When it is not in a continuous upward state, the moment when the safety threshold is exceeded will be used as the warning time range and a warning signal will be generated.

8. A construction site safety situation awareness system based on deep learning networks, used to implement the method described in any one of claims 1-7, characterized in that, The system includes: The data processing module is used to acquire multi-source sensing data from the construction site and extract features from the multi-source sensing data through a pre-built deep learning network to obtain multi-dimensional feature vectors. The stage identification module is used to identify the current construction stage based on the feature components in the multi-dimensional feature vector, and obtain the work stage identifier. The coupling calculation module is used to extract the hazard source feature components corresponding to multiple hazard sources within a preset spatial range in the multidimensional feature vector, construct an interaction matrix based on the hazard source feature components, and perform coupling calculations on the hazard source feature components and the interaction matrix to obtain the coupling risk factor. The risk quantification module is used to retrieve the corresponding risk propagation rule based on the work stage identifier, substitute the coupled risk factors into the risk propagation rule for calculation, and obtain the risk quantification value of the current work stage. The evolution prediction module is used to determine the corresponding stage evolution sequence based on the operation stage identifier, calculate the risk quantification value of the next operation stage based on the risk quantification value of the current operation stage and the risk propagation rule of the next operation stage in the stage evolution sequence, and generate the risk evolution sequence. The early warning generation module is used to traverse the risk evolution sequence. When the risk quantification value in the risk evolution sequence exceeds the safety threshold of the corresponding operation stage, the module extracts the moment when the safety threshold is exceeded and generates an early warning signal.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7.

Images