A construction engineering safety risk intelligent early warning and management and control method

Abstract

The application relates to the technical field of construction safety management, and discloses a building engineering safety risk intelligent early warning and management and control method, which comprises the following steps: firstly, a three-dimensional digital scene containing static background and dynamic entity and a cause-effect graph representing a risk cause-effect relationship are constructed; secondly, a risk chain activated is identified by using entity state prediction, and risk saliency is calculated; thirdly, when the risk exceeds a threshold value, an alternative intervention instruction set is generated, and parallel counterfactual simulation is performed on each instruction to calculate global risk after intervention; fourthly, an integrated utility score is calculated by combining intervention cost and global risk, and an optimal management and control instruction is selected and issued for execution; and finally, factual data after instruction execution is collected to calculate prediction deviation. The application quantitatively evaluates management and control effect through counterfactual deduction, continuously optimizes model parameters based on execution feedback, effectively solves the problems of risk coupling difficulty in prediction and poor model adaptability in a dynamic construction environment, and balances safety and efficiency.

Description

Technical Field

[0001] This invention relates to the field of construction safety management technology, specifically to an intelligent early warning and control method for safety risks in building engineering. Background Technology

[0002] With the digital transformation of the construction industry, various sensing devices and monitoring systems are widely used in safety management at construction sites. Current engineering safety management mainly relies on video surveillance, sensor monitoring, and deep learning-based target detection technology to identify obvious violations such as not wearing safety helmets and intrusion into dangerous areas.

[0003] However, existing safety early warning technologies primarily rely on statistical correlation or static rule matching for judgment. While these methods can identify current hazardous states, they lack the ability to causally infer the evolutionary mechanisms of risk events. When faced with complex human-machine collaboration and a constantly changing environment at construction sites, correlation analysis alone is insufficient to accurately predict the actual effectiveness of intervention measures. For example, when a system detects a vehicle collision risk, traditional prediction models often fail to accurately distinguish between the different consequences of "emergency braking" and "evasive steering" under specific working conditions, resulting in control instructions lacking specificity and effectiveness.

[0004] Furthermore, construction sites are unstructured environments in a constant state of dynamic change, with work surfaces, road conditions, and equipment performance all evolving as the project progresses. Most existing early warning models rely on offline training or fixed parameter settings, making them ill-suited to this time-varying nature. When the site environment changes, the models often fail to promptly correct their judgment logic, leading to increased false alarm rates or missed hazard warnings. To avoid potential liability, existing control strategies tend to adopt conservative, indiscriminate work stoppages or widespread alarms. While this approach ensures safety to some extent, it severely disrupts normal construction progress and fails to achieve synergistic optimization of safety and production efficiency. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent early warning and control method for safety risks in construction engineering. It solves the problem that most existing early warning models rely on offline training or fixed parameter settings, making them ill-suited to time-varying characteristics. Furthermore, when the on-site environment changes, the models often fail to promptly correct their judgment logic, leading to increased false alarm rates or missed hazard warnings.

[0006] To achieve the above objectives, this invention provides the following technical solution: an intelligent early warning and control method for safety risks in construction engineering, comprising first constructing a three-dimensional digital scene containing a static background and dynamic entities, and establishing a causal graph representing the causal relationships of risk events. The three-dimensional digital scene, as a digital mapping of physical space, provides the foundation for spatial and state data; the causal graph, based on a Bayesian network structure, explicitly expresses the probabilistic dependencies between risk factors, intermediate states, and accident consequences through nodes and directed edges.

[0007] Based on this, the system predicts the state of dynamic entities in a 3D digital scene, generating future state trajectories. The predicted future state data is then input into a causal graph for matching, identifying activated risk chains. By extrapolating along the causal path and combining conditional probability parameters, the system calculates the probability of risk events occurring and correlates them with potential losses, thereby quantifying the risk salience.

[0008] When the detected risk significance exceeds a preset threshold, the system dynamically generates a set of alternative intervention instructions. For each alternative intervention instruction in this set, the system performs a parallel counterfactual simulation. This simulation is conducted in an isolated copy of the causal state space. By forcibly setting the value of the intervention variable and blocking the causal influence of its parent node, the system deduces the system evolution state under the assumption of applying the intervention measure, and then calculates the sum of global risk significance after the instruction is applied. This step achieves a preliminary quantitative assessment of the effectiveness of control measures.

[0009] To achieve optimal decision-making, the system calculates a comprehensive utility score by combining the intervention costs of each alternative intervention command with the sum of the significance of the overall risk. The intervention cost is used to quantify the impact of command execution on construction progress or resource consumption. The system determines the optimal control command by solving a problem that minimizes the comprehensive utility score and then issues it to the terminal for execution.

[0010] After the command is executed, this method introduces a closed-loop correction mechanism. It collects factual data after the execution of the optimal control command and calculates the prediction deviation between the actually observed risk state and the counterfactual simulation prediction state. Based on this deviation, the parameters of the causal graph are adaptively updated, enabling the model to continuously fit the causal laws of the real environment as the construction progresses.

[0011] In one optional implementation, the process of identifying activated risk chains and calculating risk significance is as follows: Based on the state prediction results of dynamic entities, the system identifies nodes with abnormal states in the causal graph as activated risk factor nodes. Subsequently, the algorithm starts from this node and searches downstream along directed edges until it reaches the defined risk event node, forming a complete risk chain. The calculation of risk significance is based on the product of the conditional probability parameters of each node in the chain, and the impact of the potential loss of the risk event is weighted in, thereby achieving a quantitative measurement of the degree of risk.

[0012] In one alternative implementation, the parallel counterfactual simulation employs intervention logic from causal inference. Specifically, for each alternative intervention instruction, the system creates an independent copy of the causal state space. In this copy, the system forcibly sets the variable node corresponding to the instruction to the instruction's target value, while simultaneously severing the connection between this node and its original parent node to eliminate the influence of confounding factors. Probability propagation is then performed again on this basis, and the sum of residual risks is calculated. This mechanism distinguishes between correlation and causality, reducing false alarms caused by spurious correlations.

[0013] In one alternative implementation, a cost function is introduced into the calculation of the overall utility score. This cost function maps instructions to quantified intervention costs based on the atomic operation type and specific parameters of the instruction. For instructions involving speed adjustment, the intervention cost is proportional to the speed adjustment magnitude to reflect the disturbance to equipment wear and subsequent processes caused by abrupt deceleration. The system uses preset risk weights and cost weights to perform a weighted summation of the total global risk saliency and intervention cost, and selects the optimal instruction by minimizing this objective function.

[0014] In one optional implementation, the closed-loop correction of the causal graph employs a gradient-based parameter update strategy. The system first recalculates the significance of factual risk based on the post-execution factual data, and then calculates the difference between this recalculation and the simulated prediction value to obtain the prediction bias. Subsequently, the conditional probability parameters related to risk in the causal graph were analyzed using the following formula. Update: ; in, and These are the conditional probability parameters before and after the update, respectively; The preset learning rate is used to control the update step size; Gradient of risk significance with respect to the conditional probability parameter Furthermore, to reduce computational complexity and improve the relevance of updates, gradient calculation is implemented through backward tracing. Starting from the risk event node, the system traverses backward along the causal graph, reversing all causal paths that lead to the activation of that risk. The algorithm only calculates non-zero gradients for the conditional probability parameters of edges located on these activation paths, while setting the gradients of path parameters not involved in the current risk evolution to zero, thereby achieving precise adjustment of local parameters.

[0015] In one optional implementation, the distribution process includes instruction encapsulation and status feedback. The system encapsulates the optimal control instruction into a standardized message format containing the target identifier, action type, and parameters, and routes it to the corresponding target terminal device through the execution gateway. The system is equipped with a timeout monitoring mechanism; if no status confirmation message containing the execution result is received from the terminal within a preset time, an abnormal handling process such as retransmitting the instruction or escalating an alarm is triggered.

[0016] In an optional implementation, the method further includes a tiered control strategy. After determining the optimal control instruction, the system compares the overall utility score of the instruction with a preset execution threshold. If the overall utility score is less than the threshold, the system automatically issues the instruction; if the overall utility score is not less than the threshold, indicating a high risk or significant handling cost, the system determines that manual intervention is required, automatically generates a structured risk report containing risk chain information and simulation assessment data, and pushes it to the management terminal to assist in manual decision-making.

[0017] This invention provides an intelligent early warning and control method for safety risks in construction engineering. It has the following beneficial effects: 1. This invention effectively solves the technical challenge of accurately predicting the effectiveness of interventions in traditional early warning models by constructing a causal graph and performing parallel counterfactual simulations. The system creates an independent copy of the causal state space for each alternative instruction, forcibly applying intervention conditions and blocking the influence of the original parent node, thereby enabling the simulation of the risk evolution state after intervention in a virtual environment. This mechanism can accurately distinguish between correlation and causality between data, quantitatively assess the actual blocking effect of control measures on the risk propagation chain, and avoid ineffective early warnings or erroneous control due to false correlations.

[0018] 2. This invention introduces a decision optimization mechanism based on comprehensive utility scores, achieving a dynamic balance between construction safety and operational efficiency. The system maps intervention commands (such as deceleration, steering, and emergency stops) into quantified execution costs through a cost function, and then weights these costs with the significance of the global risk after the intervention. This allows the system to prioritize the optimal command that minimizes the impact on construction progress and resource consumption, while ensuring that risks are controllable, thus avoiding the production stagnation caused by the "one-size-fits-all" shutdowns common in traditional methods.

[0019] 3. This invention establishes a closed-loop correction mechanism for the causal graph based on execution results, endowing the early warning model with the ability to adapt to dynamic environments. The system collects factual data after command execution, calculates the deviation between the actual risk state and the simulated predicted value, and automatically fine-tunes the conditional probability parameters in the causal graph using the gradient descent algorithm. This mechanism enables the model to continuously correct its causal logic as construction data accumulates, automatically adapting to the risk evolution patterns under different construction stages and environmental changes, and maintaining long-term prediction accuracy. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the causal graph structure in an embodiment of the present invention; Figure 3 This is a diagram illustrating the risk salience calculation and activation path search of the present invention; Figure 4 This is a schematic diagram illustrating the parallel counterfactual simulation and intervention principle of the present invention; Figure 5 This is the optimal control command decision logic diagram of the present invention; Figure 6 This is a flowchart of the closed-loop correction process based on prediction bias according to the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example: Please see the appendix Figure 1 -Appendix Figure 6 This invention provides an intelligent early warning and control method for safety risks in construction projects, comprising the following steps: S101, Construct and update the causal state space in real time; the causal state space is a multi-layered data structure that integrates static physical data, dynamic entity data, environmental field data and causal graphs of the construction site.

[0023] S102, based on the current state in the causal state space, predict the future short-term state of the dynamic entity to generate a trajectory cluster.

[0024] S103, match the predicted future state with the causal graph, identify potential risk transmission chains, and calculate a risk significance index that includes the probability of risk occurrence, the severity of consequences, and the probability time gradient.

[0025] S104, determine whether the risk significance index exceeds a preset threshold. If it does, generate a set of alternative intervention instructions for the high-significance risk.

[0026] S105, establish a parallel counterfactual simulation environment for each instruction in the set of alternative intervention instructions, apply the corresponding instruction as an intervention condition in the environment, and re-perform risk simulation to assess the global risk level after intervention.

[0027] S106. Compare the global risk levels of each counterfactual simulation environment and select the instruction that minimizes the global risk level as the optimal control instruction.

[0028] S107, execute the optimal control command and use the actual result of the command execution as feedback data to correct and update the causal graph to form a closed-loop learning.

[0029] The construction and updating of the causal state space aims to build a spatiotemporally unified digital model that comprehensively represents the construction site. This model serves as the data foundation for subsequent risk simulations. Its construction process includes the collection, processing, and fusion of multi-source heterogeneous data. The specific implementation method is as follows: First, the collection of multi-source heterogeneous data is carried out, specifically including: For static environmental data, design information on the building's structure, component types, and material properties is extracted by parsing Building Information Modeling (BIM) files. The construction site is periodically scanned using a terrestrial 3D laser scanner or a LiDAR system mounted on a drone to acquire high-density 3D point cloud data reflecting the actual completed state of the site.

[0030] For dynamic entity data, UWB (Ultra-Wideband) positioning tags are configured for construction workers, and GPS modules or UWB tags are installed on large machinery and vehicles to obtain their 3D coordinates in the scene. Cameras are deployed in key areas, and a deep learning-based human pose estimation model is used to detect human targets in the video stream, extracting the coordinates of key skeletal points, including the head, shoulders, elbows, wrists, hips, knees, and ankles, thereby constructing limb pose vectors describing the human's movement characteristics.

[0031] For environmental field data, meteorological stations, gas concentration sensors, vibration sensors, and noise sensors are deployed at predetermined locations on site to collect parameters such as wind speed, rainfall, harmful gas concentration, and structural vibration frequency and amplitude.

[0032] After data acquisition is completed, the data is preprocessed and fused to construct a unified 3D scene, which includes the following sub-steps: S1011, Perform spatiotemporal benchmark unification. To ensure that data from different sensors are analyzed within a unified framework, their coordinate systems and timestamps need to be aligned.

[0033] Regarding spatial benchmark unification, the coordinate system of the BIM model is selected as the global coordinate system. Measurement values ​​from other data sources are transformed to this global coordinate system using a coordinate transformation matrix. This transformation employs a rigid body transformation model, and its mathematical expression is as follows: ; in, The coordinates of the entity in the global coordinate system. Let its coordinates be in the local coordinate system. Given a 3x3 rotation matrix, It is a 3x1 translation vector. (Matrix) with vector This is obtained by pre-calibrating the equipment.

[0034] Regarding time standardization, all data acquisition devices and servers within the system synchronize their time using the Network Time Protocol (NTP) to ensure that all uploaded data carries a unified timestamp.

[0035] S1012, perform scene model fusion.

[0036] The point cloud data, after coordinate transformation, is registered with the BIM model. The Iterative Closest Point (ICP) algorithm is used to align the point cloud with the BIM model surface. After alignment, the model is dynamically corrected by comparing the differences in the geometric boundaries between the point cloud data and the BIM model. Specifically, a voxel rasterization method is used to map the point cloud to the model space, marking areas occupied by point clouds but empty in the BIM model. These areas are identified as newly added temporary facilities or excavation areas, thereby generating a static background model reflecting the actual site conditions.

[0037] The spatiotemporally unified dynamic entity data is mapped onto the static background model. The worker's 3D coordinates, calculated from the UWB positioning tags, are used for localization within the fused 3D scene. For the 2D skeletal keypoint coordinates extracted by computer vision, combined with the camera's intrinsic and extrinsic parameter matrices and depth information, they are projected into 3D space and attached as pose attributes to the corresponding human entity objects.

[0038] Spatial interpolation is performed on the environmental field data. For discretely distributed environmental sensor data, Kriging interpolation or inverse distance weighted (IDW) interpolation methods are used to establish a voxelized environmental field covering the 3D scene, and corresponding environmental parameter values ​​are assigned to any spatial coordinate point in the scene.

[0039] Through the above steps, the system generates a unified, dynamically updated 3D scene model. A snapshot of this model at any given time constitutes the physical state layer in the causal state space, providing data input for risk deduction in subsequent steps S102 and S103. Regarding the specific algorithms involved in the data acquisition, coordinate transformation, and spatial interpolation, those skilled in the art can select mature solutions based on actual accuracy and computing power requirements; their implementation is well-known technology in the field and will not be elaborated upon here.

[0040] As the core step in constructing the causal state space in step S101, the construction of the causal graph provides the logical basis for risk transmission in the system. This process specifically includes the graph initialization step S1013 based on knowledge extraction, and the graph adaptive update step S1014 based on on-site data feedback.

[0041] S1013, perform initialization of the causal graph.

[0042] The system establishes a directed acyclic graph structure. As an initial causal graph. Among them, It is a set of nodes, representing various observable or inferable risk factors; The set of directed edges represents the causal relationships between factors. Natural language processing (NLP) techniques are used to process unstructured text data in the field of building safety. This text data includes safety technical specifications, industry standard operating procedures, and historical accident investigation reports. A deep learning-based entity recognition model (such as BiLSTM-CRF) is used to extract entities from the text data. During the extraction process, a thesaurus of building-related terms is introduced, mapping different words with the same semantic meaning (e.g., "crane," "gantry crane," "tower crane") to unique standard concept nodes. This eliminates semantic ambiguity. The identified entities encompass environmental conditions, personnel behavior, equipment status, and accident consequences.

[0043] A relation extraction algorithm is used to identify semantic relationships between entities, constructing knowledge triples in the form of "cause-effect". For each extracted causal relationship, a directed edge is established from the cause node to the result node. For each directed edge Assign initial conditional probabilities This probability is calculated based on statistical methods, specifically by analyzing historical corpora when a causal event occurs. When it occurs, the resulting event The frequency of simultaneous occurrences was also considered. For edges lacking statistical data, an expert scoring method was used to assign initial weights, thus completing the construction of the static graph.

[0044] S1014 performs adaptive updates to the causal graph.

[0045] This step utilizes on-site measured data to correct parameters and reconstruct the structure of the graph, adapting it to the actual working conditions of a specific construction site. For parameter correction, the system employs a Bayesian inference mechanism to dynamically update the conditional probabilities of directed edges. Each edge... conditional probability It can be considered as a random variable following a Beta distribution, i.e. ,in and Let be the hyperparameters of the distribution, and let be the number of positive samples and the number of negative samples, respectively. When the system detects a node... Event log At that time, update the corresponding hyperparameters according to the actual evolution of the event: If observed It happened and then If it also occurs, then update. ; If observed It happened but If it has not occurred, update. .

[0046] Updated expected value of conditional probability The calculation is as follows: ; The expected value Assigned to the edges in the graph As a new weight, the probability values ​​in the graph gradually approach the actual statistical patterns of the construction site.

[0047] In terms of structural reconstruction, the system processes cognitive bias events to improve the graph structure. Structural reconstruction is triggered when the prediction results of the risk saliency inference engine are inconsistent with the actual events that occur on site, or when the data mining module discovers a strong correlation between two nodes that are not directly connected in the graph.

[0048] Specifically, system computing nodes and Time series mutual information .like Exceeding the preset association threshold, and there is no correlation in the graph from arrive The system generates a candidate causal edge for the path. , the candidate edge The graph is temporarily added and assigned an initial confidence level. During subsequent time windows, the system monitors... and The co-occurrence of [variables]. If the data from new observations continue to support this causal hypothesis, i.e., satisfying the Bayesian factor [condition], then [the causal hypothesis] is considered. If the candidate edge is found to be true, then it is solidified into a permanent structure of the graph; otherwise, it is removed.

[0049] The Bayesian factor An approximate calculation is performed using the difference calculated by the Bayesian Information Criterion (BIC): ; in, For including candidate edges The model, This is a model that does not include this edge. , This represents the maximum likelihood function value of the model. The number of model parameters. The sample size is used. By comparing the BIC values ​​of the two models, it is quantitatively determined whether introducing new edges increases the explanatory power of the models. Through the above parameter correction and structural reconstruction, the causal graph achieves adaptive evolution for specific field environments.

[0050] Step S102 uses real-time motion data extracted from the causal state space to predict the motion trajectory of each dynamic entity at the construction site over a future period. This prediction generates a trajectory cluster containing deterministic paths and quantified positional uncertainties. For the different motion characteristics of personnel and machinery, corresponding state space models are established, and their specific implementation methods are as follows: S1021, Construct the kinematic state equations. For the construction worker entity, define its state at time... state vector ,in For position coordinates, Let be the velocity component. Using the constant velocity (CV) model, the state transition equation is: ; in Here is the state transition matrix. The process noise follows a Gaussian distribution.

[0051] For mechanical equipment entities (such as excavators and cranes), their state vector is defined as follows: ,in The magnitude of linear velocity, For heading angle, Let be the angular velocity. A cooperative turning (CT) model is adopted, with its nonlinear state transition function... The specific form is as follows: ; ; ; ; When external control input cannot be directly obtained At this time, it is assumed that the control input is zero, that is, the device maintains the current motion trend.

[0052] S1022, Perform Kalman filter prediction. The state of the entity is iteratively extrapolated using an extended Kalman filter (EKF). In the prediction extrapolation phase, based on the current time... Posterior state estimation and error covariance To project into the future Step. State prediction: ; Covariance prediction: ; in, State transition function Relative to the state vector The Jacobian matrix, whose elements are composed of The calculation yields the mean of the predicted state at a series of future moments through the above iterations. Covariance Matrix .

[0053] S1023, Generate trajectory clusters.

[0054] The system is based on the predicted covariance matrix An ellipsoidal region describing the location uncertainty is generated at each prediction time. This region is defined by the following equation: ; in For a degree of freedom of d and a confidence level of The chi-square distribution threshold.

[0055] To construct the trajectory family, a Sigma point sampling strategy from the Unscented Transform is employed. At each time step... According to the mean Generation of covariance P There are Sigma points (L is the state dimension). These Sigma points are connected in the time series to form multiple micro-paths covering the edge of the probability distribution. These paths, together with the mean path, constitute a trajectory cluster, which serves as the input to the subsequent risk identification module. The basic theory of the Kalman filter algorithm is well-known in this field and will not be elaborated here.

[0056] Step S103 connects the predicted data with causal knowledge to achieve a quantitative assessment of potential risks. It aims to identify risk paths that may be triggered in future states and calculate a significance index to measure the value of risk concern. This process includes a risk transmission chain matching and activation step S1031, and a multidimensional risk significance quantification step S1032.

[0057] S1031, Perform matching activation of the risk transmission chain. Map the dynamic entity trajectory cluster generated in step S102 and the environmental field data in S101 to the causal graph. In this process, the activation state of each risk factor node in the graph is determined. A node activation function is defined. For observation nodes Its activation state is determined by the degree of matching between the predicted data and the preset rules. For spatial location-related risks, the entity is calculated. Predicted trajectory points Dangerous areas minimum Euclidean distance .like Then the node is activated, and the activation confidence level is increased. ,in The preset distance attenuation coefficient, This is the safe distance threshold.

[0058] For interactive risks (such as human-machine co-operation), compute entities and The predicted trajectory overlap is determined. The Separating Axis Theorem (SAT) is used to determine the overlap between two entities at time [time value missing]. The uncertainty lies in whether the ellipsoids intersect. Specifically, if there exists an axis such that the projection intervals of the two ellipsoids on that axis do not overlap, they are determined to be non-intersecting; otherwise, they are determined to intersect, and the volume ratio of the overlapping region is used as the confidence level for node activation. For environmental risks, if the environmental field value at the prediction time exceeds the threshold specified by the national standard, the corresponding node is activated, and the confidence level is set to 1.

[0059] After determining the activation state of the observed node, forward reasoning is performed using the causal graph to calculate the probability P(R,t) of the occurrence of the downstream hidden node (i.e., the potential risk event R).

[0060] For nodes The Noisy-OR model is used to calculate its parent node set. The overall probability of its occurrence: ; in parent node Activation confidence level; The connection strength parameter indicates the connection strength when only the parent node is present. When it occurs, it causes child nodes The probability of occurrence is derived from the edge weights stored in the causal graph. The probability of all terminal risk nodes occurring at future time t is calculated recursively. .like If the probability exceeds a preset probability threshold, the risk transmission chain is considered activated.

[0061] S1032, Perform risk significance quantification calculation. Introduce risk significance indicators. Taking into account probability, consequences, and urgency, the significance calculation formula is as follows: ; in: Let R be the probability of the risk event R occurring at the predicted time t, with a value in the range [0,1].

[0062] Risk event The severity coefficient of the consequences is pre-stored in the risk knowledge base. The quantitative grading standards are as follows: no harm or minor property loss corresponds to... ; Corresponding to general injury or general property damage Serious injury or significant property damage corresponds to ; Response to death or major accidents .

[0063] The time gradient of the risk probability characterizes the urgency of risk evolution. The discretized calculation form is: ; in This represents the time step in risk projection. This term reflects the rate of increase in the probability of a risk event; when the probability of a risk event occurs rapidly, this gradient term will significantly increase the indicator value.

[0064] , , These are the probability weights, consequence weights, and gradient weights, respectively, and satisfy the following conditions: The system calculates all activated risk events. The results are then output to the decision module.

[0065] The specific implementation method for generating the high-significance risk triggering and intervention instruction set described in step S104 is as follows: When any risk significance index is calculated in step S103 Exceeding the preset dynamic trigger threshold At that time, the system initiates the intervention decision-making process.

[0066] S1041, Execution risk trigger determination.

[0067] The dynamic trigger threshold It is not a fixed value, but is adjusted according to the macroscopic conditions of the construction site. The calculation method is as follows: ; in, Basic threshold; This represents the number of active entities within the current work area. and These are the weighting coefficients; This is a time-influence coefficient, the value of which is obtained from a preset time-coefficient value lookup table based on the construction plan. For example, the coefficient for nighttime construction periods (22:00-06:00) is set to a higher value, while the coefficient for normal working hours is set to a lower value. Through this calculation method, the sensitivity of the early warning trigger can be dynamically adjusted according to the intensity of on-site work and specific time periods.

[0068] S1042, Execute risk tracing of key entities.

[0069] After a risk is triggered, the system identifies the high-significance risk event. Relevant key entities. This process is achieved through causal mapping. The reverse tracing is implemented from the activated terminal risk node. Initially, the system traverses all parent nodes along the directed edges in reverse order, recursively tracing upwards until it reaches the initial observation node in the large graph. These initial observation nodes are directly associated with specific dynamic entities or static regions in the 3D scene. The system records the IDs of all associated entities along this tracing path, forming a set of key entities. .

[0070] S1043, Generate alternative intervention instruction set.

[0071] Based on the set of key entities obtained through tracing, the system queries a pre-defined intervention strategy database and generates one or more alternative intervention instructions for each key entity.

[0072] The intervention strategy database is a structured knowledge base, indexed by "risk type-entity type" tuples, storing a list of corresponding atomic operation instructions. For example: For the index ("collision risk", "vehicle"), the database returns a list of atomic operations: ["decelerate to 0", "avoid according to preset trajectory", "honk horn warning"].

[0073] For the index ("Fall from Height Risk", "Personnel"), the database returns a list of atomic operations: ["Helmet Vibration Alarm", "Voice Warning of Danger"].

[0074] The system combines the retrieved atomic operation instructions with key entity IDs to generate structured alternative intervention instructions. Each instruction Defined as a data structure: Here, TargetID is the unique identifier of the key entity; ActionType is the atomic operation type; and Parameters are the specific parameters required to execute the operation. For example, for the "decelerate" operation, the parameter is the target speed value; for the "avoid according to preset trajectory" operation, the parameter can be the unique identifier of the avoidance trajectory or a series of target path point coordinates. Finally, the system summarizes all generated alternative intervention instructions to form an alternative intervention instruction set. This serves as the input for the next step of counterfactual simulation.

[0075] Step S105 assesses the effectiveness and potential negative impacts of each instruction by simulating the consequences of different intervention measures in a virtual environment. The specific implementation method is as follows: S1051, Create a parallel simulation environment.

[0076] For the alternative intervention instruction set generated in step S104 The system provides each of the alternative instructions. Create a separate, temporary copy of the causal state space (CSS). This copy is generated by replicating the complete data state of the CSS at the current moment, including the 3D scene model, the state vectors and covariance matrices of all dynamic entities, and the node activation states of the causal graph. This process establishes... A parallel simulation branch with a consistent initial state is used to ensure that the derivation of each branch does not interfere with each other.

[0077] S1052, Applying counterfactual conditions.

[0078] In each independent simulation environment copy In the middle, the system forcibly executes the corresponding alternative intervention command. This involves modifying the future state of an entity. For instructions that change kinematic parameters (such as "decelerate" or "stop"), the state vector of the target entity TargetID is directly used during state prediction in step S102. The velocity component in the command is modified to the value specified in the Parameters parameter. For path planning commands (such as "avoid according to preset trajectory"), the path point coordinate sequence defined in the Parameters parameter is directly used to replace all position coordinates of the target entity TargetID in the future prediction time step. For information transmission commands (such as "voice alarm" and "vibration alert"), in the simulation, this command is implemented by modifying the conditional probability of specific nodes in the causal graph. For example, for an alarm targeting personnel, the system multiplies the probability of occurrence of risk factor nodes such as "failure to notice danger" related to that personnel by a correction coefficient less than 1 (e.g., 0.5). This coefficient is preset in the human factors engineering database, indicating that the alarm can effectively reduce the probability of human negligence.

[0079] S1053, Perform risk simulation and global risk assessment. After applying counterfactual conditions, the system performs simulations in each replica of the simulation environment. In this process, the complete procedures of steps S102 and S103 are re-executed. This risk reassessment is global, not only recalculating the original high-significance risk. Significance indicators It will also calculate new risks that may be induced by all other entities and areas in the scene. The significance of the intervention measures is used to quantify the potential secondary risks.

[0080] Ultimately, the system provides each simulation branch with... Calculate a global risk significance sum The calculation formula is as follows: ; in, For the current moment, The preset simulation time domain; This represents the total number of all possible risk events in the scenario. This indicates that an intervention instruction has been issued. Under the premise of risk events In the future The significance index. In the discrete-time implementation, this integral is replaced by summation. The value represents the number of steps taken. After these intervention measures, the overall risk level of the entire construction scenario in the future.

[0081] The execution of the optimal control command transforms the decision-making results into actual intervention actions in the physical world and ensures that the command is effectively executed. The specific implementation method is as follows: S1071, execution instruction encoding and routing.

[0082] The system first selects the optimal control command in step S106. The instruction is encoded, transforming its internal data structure into a standardized message format. It is encapsulated in JSON format, with the JSON message body containing multiple key-value pairs corresponding to the instruction's target ID, action type, specific parameters, and a unique message ID for tracking. The instruction is then routed and distributed through an execution gateway. This gateway uses the MQTT protocol to publish the encoded instruction message to a specific topic. The topic name follows a predefined hierarchical structure; for example, its path may include fields such as instruction category, entity type, and entity ID, ensuring that the instruction is accurately routed to the corresponding end device for subscription.

[0083] S1072, Execution of Terminal Parsing and Execution. The smart terminal device bound to the target entity continuously subscribes to its corresponding MQTT topic. Upon receiving the instruction message, the terminal device first parses the JSON message body to extract the action type and parameters. The execution method differs for different types of terminals: If the target entity is a person, after receiving the "vibration alarm" instruction, the built-in microcontroller of the smart safety helmet will drive the vibration motor to work at the specified intensity and duration. If the target entity is construction machinery with an automatic control interface (such as an autonomous truck), after its onboard computing unit receives the "deceleration" instruction, it will convert the high-level instruction into a vehicle-level control signal, for example, sending a braking pressure instruction to the vehicle's electronic brake controller via the CAN bus until the speed feedback from the vehicle speed sensor reaches the target speed.

[0084] S1073, Execution Status Feedback and Anomaly Handling. To ensure a closed-loop control system, the terminal device must send a status confirmation message back to the system after executing a command. This confirmation message, also in JSON format, contains the message ID of the original command and its execution status (e.g., "Executed," "Execution Failed," "Parameter Exceeded," etc.) and is published to a dedicated feedback topic. The naming convention for this feedback topic is also associated with the target ID, facilitating the system's matching of feedback information with the original command. Upon receiving the confirmation message, the system updates the command's status to "Completed." If no confirmation message is received for a specific command within a preset timeout period, the system marks the command as "Execution Timeout" and triggers an anomaly handling process. This process may include resending the command, or, if failures occur consecutively, automatically escalating the event to a high-level alarm, notifying a human administrator to intervene and check the target terminal's communication or functional status.

[0085] After the control command in step S107 is executed, the system continues to acquire data on the actual state evolution of the target entity and its surrounding environment after intervention through the data acquisition module in step S101. This data constitutes the "fact" trajectory, serving as a benchmark for evaluating the accuracy of the prediction.

[0086] The system will collect the factual data and compare it with the optimal instruction in step S105. The prediction results obtained from counterfactual simulation are compared. Specifically, the system uses factual data to rerun the risk significance calculation process in step S103, obtaining a risk significance index based on real-world results. Then, the prediction bias was calculated. in, This refers to the significance of the risk predicted in counterfactual simulations after intervention measures are taken. If This indicates that the system underestimated the risks or that the intervention measures were less effective than expected; if This indicates that the system overestimated the risks or that the intervention measures were more effective than expected.

[0087] Then, regarding the probability parameters (As defined in S1031) Perform an update. The update rules are as follows: ; The system calculates the prediction deviation. , on causal graph Medium and risk events The relevant parameters are adjusted. This invention uses a gradient descent-based approach to fine-tune the parameters, specifically for the connection to the parent node. and child nodes The weight of the edge, i.e., the condition, is given by... and These are the conditional probability parameters before and after the update, respectively; It is a preset learning rate used to control the step size of each update; This is the gradient of the risk saliency with respect to the conditional probability parameters. This gradient is calculated in the causal graph using the chain rule. Specifically, the system first traces backward along the causal graph from the risk node R, following all causal paths that lead to the activation of that risk. The gradient is calculated only with respect to the weight parameters of the edges located on these paths. The gradient is non-zero for parameters not on the path, and therefore not included in the update. Through this update process, if the system underestimates the risk ( The weights on the causal paths that have a positive impact on risk are then determined. The value will be increased; conversely, it will be decreased. This adjustment process is performed after each effective intervention event, thereby achieving continuous optimization of the model.

Claims

1. A construction engineering safety risk intelligent early warning and management method, characterized in that, Includes the following steps: Construct a 3D digital scene containing static backgrounds and dynamic entities, and establish a causal graph representing the causal relationships of risk events; State prediction is performed on dynamic entities in the three-dimensional digital scene to generate future state trajectories; The predicted future state is matched with the causal graph, and the activated risk chains are identified, thereby calculating the risk salience. When the significance of the risk exceeds a preset threshold, a set of alternative intervention instructions is generated to address the risk. Parallel counterfactual simulations are performed on each alternative intervention instruction in the set of alternative intervention instructions to calculate the sum of global risk significance after applying the instruction; By combining the intervention costs of each alternative intervention instruction with the sum of the global risk significance, a comprehensive utility score is calculated, and the optimal control instruction is determined based on the comprehensive utility score. The optimal control command is executed, factual data after the execution of the optimal control command is collected, the prediction deviation is calculated, and the causal graph is closed-loop corrected based on the prediction deviation.

2. The construction engineering safety risk intelligent early warning and management method according to claim 1, characterized in that, The step of identifying activated risk chains in a causal graph and then calculating the risk significance of the risk includes: Based on the state prediction results of the dynamic entity, the activated risk factor nodes are identified in the causal graph. Starting from the risk factor node, the risk chain is formed by searching along the causal path to the risk event node. Based on the conditional probability parameters of each node in the risk chain, the probability of occurrence of the risk event node is calculated. The significance of the risk is calculated by considering the potential loss of the risk event.

3. The construction engineering safety risk intelligent early warning and management method according to claim 1, characterized in that, The steps for performing parallel counterfactual simulations for each alternative intervention instruction include: For each alternative intervention instruction in the set of alternative intervention instructions, create an independent copy of the causal state space; In each independent copy of the causal state space, the corresponding alternative intervention instruction is forcibly applied as a counterfactual condition, and the influence of the original parent node of the node targeted by the instruction is blocked. In a copy of the causal state space with counterfactual conditions applied, risk deduction is performed again, the significance of all potential risks is calculated, and the sum of the global risk significance is obtained.

4. The construction engineering safety risk intelligent early warning and management method according to claim 1, characterized in that, The step of calculating a comprehensive utility score by combining the intervention costs of each alternative intervention instruction with the sum of the global risk significance, and determining the optimal control instruction based on the comprehensive utility score, includes: Based on a preset cost function, the intervention cost of each alternative intervention instruction is calculated; Using preset risk weights and cost weights, the total global risk significance is weighted and summed with the intervention cost to obtain the comprehensive utility score. Compare the overall utility scores of all candidate intervention instructions, and select the candidate intervention instruction with the lowest overall utility score as the optimal control instruction.

5. The construction engineering safety risk intelligent early warning and management method according to claim 4, characterized in that, The cost function maps the candidate intervention instructions to quantified intervention costs based on the atomic operation type and specific parameters of the instructions; wherein, for instructions involving speed adjustment, the intervention cost is proportional to the speed adjustment magnitude.

6. The construction engineering safety risk intelligent early warning and management method according to claim 1, characterized in that, The steps for closed-loop correction of the causal graph include: Collect factual data after the execution of the optimal control command, and recalculate the significance of factual risk based on the factual data; The prediction bias is obtained by calculating the difference between the factual risk significance and the risk significance predicted for the optimal control command in the parallel counterfactual simulation. Based on the prediction bias, the risk-related conditional probability parameters in the causal graph are updated.

7. The intelligent early warning and control method for safety risks in construction projects according to claim 6, characterized in that, The step of updating the risk-related conditional probability parameters in the causal graph based on the prediction bias includes: Calculate the gradient of risk significance with respect to the conditional probability parameter; The parameter adjustment amount is obtained by calculating the product of the preset learning rate, the prediction bias, and the gradient. The adjusted parameters are then added to the original conditional probability parameters to obtain the updated conditional probability parameters.

8. The construction engineering safety risk intelligent early warning and management method according to claim 7, characterized in that, Calculating the gradient of risk significance with respect to the conditional probability parameters includes: Starting from the risk event node, trace back along the causal graph to all causal paths that led to the activation of the risk; Non-zero gradients are computed only for the conditional probability parameters of edges located on the causal path; for parameters not located on the causal path, their gradients are set to zero.

9. The construction engineering safety risk intelligent early warning and management method according to claim 1, characterized in that, The steps for issuing and executing optimal control instructions also include: The optimal control command is encapsulated into a standardized message format that includes target identifier, action type, and parameters; By executing the gateway, the message is routed to the corresponding target terminal device based on the target identifier; The system receives a status confirmation message from the target terminal device after executing the instruction. If the status confirmation message is not received within a preset time, an exception handling process is triggered.

10. The construction engineering safety risk intelligent early warning and management method according to claim 1, characterized in that, The method further includes: After determining the optimal control instruction, the comprehensive utility score of the instruction is compared with the preset execution threshold; If the overall utility score is less than the execution threshold, the optimal control instruction will be automatically issued. If the overall utility score is not less than the execution threshold, it is determined that manual intervention is required, a structured risk report containing risk chain information and simulation evaluation results is generated, and pushed to the management terminal.

Images