Quality safety management method and system for water conservancy construction project

Abstract

The present application relates to the technical field of engineering construction management, in particular to a quality and safety management method and system for water conservancy construction engineering. The present application obtains multi-source heterogeneous data of construction engineering and builds a knowledge graph containing physical domain and non-physical domain nodes, and establishes a dynamic causal model integrating physical causal layer and situational causal layer. The model can be dynamically reconstructed and incrementally learned according to the changes of construction stages and the inflow of new data. When the deviation between the predicted value and the actual value of the model exceeds the threshold value, the system determines it as a risk event, and traces back to the root cause from the physical domain anomaly point to the non-physical domain through hierarchical causal tracing. Finally, the expected impact of different control measures on the construction risk index is evaluated through forward-looking deduction, and the optimal control measures are generated and issued, solving the problem that the prior art cannot perform deep causal attribution, dynamic adaptation and quantitative decision-making.

Description

Technical Field

[0001] This invention relates to the field of engineering construction management technology, specifically to a quality and safety management method and system for water conservancy construction projects. Background Technology

[0002] With the development of information technology, engineering quality and safety management has entered a new stage of digitalization and intelligence. Building Information Modeling (BIM) enables visualized simulation of the construction process, design conflict detection, and resource optimization by establishing three-dimensional models of projects, thereby improving the level of precision in project management to a certain extent. At the risk analysis level, some technologies use machine learning to predict the evolution trend of structural health status and simulate the transmission and diffusion of risks in the engineering physical structure. Some technical solutions explicitly propose to explore the causal relationship between risk factors and identify common causes and mutual inducing mechanisms by using statistical methods such as correlation analysis and regression analysis of historical data. Although existing technologies have made significant progress in digitalization and intelligence, some problems still exist in terms of in-depth risk cognition and proactive safety control.

[0003] Existing risk analysis models struggle to achieve fundamental causal attribution for cross-domain, multimodal risks. Construction safety risks are often the result of the combined effects of cross-domain factors such as management, personnel, equipment, environment, and materials. While existing technologies can identify statistical correlations between risk phenomena, they cannot fundamentally distinguish true causality. That is, the system cannot integrate multimodal heterogeneous data within a unified model and deduce the root causes of risks. This results in risk management often only addressing symptoms and superficial issues without addressing the root causes. Furthermore, existing models lack the ability to autonomously discover and learn causal structures in dynamic environments. Construction sites are dynamically evolving systems; new risk factors and unknown causal chains emerge continuously as construction progresses. Current technical solutions are typically based on fixed risk frameworks or... Analysis based on predefined knowledge graphs of physical connections is a static or semi-static model that cannot adapt to the dynamic changes at construction sites. It lacks the ability to learn autonomously from continuously generated data, discover and update causal structures, thus exhibiting cognitive lag when new risks emerge. Existing technologies generally lack the ability to make counterfactual inferences based on causal models, making it difficult to conduct forward-looking quantitative assessments of control measures. Current risk response mechanisms mainly focus on issuing warnings and recommending static, pre-set emergency plans. This lack of "counterfactual inference" ability makes safety decisions heavily reliant on the personal experience of managers, making it impossible to quantitatively compare and scientifically select the future effects of different intervention measures in complex situations. This limits the profound transformation of safety management from passive response to proactive prevention and intelligent decision-making.

[0004] Therefore, a quality and safety management method and system for water conservancy construction projects is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a quality and safety management method and system for water conservancy construction projects. By acquiring multi-source heterogeneous data from the construction project, a knowledge graph containing nodes from both the physical and non-physical domains is constructed. A dynamic causal model integrating physical and contextual causal layers is established. This model can dynamically reconstruct and incrementally learn based on changes in the construction stage and the influx of new data. When the deviation between the model's predicted value and the actual value exceeds a threshold, the system identifies it as a risk event and uses hierarchical causal tracing to trace back from the anomaly point in the physical domain to the root cause in the non-physical domain across modalities. Through forward-looking extrapolation, the expected impact of different control measures on the construction risk index is evaluated, and the optimal control measures are generated and issued. This solves the problem that existing technologies cannot perform deep causal attribution, dynamic adaptation, and quantitative decision-making.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A quality and safety management method for water conservancy construction projects includes: Acquire real-time and historical multi-source heterogeneous data of construction projects, and generate causal feature datasets by constructing a knowledge graph containing physical domain and non-physical domain nodes; A physical causal layer is constructed for the physical domain based on prior physical knowledge, and a contextual causal layer is constructed for the non-physical domain based on causal discovery and latent variable inference models. The two layers are then integrated into a dynamic causal model. When the engineering stage changes, nodes and causal relationship edges of the physical causal layer are dynamically added, deleted, and reconstructed. For newly flowing non-physical feature parameters, an incremental learning mechanism is used to update the local subgraph of the contextual causal layer in real time. The system uses a dynamic causal model to generate predicted values ​​and calculate prediction bias. When the prediction bias exceeds the threshold, it is identified as a risk event. Through hierarchical causal tracing, it starts from anomalies in the physical domain and backtracks across modes to non-physical domains to generate symptom points and quantifies their contribution. For the symptom points with the highest contribution, the system obtains the corresponding control measures from the preset knowledge base. Through forward-looking extrapolation, it evaluates the expected change in the construction risk index after taking control measures and generates the optimal control measures.

[0007] Preferably, the physical domain refers to entities, materials, and processes with measurable physical properties at the construction site; the unphysical domain refers to abstract concepts, events, and attributes related to human decision-making, behavior, and management processes; for unstructured image data and text data, a pre-trained model is invoked to identify and extract the engineering entities, construction activities, management events, and status attributes contained therein; the engineering entities include at least on-site workers, construction machinery and equipment, and component IDs in the BIM model; the construction activities include at least concrete pouring, formwork support, and rebar tying; the management events include at least design change approval, safety technical briefing, and work stoppage order issuance; the status attributes include at least personnel violations, abnormal equipment conditions, and apparent defects in materials; Based on a pre-built knowledge graph ontology of pump station water conservancy construction projects, engineering entities, construction activities, state attributes, and structured design data and physical quantity time series data are associated and mapped. Causal feature datasets are generated by transforming data from different sources into entity nodes, event nodes, and relation edges with unique identifiers and timestamps in the knowledge graph.

[0008] Preferably, a physical causal layer is constructed using a physical model that describes deterministic relationships between physical variables; a contextual causal layer is constructed using a data-driven probabilistic graphical model, wherein the nodes of the contextual causal layer correspond to non-physical domains in the knowledge graph, and the causal relationship edges and probability parameters are obtained by causal discovery on a causal feature dataset; the two layers are fused together through shared nodes. In the contextual causal layer, non-physical factors are modeled and quantified contextual influence factors are output; in the physical causal layer, selected parameters are defined as parameter correction nodes, and the parameter correction nodes contain the contextual influence factors. The causal discovery is a hybrid-driven algorithm combining scoring search and conditional independence testing. A heuristic search is performed in the causal feature space using the Bayesian information criterion scoring function to generate candidate causal network skeletons. For each causal relationship edge in the candidate skeleton, the G² test is used to verify the authenticity of the relationship and eliminate false relationships that do not meet the conditions. The latent variable inference model infers the state of latent variables by analyzing the collaborative change patterns of observable causal features and using a variational autoencoder. The latent variables include at least personnel fatigue inferred by analyzing continuous working hours, frequency of operational errors, and behavioral postures in videos, and project management pressure inferred by analyzing the frequency of schedule adjustments, number of resource allocation changes, and quality inspection exemption records in management instructions.

[0009] Preferably, the system receives and parses construction data from the BIM model in real time; when a change in a construction phase event is detected, a graph structure change script associated with the event is triggered; the script performs add, delete, and modify operations on nodes and causal relationship edges of the dynamic causal model according to preset rules; the incremental learning mechanism includes: setting up a data change monitoring unit to monitor the distribution difference between newly incoming causal feature datasets and historical datasets in real time; when a significant change in data distribution is detected, the variable nodes directly associated with the data change are located and the causal discovery algorithm is re-executed only in the local subgraph region formed by these nodes and their adjacent nodes.

[0010] Preferably, a target node to be predicted is selected according to a preset monitoring strategy, and the actual observation values ​​of evidence nodes that have a causal relationship with the target node are obtained. The actual observation values ​​of the evidence nodes are used as input, and a probabilistic causal inference algorithm is executed on the dynamic causal model. Based on the given state of the evidence node, the posterior probability distribution of the target node is calculated through a Bayesian network. The expected value of the posterior probability distribution is extracted and used as the predicted value of the target node in the current causal context. The actual observation value of the target node at the same time is obtained from the causal feature dataset, and the prediction deviation between the actual observation value and the predicted value is calculated. If the prediction deviation exceeds a preset risk threshold, it is determined as a risk event. The preset risk threshold is determined according to the risk level of different monitoring targets.

[0011] Preferably, reverse reasoning is performed in the physical causal layer. Starting from the risk event of physical anomaly, the direct physical cause leading to the anomaly is identified and output as a physical cause report along the determined physical law path. Using the direct physical cause as a new starting point for reasoning, reverse reasoning is continued in the situational causal layer. The root cause leading to the non-physical domain is identified and output as a root situational cause report along the causal path learned by data-driven learning.

[0012] Preferably, the construction risk index is calculated as follows: risk nodes in the dynamic causal model are identified, the probability of occurrence of each risk node under the current conditions is calculated based on the probability distribution of the model, and the probability values ​​of all risk nodes are weighted and summed according to their corresponding risk level weights to obtain the construction risk index; the risk level weights are determined using the analytic hierarchy process (AHP) to determine the relative importance of multiple impact dimensions such as safety, quality, schedule, and cost, and the quantitative impact of specific risks on these dimensions is weighted and summed. The project uses a dynamic causal model to perform forward-looking simulations, calculating the expected change in the construction safety risk index after the implementation of control measures. This also includes assessments of potential secondary risks and resource impacts that the control measures may trigger. The secondary risk assessment refers to the simultaneous calculation of risk index changes for other non-target nodes in the causal network during the forward-looking simulation. The resource impact assessment refers to evaluating the impact of the measures on construction costs and schedule. The optimal control measures are determined based on a multi-objective optimization function aimed at minimizing the risk index, secondary risks, and resource impacts.

[0013] Preferably, a closed-loop management instruction for the construction period is generated based on the optimal control measures, and the instruction is issued to the corresponding management and personnel terminals. The closed-loop management instruction includes a dynamic digital work permit that is linked in real time with on-site personnel authentication and equipment management. When the instruction is issued, it includes the minimum personnel qualification level, specific equipment model, and health status threshold required to execute the instruction. Before performing the operation, on-site personnel need to scan the equipment ID and personal identification ID through a mobile terminal to request activation of the permit. The backend will verify in real time whether the personnel qualification and equipment status meet the instruction requirements. Only when all conditions are fully matched will the digital work permit be authorized and activated, and the operation permission of the corresponding equipment be unlocked.

[0014] A quality and safety management system for water conservancy construction projects, comprising: Causal modeling module: Acquire multi-source heterogeneous data from the construction site, generate causal feature dataset by constructing a knowledge graph containing nodes of physical domain and non-physical domain; construct a physical causal layer for the physical domain based on prior physical knowledge, construct a contextual causal layer for the non-physical domain based on causal discovery and latent variable inference models, and merge the two layers into a dynamic causal model; Risk prediction module: When the engineering stage changes, the physical causal layer nodes and causal relationship edges are dynamically added, deleted and reconstructed; for newly incoming non-physical feature parameters, an incremental learning mechanism is used to update the local subgraph of the contextual causal layer in real time; the dynamic causal model is used to generate predicted values ​​and calculate prediction deviations; when the prediction deviation exceeds the threshold, it is judged as a risk event. Risk management module: Starting from anomalies in the physical domain through hierarchical causal tracing, backtracking across modalities to non-physical domains to generate symptom points and quantify their contribution; for the symptom points with the highest contribution, obtain the corresponding control measures from the preset knowledge base, and through forward-looking extrapolation, evaluate the expected change in the construction risk index after taking control measures, and generate the optimal control measures.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a dynamic causal model that integrates physical and non-physical domains. When a risk event occurs, this method supports tracing back from physical anomalies to deeper non-physical risk causes across modalities. This mechanism helps engineering managers to more comprehensively understand the risk transmission chain and thus adopt more targeted control strategies.

[0016] 2. Addressing the dynamic evolution of construction sites over time, this invention proposes an adaptive update mechanism for the causal structure. When changes occur in the engineering phase, the model can update its physical causal layer according to preset rules; simultaneously, leveraging an incremental learning mechanism, the model locally updates the contextual causal layer by monitoring changes in the distribution of newly incoming data. This design enables the system to continuously adapt to changes in the site environment as construction progresses, promptly identifying and learning newly emerging risk factors and their causal relationships.

[0017] 3. In the assessment and formulation of control measures, this invention introduces a forward-looking extrapolation mechanism, which can quantitatively assess the expected changes in the future construction risk index of different intervention measures. Combining a comprehensive consideration of potential secondary risks and the impact on project costs and schedules, the system utilizes a multi-objective optimization function to generate reasonable control measures that take into account multiple dimensions, thereby providing scientific quantitative decision-making references for construction sites and promoting the transformation of safety management from response and handling to pre-emptive prevention. Attached Figure Description

[0018] Figure 1 A flowchart of a quality and safety management method for water conservancy construction projects provided in this embodiment of the invention; Figure 2 A schematic diagram of the structure of a quality and safety management system for water conservancy construction projects provided in an embodiment of the present invention; Figure 3 A flowchart is provided to illustrate the specific implementation method of dynamic updating of the dynamic causal model in the embodiments of the present invention. Detailed Implementation

[0019] The technical solutions of 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.

[0020] Please see Figures 1 to 3 This invention provides a method for quality and safety management in water conservancy construction projects, the technical solution of which is as follows: A quality and safety management method for water conservancy construction projects includes: Acquire real-time and historical multi-source heterogeneous data of construction projects, and generate causal feature datasets by constructing a knowledge graph containing physical domain and non-physical domain nodes; A physical causal layer is constructed for the physical domain based on prior physical knowledge, and a contextual causal layer is constructed for the non-physical domain based on causal discovery and latent variable inference models. The two layers are then integrated into a dynamic causal model. When the engineering stage changes, nodes and causal relationship edges of the physical causal layer are dynamically added, deleted, and reconstructed. For newly flowing non-physical feature parameters, an incremental learning mechanism is used to update the local subgraph of the contextual causal layer in real time. The system uses a dynamic causal model to generate predicted values ​​and calculate prediction bias. When the prediction bias exceeds the threshold, it is identified as a risk event. Through hierarchical causal tracing, it starts from anomalies in the physical domain and backtracks across modes to non-physical domains to generate symptom points and quantifies their contribution. For the symptom points with the highest contribution, the system obtains the corresponding control measures from the preset knowledge base. Through forward-looking extrapolation, it evaluates the expected change in the construction risk index after taking control measures and generates the optimal control measures.

[0021] Example 1: This embodiment will describe in detail the specific application of the present invention in a typical water conservancy construction scenario—the large-volume concrete pouring operation of the deep foundation pit bottom slab of a large pumping station.

[0022] As one embodiment of the present invention, refer to Figure 1 A flowchart of a quality and safety management method for water conservancy construction projects, referring to... Figure 2 A schematic diagram of the structure of a quality and safety management system for water conservancy construction projects, with reference to... Figure 3 The flowchart illustrates the specific implementation method of dynamic causal model dynamic updating.

[0023] Furthermore, for unstructured image and text data, a pre-trained model is invoked to identify and extract the engineering entities, construction activities, management events, and status attributes contained therein; the engineering entities include at least on-site workers, construction machinery and equipment, and component IDs in the BIM model; the construction activities include at least concrete pouring, formwork support, and rebar tying; the management events include at least design change approval, safety technical briefing, and work stoppage order issuance; the status attributes include at least personnel violations, abnormal equipment status, and material defects. Based on a pre-built knowledge graph ontology of pump station water conservancy construction projects, engineering entities, construction activities, state attributes, and structured design data and physical quantity time series data are associated and mapped. Causal feature datasets are generated by transforming data from different sources into entity nodes, event nodes, and relation edges with unique identifiers and timestamps in the knowledge graph.

[0024] The knowledge graph ontology defines the following core entity classes and their key attributes: 1. "Personnel" entity node includes attributes: personnel identifier, qualification level (range: primary, intermediate, advanced, special type of work), work team, and cumulative working hours for the day (unit: hours, accuracy: 0.1 hours); 2. "Construction Machinery and Equipment" entity node includes attributes: equipment identifier, equipment model, access protocol type, and normalized health status (range: 0 to 1); 3. "Construction Activity" event node includes attributes: activity type, start timestamp, end timestamp, and associated component identifier; 4. "Management Event" event node includes attributes: event type, timestamp, and associated responsible person identifier; 5. "Physical Quantity" node includes attributes: sensor identifier, measurement type, value, unit, and timestamp. The node classification rules are as follows: nodes containing physical quantity attributes and whose objective values ​​can be directly collected by sensors are classified into the physical domain; nodes whose main attribute source is text events or personnel behavior decisions and whose state requires semantic analysis are classified into the non-physical domain. Cross-domain shared nodes are retained in both layers of the network to achieve inter-layer fusion in the form of shared nodes. Relationship types are divided into two categories: direct physical causality for the physical causality layer and probabilistic causality for the contextual causality layer. Different types of relationship edges cannot be mixed.

[0025] Specifically, the first step is to access and integrate multi-source heterogeneous data from the construction site, including structured data such as component IDs and design parameters from the BIM model; time-series data such as steel reinforcement stress (e.g., 35.2 MPa, every 15 minutes) and concrete internal temperature (e.g., 45.8°C, every 10 minutes) collected by IoT sensors, as well as environmental temperature and humidity data; and unstructured data such as video streams, construction logs, safety technical briefing meeting minutes, and design change approval forms collected by high-definition cameras deployed on site.

[0026] The system calls pre-trained deep learning models to process unstructured data. The YOLOv5 model is used to identify on-site workers, concrete pump trucks, and other engineering entities in the video, as well as status attributes such as personnel violations (e.g., not wearing safety helmets). The BERT model is used to extract construction activities such as concrete pouring and formwork support, and management events such as stop-work orders from construction logs. Based on a pre-built knowledge graph ontology for pump station water conservancy construction engineering, which predefines core entity categories such as 'personnel,' 'equipment,' 'construction activities,' and 'management events,' and their logical relationships such as 'operation' and 'impact,' it provides a unified structural framework for data fusion. The extracted entities, activities, events, and attributes are mapped to structured data. For example, the ID of a worker performing a pouring operation identified by a camera is associated with the ID of the pump truck they are operating, the corresponding component ID in the BIM model, and the real-time temperature data measured by sensors in that area. All associated data is uniformly formatted to form a causal feature dataset with unique identifiers and timestamps, providing high-quality input for subsequent causal modeling.

[0027] By unifying multi-source heterogeneous data under a knowledge graph framework, the problem of data silos at construction sites is solved, providing a high-quality, semantic data foundation for subsequent causal analysis, which is a prerequisite for achieving deep causal discovery.

[0028] Furthermore, a physical causal layer is constructed by a physical model describing deterministic relationships between physical variables; a contextual causal layer is constructed by a data-driven probabilistic graphical model, wherein the nodes of the contextual causal layer correspond to non-physical domains in the knowledge graph, and the causal relationship edges and probability parameters are obtained by performing a causal discovery algorithm on the causal feature dataset; the two layers are fused through shared nodes. In the contextual causal layer, non-physical factors are modeled and quantified contextual influence factors are output. In the physical causal layer, selected parameters are defined as parameter correction nodes, and the parameter correction nodes contain the contextual influence factors.

[0029] The construction of the physical causal layer further includes: incorporating geotechnical engineering data from the foundation pit monitoring system into multi-source heterogeneous data, wherein the geotechnical engineering data includes at least deformation data of the retaining structure, axial force data of the supporting structure, and seepage pressure data of the groundwater level; and establishing in the physical causal layer the physical causal relationship between the thermal expansion and contraction during the pouring of the large-volume concrete and the additional stress on the retaining structure and supporting structure, as well as the physical causal relationship between the deformation of the retaining structure and the early internal stress distribution of the concrete.

[0030] By elevating risk management from individual components to the systemic safety level of the interaction between the construction body and the deep foundation pit environment, it can proactively warn of deep-seated catastrophic risks such as instability of the retaining structure and overload of the supports caused by pouring, providing a global decision-making basis for dynamic construction adjustments, and avoiding the drawback of focusing only on local quality while ignoring overall safety.

[0031] The causal discovery algorithm is a hybrid-driven algorithm combining scoring search and conditional independence testing. It uses the Bayesian information criterion scoring function to perform a heuristic search in the causal feature space to generate candidate causal network skeletons. For each causal relationship edge in the candidate skeleton, the G² test is used to verify the authenticity of the relationship and eliminate false relationships that do not meet the conditions. The latent variable inference model analyzes the collaborative change patterns of observable causal features and uses a variational autoencoder to infer the state of latent variables. The latent variables include at least personnel fatigue inferred by analyzing continuous working hours, frequency of operational errors, and behavioral postures in videos, and project management pressure inferred by analyzing the frequency of schedule adjustments, number of resource allocation changes, and quality inspection exemption records in management instructions. To address the issues of highly non-stationary data distribution and small effective experimental sample size during water conservancy construction, the causal discovery algorithm of this invention introduces the aforementioned 'prior physical knowledge' and expert rules as strong constraints before performing heuristic search, constructing a blacklist and whitelist for causal graph search to significantly reduce the search space of the network skeleton. Furthermore, the algorithm employs a sliding time window strategy to handle non-stationary drift and uses a G-test with a Laplace smoothing penalty term to mitigate the risk of statistical overfitting caused by sparse and small samples.

[0032] The training process of the variational autoencoder model is as follows: The encoder input is a concatenation of continuous work duration, operational error frequency, and behavioral posture feature vectors (a sequence of keypoint coordinates extracted by the target detection model, with dimensions of 17 x 2 x time step, and a time step of at least 30 frames). The encoder consists of two fully connected layers (256 and 128 neurons respectively), outputting the mean vector and log-variance vector of personnel fatigue (the latent space dimension is 1, i.e., scalar fatigue). The decoder takes the latent fatigue sample value as input and reconstructs the input features using two fully connected layers (128 and 256 neurons respectively). Training employs an unsupervised self-reconstruction objective, with the loss function being the sum of the reconstruction mean square error and the relative entropy, and the relative entropy weight coefficient set to 0.1. The model is pre-trained on historical data from at least 500 complete work shifts; the inferred latent variables are mapped to the 0-1 interval using an S-shaped growth curve function, with 0.7 as the high fatigue warning threshold. The same architecture is used to infer project management stress, but the input features are replaced with time series of schedule adjustment frequency, number of resource allocation changes, and quality inspection exemption records.

[0033] The construction of a contextual causal layer in the non-physical domain based on the causal discovery and latent variable inference model further includes: using the latent variable inference model, taking time series data such as temperature, humidity and strain collected from sensors inside the concrete as input features; and taking the degree of hydration reaction, equivalent age, long-term strength development trend and impermeability of the concrete as latent variables to be inferred; thereby realizing online inference and early prediction of the final performance indicators of large-volume concrete.

[0034] By transforming quality assessment from post-construction local sampling to real-time, global online inference during construction, the problem of severe delays in concrete quality acceptance has been fundamentally changed. By predicting the final strength and impermeability in advance, remedial measures can be proactively taken during the critical curing period to nip quality defects in the bud and significantly reduce later risks and costs.

[0035] Specifically, a physical causal layer is constructed based on prior physical knowledge (including physical law models expressed by mathematical equations and qualitative constraints expressed by engineering mechanism rules). This physical causal layer uses the following physical equations to describe the deterministic causal relationships between nodes: 1. For the temperature rise process of concrete hydration heat, an adiabatic temperature rise model based on equivalent age is used, with cement dosage, water-cement ratio, and ambient temperature at the time of pouring as input nodes, and the temperature rise rate (unit: degrees Celsius per hour) at any location inside the concrete as the output node; 2. For the internal and external temperature difference constraints of large-volume concrete, the highest internal temperature rise and surface heat dissipation coefficient are used as input nodes, and the surface cracking risk (normalized from 0 to 1) is used as the output node, with the threshold determined according to the specification that the internal and external temperature difference should not exceed 25 degrees Celsius; 3. For the influence of additional stress on the retaining structure, the temperature expansion strain of the concrete in the pouring area is used as the input node, and the additional contribution to the displacement of the top of the retaining structure is calculated using the principle of superposition of elasticity, as the output node. All of the above physical relationships are unidirectional causal directions, with no feedback loops. For example, based on heat transfer and the kinetics of concrete hydration heat reaction, a physical model can be established to describe the relationship between cement grade, aggregate ratio, ambient temperature and the rate of temperature rise inside the concrete, and the causal relationship is deterministic.

[0036] For non-physical domains such as human decision-making and management processes, a data-driven approach is used to construct a contextual causal layer. A variational autoencoder model is used to analyze observable indicators to infer latent variables that are difficult to measure directly. For example, by analyzing workers' continuous working hours (e.g., exceeding 8 hours), the frequency of operational errors obtained from equipment operation logs, and behavioral postures in videos (e.g., slow movements, lack of concentration), worker fatigue is quantified as a value between 0 and 1, such as 0.85. By analyzing management instruction data such as the frequency of schedule adjustments and the number of resource allocation changes, a project management stress index is inferred, such as 0.92. The system employs a hybrid driving algorithm combining score search (specifically, the Bayesian information criterion scoring function) and conditional independence testing (specifically, the G² test) to analyze causal feature data containing latent variables. Learning from the data set reveals causal relationships between nodes in the non-physical domain. For example, the algorithm might discover a causal chain like project management pressure → reduced safety training time → increased personnel violation rates. By sharing nodes, the two-layer network is merged. Quantified situational influencing factors output by the situational causal layer, such as personnel fatigue, are introduced into the physical causal layer as a parameter correction node. For example, increased personnel fatigue might correct the vibration operation quality coefficient in the physical model, thus affecting the physical prediction of concrete density. By integrating physical mechanisms and data-driven approaches, a hybrid model is created that can simultaneously understand the deterministic laws of the physical world and the complex randomness of the non-physical world. By inferring latent variables such as personnel fatigue, the model can capture human and management factors that were previously unquantifiable but are the root causes of accidents.

[0037] When the dynamic causal model performs inter-layer fusion, the dimensionless latent variables output by the contextual causal layer (such as personnel fatigue) do not change the structure and material constants of the objective physical law equations in the physical causal layer. Instead, they are transformed into 'operational quality deviation coefficients' as conditional variables in the probabilistic causal network through a preset empirical mapping function. This coefficient is generated based on qualitative constraints constituted by engineering mechanism rules and acts on the final output node of the physical equation in a probabilistic form. For example, an increase in personnel fatigue does not change the heat transfer equation itself, but rather, based on historical empirical statistical laws, increases the variance and down-biased penalty term of the 'expected value of concrete density prediction due to improper vibration,' thereby achieving a quantitative impact of management behavior on physical results while strictly maintaining the dimensional consistency of the physical system.

[0038] Furthermore, the physical causal layer receives and parses the construction data in the BIM model in real time. When a change in the construction phase event is detected, a preset graph structure change script is triggered to modify the nodes and causal relationship edges of the physical causal layer. The contextual causal layer monitors the distribution difference between the newly inflowing causal feature dataset and the historical dataset in real time through the data change monitoring unit. When a significant change in data distribution is detected, the variable nodes directly related to the data change are located and the causal discovery algorithm is re-executed only in the local subgraph region formed by these nodes and their adjacent nodes.

[0039] In the real-time update process of the local subgraph of the contextual causal layer, "real-time" means completion within a set engineering control business cycle (e.g., 15 to 30 minutes). The size of the local subgraph region is strictly limited by the number of hops, specifically, it is a set of adjacent nodes that expand outward along the existing causal path from the variable node directly affected by the significant distribution change as the center, without exceeding two hops. This significant limitation on the size of the relearned nodes allows the causal discovery algorithm to meet the business response time constraints on edge computing servers with limited on-site resources.

[0040] The data change monitoring unit uses a multivariate Kolmogorov-Smirnov test (significance level set at 0.05) to compare the distribution of newly incoming data with that of historical data within the sliding window. Simultaneously, it calculates the difference in Frobenius norm between the covariance matrices of the newly incoming data and historical data. When this norm difference exceeds 30% of the historical variance mean, it is considered a significant distribution change, triggering a local causal relearning mechanism. An update is triggered if either of these two conditions is met.

[0041] Specifically, when the BIM model data shows that the project has moved from the foundation slab pouring stage to the wall construction stage, a preset graph structure change script is triggered to automatically modify the nodes and edges in the physical causal layer. For example, the foundation slab temperature sensor node is deleted while new nodes and causal relationship edges related to the stress on the wall formwork and the status of vertical transportation equipment are added. The data change monitoring unit continuously monitors the new incoming data and calculates the distribution difference with the historical data to update the contextual causal layer in real time. For example, when the project temporarily introduced a batch of new quick-setting concrete to cope with the flood season, the monitoring unit detected that its hydration heat characteristics were significantly different from the historical data and located the variable node related to the concrete material properties. The causal discovery algorithm is then re-executed only in the local subgraph region formed by the node and its adjacent nodes, thereby quickly and efficiently updating the local structure of the contextual causal layer without retraining the entire huge model.

[0042] By endowing the model with adaptive and self-learning capabilities, the problem of rapid failure of traditional static models in dynamic construction environments is solved, ensuring the continuous effectiveness of the causal network and the ability to capture new risks in a timely manner.

[0043] Further, a target node to be predicted is selected according to a preset monitoring strategy, and the actual observation values ​​of evidence nodes that have a causal relationship with the target node are obtained. Using the actual observation values ​​of the evidence nodes as input, a probabilistic causal inference algorithm is executed on the dynamic causal model. Based on the given state of the evidence node, the posterior probability distribution of the target node is calculated using a Bayesian network. The expected value of the posterior probability distribution is extracted and used as the predicted value of the target node in the current causal context. The actual observation values ​​of the target node at the same time are obtained from the causal feature dataset, and the prediction deviation between the actual observation values ​​and the predicted values ​​is calculated. If the prediction deviation exceeds a preset risk threshold, it is determined as a risk event. The preset risk threshold is determined according to the risk level of different monitoring targets. For example, it can be statistically set based on the normal fluctuation range of historical data, or specified by expert experience with reference to relevant industry standards.

[0044] The preset threshold is determined according to the following method: the 95th percentile of the deviation between the predicted value and the actual observed value of the target node during a period in which there are no historical anomalies is taken as the baseline threshold; for monitoring targets with high risk level, the baseline threshold is tightened by multiplying it by a coefficient of 0.8; for monitoring targets with medium risk level, the baseline threshold is used; for monitoring targets with low risk level, the baseline threshold is relaxed by multiplying it by a coefficient of 1.2.

[0045] Specifically, the system selects the risk of concrete surface cracking as the target node to be predicted based on a preset monitoring strategy. It acquires evidence nodes causally related to this target node, such as real-time ambient temperature, wind speed, internal temperature difference, and actual observations of curing measures. These observations are used as input to perform probabilistic causal inference on a dynamic causal model. For example, it uses a message passing algorithm based on a Bayesian network to calculate the posterior probability distribution of the concrete surface cracking risk at the target node and extracts its expected value as the predicted value; for example, the predicted cracking risk probability is 85%. Simultaneously, the system obtains the actual observation value of the node through UAV inspection image analysis or surface strain gauge data. For example, if a microcrack is detected through image recognition algorithms, based on a preset 'state-risk' mapping rule base, the 'microcrack' state is directly mapped to a 95% risk level, thus determining the actual risk to be 95%. Since the prediction deviation is 10%, exceeding the preset risk threshold of 5%, the system determines this to be a risk event requiring immediate intervention.

[0046] This paper presents a deep prediction method based on a causal model. Compared with the traditional alarm method based on a single data threshold, it identifies risks by analyzing the deviation between the predicted and actual values. This method can discover anomalies hidden under complex relationships earlier and more accurately, significantly improving the sensitivity and accuracy of risk identification.

[0047] Furthermore, in the physical causal layer, reverse reasoning is performed. Starting from a risk event with an abnormal physical state, the system traces back along a defined physical law path to identify the direct physical cause leading to the abnormality and outputs it as a physical cause report. Using the direct physical cause as a new starting point for reasoning, reverse reasoning continues in the contextual causal layer. The system traces back along the causal path learned through data-driven learning to identify the root cause located in the non-physical domain and outputs it as a root contextual cause report. The risk attribution report is presented through an interpretable AI interface. This interface visualizes the causal path with the highest contribution from the root cause to the risk event in the dynamic causal model. Each node on the path is assigned a different icon and color according to its modality, and each causal relationship edge is distinguished by its causal strength in terms of thickness or color depth. When the user interactively selects any node or relationship edge, the system will simultaneously display its original data source, feature value, and confidence score of its impact on downstream nodes, presenting the complex causal analysis results in an intuitive and visual way, allowing managers to quickly understand the root cause and transmission path of the risk. Secondly, the interactive traceability function demonstrates the data supporting the AI's conclusions, making the "black box" model transparent and credible, thus enhancing user trust. Ultimately, this helps decision-makers accurately identify key issues, achieve targeted control, and thereby efficiently allocate resources and improve the effectiveness of risk management.

[0048] Specifically, the system first performs reverse reasoning at the physical causal layer. Starting from the physical anomaly of concrete surface cracking, it traces back along a defined physical law path to identify the direct physical causes as excessively rapid surface water loss and excessive internal-external temperature difference, generating a physical cause report. Next, using this direct physical cause as a starting point, it continues to trace back at the situational causal layer, discovering that the root cause of the excessively rapid surface water loss points to the failure to perform maintenance watering operations on time, while the excessive internal-external temperature difference points to untimely insulation layer coverage. Further tracing reveals that the common root cause of both is the omission of night shift handover information and the decreased attention of night shift supervisors due to project management pressure. The system ultimately generates a root cause report and quantifies the contribution of each cause by comparing the difference in probability values ​​before and after intervention; for example, the contribution of negligence in night shift handover is 60%, and the contribution of supervisor fatigue is 40%.

[0049] It enables cross-modal, penetrating attribution analysis, from physical anomalies to the root causes of non-physical problems. This allows risk management to address the root cause and formulate fundamental solutions, avoiding the drawbacks of merely dealing with superficial issues and causing risks to recur.

[0050] Furthermore, risk nodes in the dynamic causal model are identified, and the probability of occurrence of each risk node under the current conditions is calculated based on the probability distribution of the model. The probability values ​​of all risk nodes are then weighted and summed according to their corresponding risk level weights to obtain the construction risk index. The risk level weights are determined using the analytic hierarchy process (AHP) to determine the relative importance of multiple impact dimensions such as safety, quality, schedule, and cost. The quantitative impact of specific risks on these dimensions is then weighted and summed.

[0051] The construction risk index is calculated as follows: Assume there are N risk nodes in the dynamic causal model. The posterior probability of the i-th risk node under the current observation conditions is a value inferred from a Bayesian network (range 0 to 1). The components of this risk node in the four influence dimensions of safety, quality, schedule, and cost are all processed by max-min normalization to the range of 0 to 1. The weight vectors of the four dimensions of safety, quality, schedule, and cost are calculated through pairwise comparison judgment matrices and determined after verification with a consistency ratio of less than 0.1. Then, the weighted risk level of the i-th risk node is equal to the safety weight multiplied by the safety impact degree, plus the quality weight multiplied by the quality impact degree, plus the schedule weight multiplied by the schedule impact degree, plus the cost weight multiplied by the cost impact degree. The construction risk index is equal to the sum of the products of the posterior probability of all N risk nodes in the dynamic causal model and their corresponding weighted risk levels. The index ranges from 0 to 1; a larger value indicates a higher overall construction risk.

[0052] The project uses a dynamic causal model to perform forward-looking simulations, calculating the expected change in the construction safety risk index after the implementation of control measures. This also includes assessments of potential secondary risks and resource impacts that the control measures may trigger. The secondary risk assessment refers to the simultaneous calculation of risk index changes for other non-target nodes in the causal network during the forward-looking simulation. The resource impact assessment refers to evaluating the impact of the measures on construction costs and schedule. The optimal control measures are determined based on a multi-objective optimization function aimed at minimizing the risk index, secondary risks, and resource impacts.

[0053] The optimal control measure is determined using the following multi-objective decision-making method: For each candidate measure, three target values ​​are calculated after its implementation through forward-looking simulation using a dynamic causal model: the expected reduction in the construction risk index, the increase in secondary risk, and the resource impact. Each measure is scored using a weighted linear comprehensive score, which equals weight 1 multiplied by the expected reduction in the construction risk index, minus weight 2 multiplied by the increase in secondary risk, and then minus weight 3 multiplied by the resource impact, where the sum of the three weights is 1, determined according to the preset project safety priority configuration (default values ​​are 0.6, 0.3, and 0.1 respectively). The candidate measure with the highest score is the optimal control measure.

[0054] The prospective extrapolation using a dynamic causal model is based on the engineering premise that the dynamic causal network already includes the main observed variables affecting core security objectives (i.e., approximately satisfying the assumption of no unobservable confounding). This extrapolation calculates the expected change trend of the target nodes through do-calculus intervention. The output expected change in the risk index is used to provide a quantitative relative preference ranking among multiple control measures, rather than promising an absolute statistical probability regression in the physical world. The extrapolation results also output a 95% confidence interval based on Monte Carlo sampling.

[0055] Specifically, for the most significant issue of negligence during night shift handover, the system retrieves corresponding control measures from a pre-set knowledge base, such as: A. Implementing a mandatory electronic handover checklist; B. Adding a dedicated night shift coordinator; C. Holding a safety awareness meeting for all night shift personnel. The system uses a dynamic causal model for forward-looking simulation: Assessment of Measure A: Simulation shows that the construction risk index is expected to decrease by 25% after implementation. However, a secondary risk assessment reveals that it may lead to worker resistance due to unfamiliarity with the new system. The resource impact assessment shows that software procurement and training are needed, increasing costs by approximately 50,000 yuan, but without affecting the project schedule. Assessment of Measure B: Simulation shows that the construction risk index is expected to decrease by 18%, with low secondary risk, but the resource impact assessment shows an increase in monthly labor costs of 30,000 yuan. Assessment of Measure C: Simulation shows that the construction risk index is expected to decrease by 8%, with minimal effect. Through a multi-objective optimization function aimed at minimizing the risk index, minimizing secondary risk, and minimizing resource impact, and by weighting the sum according to pre-set project priority weights, the system determines Measure A as the optimal control measure.

[0056] By introducing counterfactual reasoning capabilities, the problem of traditional risk response relying heavily on personal experience and lacking scientific decision-making basis is solved. By quantitatively comparing and comprehensively evaluating different measures, including secondary risks and resource impacts, it ensures that the generated control measures are the globally optimal solution, elevating risk management from passive response to a new level of intelligent decision-making.

[0057] Furthermore, a closed-loop management instruction for the construction period is generated based on the optimal control measures, and the instruction is sent to the corresponding management system or personnel terminal. The closed-loop management instruction includes a dynamic digital work permit that is linked in real time with the on-site personnel authentication system and the equipment management system. When the instruction is sent, it includes the minimum personnel qualification level, specific equipment model, and health status threshold required to execute the instruction. Before performing the operation, on-site personnel need to scan the equipment ID and personal identification ID through a mobile terminal to request activation of the permit. The system backend will verify in real time whether the personnel qualification and equipment status meet the instruction requirements. Only when all conditions are fully matched will the digital work permit be authorized and activated, and the operation permission of the corresponding equipment be unlocked.

[0058] Specifically, the system automatically generates a closed-loop management instruction for the construction period: "Immediately activate the 'Night Shift Electronic Handover List' process, and require all night shift team leaders to complete the confirmation on the booth terminal before leaving their posts. Instruction responsible person: Project Chief Engineer Wang". This instruction is sent to the project management system and the mobile terminals of relevant personnel; at the same time, this instruction can be linked to a dynamic digital work permit. For example, for the subsequent high-altitude formwork installation operation, the instruction includes the minimum personnel qualification levels required to perform this task, such as senior scaffolders, specific equipment models, such as self-ascending scaffolds, and health status thresholds, such as equipment intact rate > 95%. Before the on-site personnel start the operation, they must scan their personal ID and equipment ID through the mobile terminal; the system background verifies the personnel qualifications and equipment status in real time. Only when all conditions are fully matched can the permit be authorized and activated, and the operation permissions of the corresponding equipment be unlocked, so as to accurately implement the intelligent decision to every operation link in the physical world, forming a management closed loop.

[0059] It打通了the "last mile" from the intelligent analysis of the digital twin to the precise execution in the physical field, ensuring that the optimal decision can be implemented without loss, in real time, and compulsorily. By linking with personnel and equipment authentication, it杜绝了the occurrence of non-compliant behaviors at the operation level, truly achieving a management closed loop.

[0060] The present invention constructs a complete intelligent management and control loop from data fusion, dynamic modeling, risk prediction, deep traceability to optimization decision-making and closed-loop execution, fundamentally solving the three major pain points of superficial risk awareness, rigid models, and blind decision-making, and realizing the proactive, precise, and forward-looking management of construction risks.

[0061] Embodiment 2: This embodiment aims to elaborate on the specific application process and technical advantages of the present invention in dealing with dynamic and unexpected changes at the construction site, especially when key equipment is temporarily replaced. The scenario setting continues the large-volume concrete pouring operation of the pump station foundation pit in Embodiment 1.

[0062] As an implementation manner of the present invention, refer to Figure 1 , a flowchart of a quality and safety management method for a water conservancy construction project, refer to Figure 2 , a schematic structural diagram of a quality and safety management system for a water conservancy construction project, refer to Figure 3 , a flowchart of the specific implementation method for the dynamic update of the dynamic causal model.

[0063] It should be noted that there are some Chinese words in the original text that seem to be incomplete or incorrect in the "打通了" and "杜绝了" parts. I have translated them as best as possible according to the context, but it may need further clarification in the original content.When the continuous concrete pouring operation reached a critical stage, the main A-type concrete pump truck, which was originally planned to be used and whose data had been connected to the system, suddenly suffered a hydraulic system failure and could not continue to work. In order to avoid cold joints in the construction and ensure the quality of the project, the project team urgently rented a spare B-type pump truck from outside. The B-type pump truck was an old model, and its working efficiency, energy consumption characteristics, sensor interface and data format were significantly different from the original A-type pump truck. Moreover, its historical operating data was not included in the initial training dataset of the system.

[0064] The system first determines that the main equipment has ceased operation based on the interruption of the IoT data stream and the "serious fault" code reported by the Type A pump truck's onboard terminal. The Type B pump truck is then urgently deployed to the site and begins operation. The causal modeling module activates: its onboard sensors perform protocol parsing and normalization through an edge computing gateway deployed on-site, and new data streams begin to flow into the system's causal feature dataset. The data change monitoring unit analyzes the newly incoming Type B pump truck data in real time and compares it with historical datasets primarily composed of Type A pump truck data. It quickly identifies significant changes in the data distribution related to variables such as pumping pressure, motor load, and hydraulic oil temperature; for example, significant shifts occur in the mean, variance, and covariance matrix between variables. This distribution change triggered an incremental learning mechanism. Instead of performing a global, time-consuming retraining of the dynamic causal model, the system precisely located the variable nodes directly related to the construction machinery and equipment entity, such as equipment status, operating efficiency, and energy consumption, and their adjacent nodes, forming a local subgraph. The causal discovery algorithm was then re-executed only for this local subgraph region using the newly inflowing B-type pump truck data. Through this rapid local update process, the model learned the causal relationships specific to the B-type pump truck. For example, the model found that for the B-type pump truck, the negative causal effect strength coefficient of hydraulic oil temperature on pumping stability was three times that of the A-type pump truck; and the correlation probability between outlet pressure fluctuation and pipeline blockage risk was 40% higher than that of the A-type pump truck.

[0065] After the local causal network was updated, the risk prediction module reassessed the current pouring operation based on the model adapted to the new equipment. The model predicted that continuing to use the Type B pump truck under the current conditions would increase the probability of pump pipe blockage from 5% to 35%, and the total pouring completion time would be extended by 1.5 hours, potentially leading to insufficient initial setting time of concrete in some areas and creating new quality hazards. Based on the risk reassessment results, the system proactively generated and issued a warning message: "Equipment replacement detected. The new equipment has low operational stability. The risk of pipe blockage is expected to increase significantly within the next 2 hours. Preventive measures are recommended." To address the predicted high risk of pipe blockage, the risk management module automatically retrieves information from the knowledge base and performs a forward-looking simulation: Measure A: Adjust the concrete mix proportions, instructing the batching plant to appropriately increase the amount of admixtures to improve concrete fluidity; Simulation result: The risk of pipe blockage can be reduced to 10%, but this will increase material costs by approximately 5% and may have a slight impact on later strength development; Measure B: Optimize pumping operation procedures, instructing on-site operators to reduce pumping speed and increase pipeline lubrication frequency; Simulation result: The risk of pipe blockage can be reduced to 15%, but this will further reduce construction efficiency and is expected to extend the total construction period by an additional 45 minutes; Through multi-objective optimization... After calculating and comprehensively assessing the risk reduction effect, cost, and schedule impact, the system determines that measure A is the optimal control measure for the current scenario. The system immediately generates and issues instructions to relevant parties. The instructions issued to the concrete mixing plant are to adjust the concrete mix ratio for a specific pouring location. The instructions issued to the on-site quality inspectors are to strengthen the inspection of the adjusted concrete condition and temperature monitoring. When facing the construction challenges of sudden equipment changes, the system transforms unexpected disturbances into a controllable and manageable process through rapid self-learning, accurate risk assessment, and scientific decision optimization, thereby maximizing the continuity, quality, and safety of construction.

[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for quality and safety management in water conservancy construction projects, characterized in that, include: Acquire real-time and historical multi-source heterogeneous data of construction projects, and generate causal feature datasets by constructing a knowledge graph containing physical domain and non-physical domain nodes; Based on prior physical knowledge, a physical causal layer is constructed for the physical domain, and a contextual causal layer for the non-physical domain is constructed based on causal discovery and latent variable inference models. The two layers are then integrated into a dynamic causal model. When the engineering stage changes, nodes and causal relationship edges in the physical causal layer are dynamically added, deleted, and reconstructed. For newly incoming non-physical feature parameters, an incremental learning mechanism is used to update the local subgraph of the contextual causal layer in real time. Based on the causal feature dataset, a dynamic causal model is used to generate predicted values ​​and calculate prediction bias. When the prediction bias exceeds the threshold, it is determined to be a risk event. By tracing the source across modalities from the anomaly point in the physical domain to the non-physical domain, symptom points are generated and their contribution is quantified. For the symptom point with the highest contribution, the corresponding control measures in the preset knowledge base are obtained. The expected change in the construction risk index after taking control measures is evaluated through prospective extrapolation, and the optimal control measures are generated.

2. The method for quality and safety management of water conservancy construction projects according to claim 1, characterized in that, The physical domain refers to entities, materials, and processes with measurable physical properties at the construction site; the non-physical domain refers to abstract concepts, events, and attributes related to human decision-making, behavior, and management processes; the multi-source heterogeneous data includes: a four-dimensional BIM model containing construction schedule plans and temporary engineering structures, various sensor data monitoring concrete hydration heat, structural stress, displacement, strain, and acceleration, image data capturing construction machinery trajectories and personnel behavior, and text data recording engineering design, construction, materials, equipment, environment, and personnel information; for unstructured image data and text data, a pre-trained model is invoked to identify and extract the engineering entities, construction activities, management events, and status attributes contained therein; Based on a pre-built knowledge graph ontology of pump station water conservancy construction projects, engineering entities, construction activities, state attributes, and structured design data and physical quantity time series data are associated and mapped. Causal feature datasets are generated by transforming data from different sources into entity nodes, event nodes, and relation edges with unique identifiers and timestamps in the knowledge graph.

3. The quality and safety management method for water conservancy construction projects according to claim 1, characterized in that, A physical causal layer is constructed using a physical model that describes deterministic relationships between physical variables; a contextual causal layer is constructed using a data-driven probabilistic graphical model, where nodes correspond to non-physical domains in the knowledge graph. Causal relationship edges and probability parameters are obtained through a causal discovery algorithm on a causal feature dataset. The two layers are fused through shared nodes, with the contextual causal layer outputting quantified contextual influence factors as parameter correction nodes to adjust the corresponding parameters in the physical causal layer. The causal discovery algorithm combines scoring search and conditional independence testing, using a Bayesian information criterion scoring function to generate candidate causal network skeletons. For each causal relationship edge in the candidate skeleton, the G² test is used to verify the authenticity of the relationship and eliminate false relationships. The latent variable inference model infers the state of latent variables by analyzing the collaborative change patterns of observable causal features and using a variational autoencoder.

4. The quality and safety management method for water conservancy construction projects according to claim 1, characterized in that, The specific implementation methods for dynamically adding, deleting, and reconstructing nodes and causal relationship edges in the physical causal layer include: receiving and parsing construction data in the BIM model in real time; triggering a graph structure change script associated with the event when a change in a construction phase event is detected; and performing add, delete, and modify operations on nodes and causal relationship edges in the dynamic causal model according to preset rules. The specific implementation methods for real-time updating of local subgraphs in the contextual causal layer include: establishing a data change monitoring unit to monitor the distribution differences between newly incoming causal feature datasets and historical datasets in real time; and when a significant change in data distribution is detected, locating the variable nodes directly related to the data change and re-executing the causal discovery algorithm only on the local subgraph region formed by the variable nodes and their adjacent nodes.

5. The quality and safety management method for water conservancy construction projects according to claim 1, characterized in that, The target node to be predicted is selected according to a preset monitoring strategy. The actual observation values ​​of evidence nodes that are causally related to the target node are obtained. The actual observation values ​​of the evidence nodes are used as input to execute a probabilistic causal inference algorithm on the dynamic causal model. Based on the given state of the evidence node, the posterior probability distribution of the target node is calculated through a Bayesian network. The expected value of the posterior probability distribution is extracted and used as the predicted value of the target node in the current causal context. The actual observation value of the target node at the same time is obtained from the causal feature dataset, and the prediction deviation between the actual observation value and the predicted value is calculated. If the prediction deviation exceeds a preset risk threshold, it is determined to be a risk event.

6. The method for quality and safety management of water conservancy construction projects according to claim 1, characterized in that, In the physical causal layer, reverse reasoning is performed. Starting from the risk event of physical anomaly, the direct physical cause leading to the anomaly is identified and output as a physical cause report along the determined physical law path. Using the direct physical cause as a new starting point for reasoning, reverse reasoning is performed again in the situational causal layer. The root cause leading to the non-physical domain is identified and output as a root situational cause report along the causal path learned by data-driven learning.

7. The quality and safety management method for water conservancy construction projects according to claim 1, characterized in that, The construction risk index is calculated as follows: risk nodes in the dynamic causal model are identified, the probability of occurrence of each risk node under the current conditions is calculated based on the probability distribution of the model, and the probability values ​​of all risk nodes are weighted and summed according to their corresponding risk level weights to obtain the construction risk index. The risk level weights are determined using the analytic hierarchy process (AHP) to determine the relative importance of the four impact dimensions of safety, quality, schedule, and cost, and the quantitative impact of specific risks on these dimensions is weighted and summed. The expected change in the construction safety risk index after the implementation of control measures is calculated through forward-looking simulation using a dynamic causal model. This also includes assessments of secondary risks and resource impacts caused by the control measures. The secondary risk assessment refers to the simultaneous calculation of risk index changes of other non-target nodes in the causal network during the forward-looking simulation. The resource impact assessment refers to the assessment of the impact of the measures on construction costs and schedule. The optimal control measures are determined based on a multi-objective optimization function aimed at minimizing the risk index, minimizing secondary risks, and minimizing resource impacts.

8. The quality and safety management method for water conservancy construction projects according to claim 1, characterized in that, Based on the optimal control measures, a closed-loop management instruction for the construction period is generated and issued to the corresponding management and personnel terminals. The closed-loop management instruction includes a dynamic digital work permit that is linked in real time with on-site personnel authentication and equipment management. When the instruction is issued, it includes the minimum personnel qualification level, specific equipment model, and health status threshold required to execute the instruction. Before performing the operation, on-site personnel need to scan the equipment ID and personal identification ID through a mobile terminal to request permission activation. The backend will verify in real time whether the personnel qualification and equipment status meet the instruction requirements. Only when all conditions are fully matched will the digital work permit be authorized and activated, and the operation permission of the corresponding equipment be unlocked.

9. A quality and safety management system for water conservancy construction projects, characterized in that, include: Causal modeling module: Acquires multi-source heterogeneous data from the construction site and generates a causal feature dataset by constructing a knowledge graph containing nodes from both the physical and non-physical domains; Based on prior physical knowledge, a physical causal layer is constructed for the physical domain, and a contextual causal layer is constructed for the non-physical domain based on causal discovery and latent variable inference models. The two layers are then integrated into a dynamic causal model. Risk prediction module: When the engineering stage changes, the physical causal layer nodes and causal relationship edges are dynamically added, deleted and reconstructed; for newly incoming non-physical feature parameters, an incremental learning mechanism is used to update the local subgraph of the contextual causal layer in real time; the dynamic causal model is used to generate predicted values ​​and calculate prediction deviations; when the prediction deviation exceeds the threshold, it is judged as a risk event. Risk management module: Starting from anomalies in the physical domain through hierarchical causal tracing, backtracking across modalities to non-physical domains to generate symptom points and quantify their contribution; for the symptom points with the highest contribution, obtain the corresponding control measures from the preset knowledge base, and through forward-looking extrapolation, evaluate the expected change in the construction risk index after taking control measures to generate the optimal control measures.

Images