Water conservancy project quality safety supervision risk assessment system

Abstract

The invention relates to the field of risk assessment, in particular to a water conservancy project quality safety supervision risk assessment system. The method comprises the specific implementation steps that firstly, nonlinear interaction of risk factors in multi-source heterogeneous data is converted into quantitative risk numerical values through a risk quantification model, and the quantitative risk numerical values and context factors are packaged into risk state vectors; then, weights are distributed for the risk state vectors of all the completed engineering stages before, risk fusion vectors are constructed by fusing the weighted risk state vectors and heterogeneous data of the current engineering stage, and then the future risk evolution trend of the water conservancy project structure is predicted. When the prediction trend deviates from the safe trajectory, the system triggers early warning in advance and automatically backtracks the model to find out the main risk factor. In order to ensure the reliability of early warning, the system generates supervision robustness at the same time, and when the supervision robustness is lower than a preset threshold value, secondary assessment of the risk is automatically initiated according to the main risk factors and the robustness.

Description

Technical Field

[0001] This invention relates to the field of risk assessment, specifically to a risk assessment system for quality and safety supervision of water conservancy projects. Background Technology

[0002] In the field of quality and safety supervision and risk assessment of water conservancy projects, accurate risk assessment is crucial to ensuring project safety and smooth implementation. However, traditional assessment methods struggle to handle the massive and diverse data generated in modern engineering projects, such as heterogeneous data from sensors, construction reports, and video surveillance. This leads to biased and outdated assessment results, preventing managers from obtaining a comprehensive view of project risks. While multi-factor fusion technology has been introduced to address the data processing challenges, its inherent limitations remain a significant challenge for the industry.

[0003] Existing methods have some problems in practical applications. Traditional methods lack dynamic processing capabilities for data fusion, and usually only fuse data according to preset fixed weights. When sudden risks occur in the project, the system cannot effectively capture the risks. When data flow is interrupted or data source conflicts occur, traditional methods cannot effectively integrate information, which leads to delayed warnings or even complete errors, seriously affecting the effectiveness of management decisions.

[0004] In summary, a method is needed that can receive and analyze multi-source data input in real time, provide a basis for risk assessment based on rich data, and generate an adaptive assessment and early warning strategy for each stage of the project, thereby solving the assessment lag problem of existing methods and ensuring the reliability and proactivity of risk management.

[0005] To address this, a risk assessment system for quality and safety supervision of water conservancy projects is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a risk assessment system for quality and safety supervision of water conservancy projects. To address the problems of existing methods, this invention first acquires heterogeneous data and contextual factors from different stages of the project. It then processes this data using a risk quantification model to identify and extract risk factors. Next, by analyzing the nonlinear interactions between risk factors, their synergistic effects are transformed into a risk value, which is then encapsulated with contextual factors into a multidimensional risk state vector. The system then constructs a risk assessment model, using the heterogeneous data of the current stage and the risk state vectors of all previously completed project stages as input. Weights are assigned to historical vectors and fused with the current data to obtain a comprehensive risk fusion vector. Subsequently, the system predicts the future risk evolution trend of the water conservancy project structure based on this fusion vector. When the predicted trend deviates from the safety trajectory, the system immediately triggers an early warning and simultaneously backtracks the model to identify the main risk factors causing the predicted risk, outputting the monitoring robustness. Finally, if the monitoring robustness is lower than a preset threshold, the system automatically reassesses the risk based on the main risk factors and the monitoring robustness, thereby ensuring the reliability of the early warning.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a risk assessment system for quality and safety supervision of water conservancy projects, the method comprising: Risk Quantification Module: Acquire heterogeneous data and contextual factors from each stage of the project, construct a risk quantification model, take heterogeneous data as input, extract risk factors, analyze the nonlinear interaction between risk factors, transform the synergistic effect of risk factors into risk values, and encapsulate the risk values ​​with contextual factors to obtain a risk state vector. Dynamic assessment module: Based on long short-term memory network, a risk assessment model is constructed. Heterogeneous data of the current engineering stage and risk state vectors of previous engineering stages are used as input. Weights are assigned to the risk state vectors and fused with heterogeneous data to obtain risk fusion vectors. The risk evolution trend of the engineering structure is predicted based on the risk fusion vectors. When the risk evolution trend deviates from the safe trajectory, an early warning is triggered. Risk Adaptive Module: SHAP retrospective risk assessment model identifies the main risk factors that lead to the predicted risk and outputs the supervisory robustness. When the supervisory robustness is lower than the preset threshold, the risk is reassessed based on the main risk factors and the supervisory robustness.

[0008] Preferably, the risk quantification model is constructed, and heterogeneous data is used as input to extract risk factors. The specific implementation process includes: inputting heterogeneous data from each engineering stage into the risk quantification model, wherein the heterogeneous data includes text data, time series data, and visual data; using an NLP-based text feature extraction sub-model, a time series data analysis-based sensor data feature extraction sub-model, and a deep learning-based visual feature extraction sub-model to process the text data, time series data, and visual data respectively, and identifying and extracting key risk factors.

[0009] Preferably, the process of converting the synergistic effect of risk factors into a risk value by analyzing the nonlinear interaction between risk factors includes: mapping the identified risk factors to a coordinate point in a multidimensional space and inputting it into a deep learning sub-model based on a multilayer perceptron; using feature engineering to encode the risk factors into a unified numerical vector; capturing the relationship between risk factors by creating interactive features; performing nonlinear processing on the risk factors through the nonlinear function of the hidden layer; and outputting a quantitative risk value between 0 and 1.

[0010] Preferably, the specific implementation process of encapsulating risk values ​​and contextual factors to obtain a risk state vector includes: the contextual factors include: project completion time, input cost, environment, number of construction personnel and equipment failure rate; all contextual factors are standardized and mapped to the range of 0 to 1; the quantified risk value is used as the first dimension of the vector, and all standardized contextual factors are concatenated as subsequent dimensions to obtain the risk state vector of the project stage.

[0011] Preferably, a risk assessment model is constructed based on a long short-term memory network. This model uses heterogeneous data from the current engineering phase and risk state vectors from previous engineering phases as input. Weights are assigned to the risk state vectors, and the model is then fused with the heterogeneous data to obtain a risk fusion vector. The specific implementation process includes: constructing a risk assessment model based on a long short-term memory network; inputting the risk state vectors of all completed engineering phases prior to the current engineering phase into the risk assessment model; using the gating mechanism of the long short-term memory network to learn and assign weights to each risk state vector; multiplying all risk state vectors from previously completed engineering phases by their corresponding weights to obtain a weighted vector; and then constructing a complete sequence with the heterogeneous data from the current engineering phase in chronological order, concatenating the two sequences to obtain the risk fusion vector for the current engineering phase.

[0012] Preferably, the risk evolution trend of the engineering structure is predicted based on the risk fusion vector, and an early warning is triggered when the risk evolution trend deviates from the safety trajectory. The specific implementation process includes: inputting the risk fusion vector into the risk assessment model to predict the risk value at the first future time point; using the prediction result and new real-time acquired heterogeneous data as new inputs to predict the risk value at the next future time point; obtaining a risk value sequence composed of discrete time points through iterative prediction; plotting the predicted risk value sequence into a curve to obtain the predicted risk evolution trend; comparing each time point on the prediction curve; and immediately triggering an early warning when there is a predicted risk value higher than a preset safety threshold at any of the time points.

[0013] Preferably, the SHAP backtesting risk assessment model, which identifies the main risk factors leading to predicted risk and outputs supervised robustness, specifically includes the following steps: Backtesting the risk fusion vector used for prediction using the SHAP algorithm; calculating the Shapley value for each dimension of the risk fusion vector; sorting all Shapley values ​​by size; analyzing the contribution of each risk factor in the risk fusion vector to the prediction result; identifying the main risk factors; subjecting the current input risk fusion vector to multiple random perturbations; inputting the perturbed new vector into the risk assessment model for prediction; calculating the standard deviation of the prediction result to quantify model stability; calculating the concentration of attribution results based on the contribution of the main risk factors to the prediction result using the Gini coefficient to quantify attribution clarity; and linearly weighting and fusing the quantified indicators to generate a supervised robustness between 0 and 1.

[0014] Preferably, when the supervisory robustness is lower than a preset threshold, the risk is reassessed based on the main risk factors and the supervisory robustness. The specific implementation process includes: setting a threshold according to the project type, comparing the generated supervisory robustness with the preset threshold, and when the supervisory robustness is lower than the preset threshold, identifying the cause of the low supervisory robustness based on the quantitative results of model stability and attribution clarity, sending data requests to the sensors and relevant personnel based on the cause, requesting supplementary data, and reassessing the risk.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention significantly improves the robustness and accuracy of risk assessment, overcoming the limitation of unstable assessment results in complex environments. Traditional risk assessment systems are prone to distorted results due to their inability to effectively handle incomplete data, conflicting data sources, or lack of analysis of nonlinear interactions. This invention constructs a risk quantification model, utilizes sub-models based on NLP, time series analysis, and deep learning to process heterogeneous data, and employs a multilayer perceptron to capture the nonlinear interactions between risk factors, transforming the synergistic effects of risk factors into unified quantitative values. Therefore, risk warnings will be based on more reliable and accurate quantitative information, ensuring the effectiveness of the assessment from the outset.

[0016] 2. Achieve efficient and stable prediction and early warning of future risk trends, solving the problem of easy interruption in existing technology assessments. Existing methods typically rely on static models for risk assessment, which are easily affected by sudden changes in the current environment and insufficient historical data, thus failing to achieve stable predictions. This invention constructs a risk assessment model based on Long Short-Term Memory (LSTM) networks, uses a gating mechanism to assign weights to the risk state vectors of completed engineering phases, and fuses them with heterogeneous data vectors of the current engineering phase to obtain a risk fusion vector for the current engineering phase. The risk evolution trend is predicted based on the risk fusion vector. When the predicted risk evolution trend deviates from the safe trajectory, the system immediately triggers an early warning, ensuring that the system can maintain stable predictions and issue early warnings even when risks change dynamically.

[0017] 3. This invention constructs an adaptive correction method based on supervised robustness. By using the SHAP algorithm to backtrack the model, the main risk factors causing the risk are identified. Simultaneously, the model stability and attribution clarity are quantified, thereby generating supervised robustness. When this robustness falls below a threshold, the system no longer passively waits but automatically requests supplementary data and initiates a secondary evaluation process based on the cause. This allows the system to quickly and efficiently reconfirm and lock in the risk, greatly improving the continuity and autonomy of the system in performing risk assessment and early warning in complex and dynamic tasks, ensuring the successful completion of the task. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a risk assessment system for quality and safety supervision of water conservancy projects; Figure 2 A schematic diagram of a risk assessment system for quality and safety supervision of water conservancy projects; Figure 3 This is a flowchart illustrating the risk quantification model proposed in this 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 Figure 1 - Figure 3 This invention proposes a risk assessment system for quality and safety supervision of water conservancy projects, the technical solution of which is as follows: A risk assessment system for quality and safety supervision of water conservancy projects, with reference to Figure 1 The specific implementation steps of the method proposed in this invention include: Heterogeneous data and contextual factors from each stage of the project are acquired, and a risk quantification model is constructed. Heterogeneous data is used as input to extract risk factors. By analyzing the nonlinear interaction between risk factors, the synergistic effect of risk factors is transformed into risk values. Risk state vectors are obtained by encapsulating risk values ​​with contextual factors. A risk assessment model is constructed based on long short-term memory networks. Heterogeneous data from the current engineering stage and risk state vectors from previous engineering stages are used as inputs. Weights are assigned to the risk state vectors, and they are fused with the heterogeneous data to obtain a risk fusion vector. The risk evolution trend of the engineering structure is predicted based on the risk fusion vector, and an early warning is triggered when the risk evolution trend deviates from the safe trajectory. The SHAP retrospective risk assessment model identifies the main risk factors that lead to the predicted risk and outputs the supervisory robustness. When the supervisory robustness is lower than a preset threshold, the risk is reassessed based on the main risk factors and the supervisory robustness.

[0021] Example 1 This embodiment provides a specific application of a risk assessment system for quality and safety supervision in water conservancy projects. A typical application scenario is the risk assessment of quality and safety supervision during the construction phase of large-scale water conservancy projects. The specific steps proposed in this invention include: S1. Acquiring heterogeneous data and contextual factors from each engineering stage, constructing a risk quantification model, and extracting risk factors using the heterogeneous data as input; S2. Analyzing the nonlinear interactions between risk factors, converting the synergistic effect of risk factors into risk values; and encapsulating the risk values ​​with contextual factors to obtain a risk state vector; S3. Constructing a risk assessment model based on a long short-term memory network, using the heterogeneous data of the current engineering stage and the risk state vectors of previous engineering stages as input, assigning weights to the risk state vectors, and fusing them with the heterogeneous data to obtain a risk fusion vector; S4. Predicting the risk evolution trend of the engineering structure based on the risk fusion vector, and triggering an early warning when the risk evolution trend deviates from the safety trajectory; S5. Using SHAP to backtrack the risk assessment model, identifying the main risk factors leading to the predicted risk and outputting the supervision robustness; S6. When the supervision robustness is lower than a preset threshold, reassessing the risk based on the main risk factors and the supervision robustness.

[0022] Furthermore, heterogeneous data and contextual factors from each stage of the project are acquired, and a risk quantification model is constructed. Using the heterogeneous data as input, risk factors are extracted. Corresponding to step S1 above, the specific process is as follows: The system acquires heterogeneous data and contextual factors in real time across different engineering phases. For example, it obtains time-series data, such as settlement rate and seepage pressure, from the sensor network at the construction site; textual data, such as daily work logs and safety inspection reports, from the construction management system; and visual data, such as images of concrete surfaces and panoramic views of the construction area, from regular drone inspections. Simultaneously, it collects contextual factors, including engineering phases such as foundation excavation, main structure pouring, input costs, number of construction personnel, and equipment failure rates.

[0023] The system utilizes multiple sub-models of the risk quantification model to process data in parallel. It employs an NLP-based text feature extraction sub-model to analyze logs, identifying key risk factors such as "hot weather" and "exposed rebar." For time-series data, the system uses a time-series data analysis sub-model to detect anomalies; for example, when the settlement rate accelerates abnormally within a short period, the system identifies it as the risk factor "abnormally accelerated settlement." The system also uses a ResNet-based object detection model to automatically identify the "concrete cracks" risk factor in visual data.

[0024] By employing NLP, time-series analysis, and deep learning sub-models, the system leverages textual, time-series, and visual data, breaking down data silos. This enables the system to better capture risk signals, ensuring that the extracted risk factors are more comprehensive and accurate, thus providing a solid data foundation for subsequent assessments.

[0025] Furthermore, by analyzing the nonlinear interactions between risk factors, the synergistic effect of risk factors is transformed into a risk value; by encapsulating the risk value with contextual factors, a risk state vector is obtained, corresponding to step S2 above. The specific process is as follows: The identified risk factors and contextual factors are input into a deep learning sub-model based on a multilayer perceptron (MLP) to capture the nonlinear interactions between all factors. For example, "abnormally accelerated settlement" alone might be classified as low risk (e.g., 0.2), and "hot weather" alone might also be classified as low risk (e.g., 0.1). However, when these factors occur simultaneously with the "main body pouring" stage, the hidden layers of the MLP can identify a high-risk combination that leads to impaired concrete strength due to the combined effect of all three. After nonlinear processing, the model outputs a quantified risk value between 0 and 1. For example, after MLP processing, the risk value of the above combination might be quantified as 0.88.

[0026] All contextual factors are standardized and mapped to a range of 0 to 1. The system uses the quantified risk value as the first dimension of the vector and then concatenates all the standardized contextual factors as subsequent dimensions. For example, concatenating the risk value of 0.88 with standardized contextual factors, such as the number of personnel (0.7) and equipment failure rate (0.5), yields the current risk state vector: [0.88, 0.7, 0.5].

[0027] Specifically, the contextual factors are standardized, and the process is as follows: Recurrent neural networks are used to process dynamic factor time-series data and output trend vectors. Learnable high-dimensional embedding vectors are assigned to discrete factors and dynamically adjusted during model training to capture the real impact of discrete factors on risk. All contextual factor vectors are concatenated to form a contextual fusion vector.

[0028] All heterogeneous contextual factors are uniformly quantized into high-dimensional, compact numerical vectors. This solves the problem of inconsistent data formats, enabling all contextual information to be seamlessly concatenated with risk factor vectors from other heterogeneous data to form a standardized, comprehensive "contextual fusion vector." This greatly simplifies the data preprocessing process and improves the processing efficiency of the entire risk assessment system.

[0029] In complex water conservancy projects, many risks are not caused by a single factor, but rather by the cumulative effect of multiple seemingly unrelated factors in a specific combination. By using deep learning models to capture the hidden and complex relationships between factors, quantitative risk values ​​can more accurately reflect the risk level and avoid the distortion of assessments caused by independent analysis.

[0030] By standardizing and integrating different types of risk values ​​and contextual factors, the problem of inconsistency across different data dimensions was solved. This also provides a standardized input format for subsequent deep learning models, greatly improving the efficiency and accuracy of model processing.

[0031] Furthermore, a risk assessment model is constructed based on a long short-term memory network. Heterogeneous data from the current engineering stage and risk state vectors from previous engineering stages are used as inputs. Weights are assigned to the risk state vectors, and they are fused with the heterogeneous data to obtain a risk fusion vector. This corresponds to step S3 above, and the specific process is as follows: A risk assessment model is built based on Long Short-Term Memory (LSTM) networks. This model can handle time series and utilizes its unique gating mechanism to intelligently learn and remember the importance of historical information. At each time step, the LTM network determines how much information from past stages to retain through a forgetting gate; it determines how much new information from the current stage to consider through an input gate; and finally, it generates a state vector through an output gate. This gating mechanism effectively assigns a dynamic weight to each historical risk state vector. For example, the system might assign higher weights to the vectors from day 179 and day 180, such as... =0.9, =0.95, while assigning lower weights to the earlier days 100 and 101, such as =0.4, =0.42, thus ensuring that the model always focuses on the historical data most relevant to the current stage.

[0032] The system stores the risk state vectors generated for each completed stage of the dam's construction since its commencement. For example, the system stores a sequence of risk state vectors for the past 180 days, denoted as... , ,..., On the same day, heterogeneous data vectors for the current stage are acquired in real time. These vectors are then processed using a risk quantification model to obtain a risk state vector, denoted as... The risk state vectors are input into the risk assessment model, weights are assigned to each engineering stage, and the risk state vectors of all previously completed engineering stages are multiplied by their corresponding weights to obtain a weighted vector.

[0033] The system concatenates all weighted historical vectors with the heterogeneous data vectors of the current stage, sequentially, to form a complete sequence—the risk fusion vector. This vector comprehensively contains information about the risk evolution of the project from its inception to the present. For example, the final risk fusion vector might be a sequence of 181 time steps.

[0034] Specifically, the process of constructing a risk assessment model is as follows: A dataset consisting of risk fusion vectors from all engineering stages and corresponding risk evolution trends is constructed. This data is then input into a risk assessment model built on a long short-term memory network. By training the model, the mapping relationship between the input and the evolution trend is learned, and future risk trends are predicted.

[0035] By using all historical risk fusion vectors and their corresponding evolutionary trends as training data, the model learns what data combinations lead to what risk evolution. This deep learning of the input-output mapping ensures the accuracy and reliability of the system's predictions of future risk trends. The system can predict risk trends over a future period based on the current state. This allows managers to move beyond passively waiting for problems to occur and instead anticipate the evolutionary path of potential risks, giving them ample time to intervene.

[0036] By inputting the risk state vectors of all completed engineering phases as a complete time-series sequence into a Long Short-Term Memory (LSTM) network, the system can learn potential trends and patterns. This provides a high-quality, highly condensed input for subsequent risk evolution trend prediction, significantly improving the accuracy and reliability of the predictions. This deep utilization of historical information overcomes the limitations of existing methods that only focus on "data snapshots," providing a solid foundation for subsequent predictions.

[0037] Furthermore, based on the risk fusion vector, the risk evolution trend of the engineering structure is predicted. When the risk evolution trend deviates from the safety trajectory, an early warning is triggered, corresponding to step S4 above. The specific process is as follows: The system inputs the constructed risk fusion vector into the risk assessment model for iterative prediction. The model first receives the entire risk fusion vector sequence and predicts the risk value for the first future time point, such as day 182. This prediction result is combined with the heterogeneous data vector acquired in real-time on the next day to form a new sequence, which is then input into the model to predict the risk value for the second future time point, such as day 183. Through continuous iteration, the system obtains a risk value sequence composed of multiple discrete time points. For example, the risk value sequence for the next 7 days might be: [0.62, 0.65, 0.68, 0.75, 0.81, 0.82, 0.83].

[0038] The system plots the predicted risk value sequence as a smooth risk evolution trend curve. This curve visually illustrates the risk trajectory of the dam project over a future period, such as the next week or month. The system compares the predicted risk value at each point in time on the curve, and immediately triggers an early warning if the predicted risk value at any point in time exceeds a pre-set safety threshold.

[0039] By using iterative prediction, a fusion vector of past and present risks is taken as input to generate risk values ​​for a series of future time points. This allows the system to plot a complete risk evolution trend curve, thereby transforming risk management from a passive response mode to a proactive prevention mode, enabling the anticipation of the evolution path of potential risks.

[0040] Furthermore, the SHAP retrospective risk assessment model identifies the main risk factors leading to the predicted risk and outputs the supervised robustness, corresponding to step S5 above. The specific process is as follows: The system automatically initiates the SHAP algorithm to calculate a Shapley value for each risk factor and context factor in the risk fusion vector, quantifying the factor's contribution to the final predicted risk value. For example, a positive Shapley value indicates that the factor increases risk, while a negative value indicates that it reduces risk. The system sorts all Shapley values ​​by their absolute values ​​to identify the primary risk factors.

[0041] The system applies small, random perturbations to the current input risk fusion vector. For example, it randomly increases or decreases certain dimensions of the vector, such as "number of construction workers" or "concrete temperature," by a small value, say ±5%, creating 100 new perturbation vectors. These 100 new vectors are then input into the risk assessment model for prediction, yielding 100 different predictions. Finally, the system calculates the standard deviation of these 100 predictions. A smaller standard deviation indicates that the model is less sensitive to small changes in the input data and has higher stability; for example, a standard deviation of 0.01. The Gini coefficient is used to calculate the concentration of major risk factors. For example, if the SHAP algorithm's analysis shows that "concrete temperature" and "number of construction workers" contribute 85% of the total contribution, its Gini coefficient will be high, say 0.85.

[0042] The system performs a linear weighted fusion of model stability and attribution clarity to obtain a supervised robustness score between 0 and 1.

[0043] By constructing supervised robustness, the stability and attribution clarity of the model are quantified, enabling the system to self-verify. This solves the problem of false positives or false negatives caused by poor data quality or poor model uncertainty, ensuring the reliability of the final early warning.

[0044] Furthermore, when the supervisory robustness falls below a preset threshold, the risk is reassessed based on the main risk factors and the supervisory robustness, corresponding to step S6 above. The specific process is as follows: When the generated supervisory robustness falls below a preset threshold, the system automatically identifies the reasons for the low robustness and analyzes the levels of model stability and attribution clarity. If the low robustness is due to low model stability (large standard deviation), it indicates that the model is uncertain about the current input data. The system will send requests to relevant sensors or management personnel, requesting supplementary data. For example, the system might request more concentrated concrete temperature data from the last 24 hours or ask construction workers to upload more on-site videos. If the low attribution clarity is due to a small Gini coefficient, it indicates that the risk is caused by numerous risk factors and the primary cause cannot be identified. The system will send requests to relevant personnel, requesting on-site verification or more background information to help the model better understand the current situation.

[0045] After acquiring supplementary data or background information, the system will automatically update the risk fusion vector with this new data and automatically initiate a reassessment of the risk.

[0046] Specifically, the process of reassessing the risks is as follows: Record the contribution of different features to the volatility of the prediction results in each disturbance test. When the model stability is low, send a data request to the sensors and management personnel to request the supplementation of feature information that contributes greatly to the volatility of the prediction results. When the attribution clarity is low, use the SHAP analysis results to identify all sets of dispersed risk factors with low but non-zero contributions, infer the missing implicit context information based on the set of factors, and send a request to the management personnel to supplement the implicit context information.

[0047] Based on accurate diagnostic results, the system can issue targeted, goal-oriented data requests, resolving the uncertainty in early warnings caused by missing or insufficient quality of key data. This ensures that final warnings are only issued under highly reliable conditions, thereby greatly enhancing managerial trust and guaranteeing the effectiveness and practicality of the entire water conservancy project quality and safety supervision system.

[0048] When robustness falls below a preset threshold, the system can identify the root cause of the problem based on whether it is "low stability" or "low attribution clarity," and automatically initiate a data supplementation request or a secondary assessment. This transforms risk management from passively receiving warnings to proactively and purposefully optimizing data and assessment processes, forming a self-evolving intelligent closed loop.

[0049] Example 2 In this application embodiment, a water conservancy project quality and safety supervision risk assessment system is applied to the quality and safety supervision of river regulation and dredging projects; see reference Figure 2 This is a schematic diagram of a risk assessment system for quality and safety supervision of water conservancy projects. The system includes a risk quantification module, a dynamic assessment module, and a risk adaptive module.

[0050] Furthermore, the risk quantification module includes a text feature extraction sub-model based on NLP, a sensor data feature extraction sub-model based on time-series data analysis, a visual feature extraction sub-model based on deep learning, and a multilayer perceptron deep learning sub-model. The risk quantification model analyzes heterogeneous data such as text data, time-series data, and visual data to identify and extract key risk factors. These risk factors are then input into the MLP sub-model. By analyzing the nonlinear interactions between risk factors, the synergistic effect of the risk factors is transformed into a quantified risk value between 0 and 1. Simultaneously, the quantified risk value, along with contextual factors such as project completion time and investment costs, is encapsulated into a multi-dimensional vector to obtain a risk state vector.

[0051] Furthermore, the dynamic assessment module includes a risk assessment model built on a long short-term memory network. This risk assessment model inputs the heterogeneous data vector of the current engineering stage and the risk state vectors of all previously completed engineering stages. Utilizing the gating mechanism of the long short-term memory network, it intelligently assigns weights to each historical risk state vector and concatenates them with the current heterogeneous data vector in chronological order to obtain a risk fusion vector for the current engineering stage. Simultaneously, it predicts future risk evolution trends based on this risk fusion vector. If the prediction result exceeds a preset safety threshold, an early warning is immediately triggered.

[0052] Furthermore, the risk adaptive module includes the SHAP algorithm. The SHAP algorithm performs backtracking analysis on the risk fusion vector, calculates the Shapley value for each risk factor and contextual factor to identify the main risk factors leading to the predicted risk, randomly perturbs the risk fusion vector, calculates the standard deviation of the prediction results to quantify model stability, and uses the Gini coefficient to calculate the concentration of the main risk factors to quantify attribution clarity. These quantified indicators are then linearly weighted and fused to generate a supervised robustness score between 0 and 1. When the supervised robustness falls below a preset threshold, the system sends requests to sensors and relevant personnel, based on the specific reasons for the low model stability or attribution clarity, requesting supplementary key feature data or implicit contextual information, and initiates a reassessment of the risk.

[0053] This invention not only captures the nonlinear interactions between current risk factors but also learns and remembers the historical evolution trends of projects, thereby accurately predicting future risk trends. This transforms risk management from a passive response to a proactive prevention approach. The system can automatically diagnose the root causes of low monitoring robustness and issue targeted data supplementation requests, enabling the system to self-optimize and ensuring the effectiveness and practicality of the entire risk assessment system.

[0054] Example 3 In large-scale river regulation and flood control projects, the application of this invention can accurately assess the quality and safety risks of the project. Its specific implementation can be as follows: The system acquires multi-source heterogeneous data in real time during the construction and operation of the dam. This includes time-series data on pore water pressure, seepage flow, dam settlement and displacement, soil moisture content, and external water level obtained from sensors deployed inside the dam and in the foundation. These data are processed by a time-series data analysis sub-model to extract features such as rate of change, peak values, and fluctuation trends. Text data obtained from daily inspection logs, geological exploration reports, and hydrological reports is also synchronized and converted into numerical vectors. High-definition images and video visual data acquired from regular UAV inspections are also included. A visual feature extraction sub-model can identify and quantify visual features such as the length and width of cracks on the dam surface, the number of seepage points, or uneven vegetation cover, and encode them into vectors. The extracted risk factor vectors are concatenated with contextual factors to form a high-dimensional input vector. This vector is then input into a deep learning sub-model based on a multilayer perceptron. The hidden layers and nonlinear activation functions of the multilayer perceptron learn and identify complex combined effects between factors.

[0055] The system utilizes the unique gating mechanism of Long Short-Term Memory (LSTM) networks to fuse risk state vectors from historical dam maintenance and operation phases with heterogeneous data vectors from the current phase, thereby predicting future flood control safety trends. When the predicted risk value exceeds the safety threshold, the system triggers an early warning and automatically initiates the SHAP algorithm. SHAP calculates a Shapley value for each dimension in the risk fusion vector used for prediction, such as pore water pressure and piping reports. System backtracking results show that 85% of the risk is caused by two factors: "sudden increase in pore water pressure" (Shapley value +0.42) and "piping reports" (Shapley value +0.35). If robustness is low, and the cause is low model stability (large standard deviation), the system will further backtrack to the perturbation test to identify which feature(s), such as "pore water pressure sensor data," caused the largest prediction fluctuation during the perturbation. Based on this diagnosis, the system will automatically send targeted data requests to relevant personnel, requesting "calibration and retesting of the pore water pressure sensor," or increasing the sampling frequency to obtain more accurate data, thereby initiating a secondary assessment.

[0056] If robustness is low and the cause has low attribution clarity (small Gini coefficient), the system will use SHAP analysis results to identify all dispersed risk factors with low but non-zero contributions. For example, the system might analyze that "external water level" and "wind speed" both contribute weakly, but are insufficient to explain the high risk. Based on this information, the system will intelligently infer potentially missing implicit contextual information, such as "historical water level fluctuation patterns" or "frequency of wind and wave impact on the dam," and request relevant supplementary information from project managers to initiate a more comprehensive secondary assessment.

[0057] By utilizing MLP and Long Short-Term Memory networks to capture the nonlinear interactions between risk factors and predict risk evolution trends, proactive risk prevention is achieved. By using supervised robustness to assess the accuracy of early warnings, self-diagnosis and self-repair are realized, thus comprehensively ensuring the effectiveness and practicality of the entire system.

[0058] 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 risk assessment system for quality and safety supervision of water conservancy projects, characterized in that, include: Risk Quantification Module: Acquire heterogeneous data and contextual factors from each stage of the project, construct a risk quantification model, use heterogeneous data as input, extract risk factors, and transform the synergistic effect of risk factors into risk values ​​by analyzing the nonlinear interaction between risk factors. A risk state vector is obtained by encapsulating risk values ​​with contextual factors. Dynamic assessment module: Based on long short-term memory network, a risk assessment model is constructed. Heterogeneous data of the current engineering stage and risk state vectors of previous engineering stages are used as input. Weights are assigned to the risk state vectors and fused with heterogeneous data to obtain risk fusion vectors. The risk evolution trend of the engineering structure is predicted based on the risk fusion vectors. When the risk evolution trend deviates from the safe trajectory, an early warning is triggered. Risk Adaptive Module: SHAP retrospective risk assessment model identifies the main risk factors that lead to the predicted risk and outputs the supervisory robustness. When the supervisory robustness is lower than the preset threshold, the risk is reassessed based on the main risk factors and the supervisory robustness.

2. The water conservancy project quality and safety supervision risk assessment system according to claim 1, characterized in that, The specific implementation process of constructing a risk quantification model and extracting risk factors by taking heterogeneous data as input includes: inputting heterogeneous data from each stage of the project into the risk quantification model, wherein the heterogeneous data includes text data, time series data, and visual data; and using an NLP-based text feature extraction sub-model, a time series data analysis-based sensor data feature extraction sub-model, and a deep learning-based visual feature extraction sub-model to process the text data, time series data, and visual data respectively, and identify and extract key risk factors.

3. The water conservancy project quality and safety supervision risk assessment system according to claim 1, characterized in that, By analyzing the nonlinear interactions between risk factors, the synergistic effect of risk factors is transformed into a risk value. The specific implementation process includes: mapping the identified risk factors to a coordinate point in a multidimensional space and inputting it into a deep learning sub-model based on a multilayer perceptron; using feature engineering to encode the risk factors into a unified numerical vector; capturing the relationship between risk factors by creating interactive features; and performing nonlinear processing on the risk factors through the nonlinear function of the hidden layer to output a quantitative risk value between 0 and 1.

4. The water conservancy project quality and safety supervision risk assessment system according to claim 1, characterized in that, The specific implementation process of encapsulating risk values ​​and contextual factors to obtain a risk state vector includes: the contextual factors include: project completion time, input cost, environment, number of construction personnel and equipment failure rate; all contextual factors are standardized and mapped to the range of 0 to 1; the quantified risk value is used as the first dimension of the vector, and all standardized contextual factors are concatenated as subsequent dimensions to obtain the risk state vector of the project stage.

5. The water conservancy project quality and safety supervision risk assessment system according to claim 1, characterized in that, A risk assessment model based on Long Short-Term Memory (LSTM) networks is constructed. Heterogeneous data from the current engineering phase and risk state vectors from previous engineering phases are used as inputs. Weights are assigned to the risk state vectors, and the model is fused with the heterogeneous data to obtain a risk fusion vector. The specific implementation process includes: constructing a risk assessment model based on LTM networks; inputting risk state vectors from all completed engineering phases prior to the current phase into the risk assessment model; using the gating mechanism of LTM networks to learn and assign weights to each risk state vector; multiplying all risk state vectors from previous completed engineering phases by their corresponding weights to obtain a weighted vector; and then constructing a complete sequence with the heterogeneous data from the current engineering phase in chronological order, concatenating the two sequences to obtain the risk fusion vector for the current engineering phase.

6. The water conservancy project quality and safety supervision risk assessment system according to claim 1, characterized in that, The risk evolution trend of the engineering structure is predicted based on the risk fusion vector. When the risk evolution trend deviates from the safety trajectory, an early warning is triggered. The specific implementation process includes: inputting the risk fusion vector into the risk assessment model to predict the risk value at the first future time point; using the prediction result and new real-time acquired heterogeneous data as new inputs to predict the risk value at the next future time point; obtaining a risk value sequence composed of discrete time points through iterative prediction; plotting the predicted risk value sequence into a curve to obtain the predicted risk evolution trend; comparing each time point on the prediction curve; and immediately triggering an early warning when there is a predicted risk value higher than a pre-set safety threshold at any of the time points.

7. The water conservancy project quality and safety supervision risk assessment system according to claim 1, characterized in that, The SHAP backtesting risk assessment model identifies the main risk factors leading to predicted risk and outputs a supervised robustness score. The specific implementation process includes: backtesting the risk fusion vector used for prediction using the SHAP algorithm; calculating the Shapley value for each dimension of the risk fusion vector; sorting all Shapley values ​​by magnitude; analyzing the contribution of each risk factor in the risk fusion vector to the prediction result; identifying the main risk factors; subjecting the current input risk fusion vector to multiple random perturbations; inputting the perturbed new vector into the risk assessment model for prediction; calculating the standard deviation of the prediction result to quantify model stability; calculating the concentration of attribution results based on the contribution of the main risk factors to the prediction result using the Gini coefficient to quantify attribution clarity; and linearly weighting and fusing the quantified indicators to generate a supervised robustness score between 0 and 1.

8. The water conservancy project quality and safety supervision risk assessment system according to claim 1, characterized in that, When the supervisory robustness is lower than the preset threshold, the risk is reassessed based on the main risk factors and the supervisory robustness. The specific implementation process includes: setting a threshold according to the project type, comparing the generated supervisory robustness with the preset threshold, and when the supervisory robustness is lower than the preset threshold, identifying the cause of the low supervisory robustness based on the quantitative results of model stability and attribution clarity, sending data requests to the sensors and relevant personnel based on the cause, requesting supplementary data, and reassessing the risk.

Images