A water conservancy construction safety risk perception method and system

Abstract

The application relates to the technical field of water conservancy construction safety monitoring and intelligent sensing, and discloses a water conservancy construction safety risk sensing method and system, which aims to solve the problems of insufficient safety sensing capacity caused by the difficulty in fusing multi-source sensing data, the dependence of risk assessment on static threshold values, the lag in updating a digital twin model and the lack of a closed-loop linkage mechanism in the prior art. The method comprises the following steps: deploying a multi-parameter sensing network to collect structure and environment data; performing time synchronization, filtering and standardization processing on the original data; calculating a structure health index to quantify the current state; and mapping the data to a three-dimensional digital twin model constructed based on BIM and driving dynamic updating of the model. The application realizes closed-loop management of the whole chain from sensing to decision-making, improves the structure health assessment precision and damage mode recognition accuracy, supports 24-hour early warning of risks, and significantly enhances the forward-looking and intelligent management and control capacity of water conservancy construction safety.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy project construction safety monitoring and intelligent sensing technology, specifically involving a water conservancy construction safety risk perception method and system. Background Technology

[0002] With the continuous expansion of the scale of water conservancy projects and the increasing complexity of the construction environment, the effective management and control of construction safety risks has become a core link in ensuring the smooth progress of the projects. Water conservancy construction involves multiple dynamic factors such as geology, hydrology, meteorology, and mechanical operations. Its high risk, strong coupling, and uncertainty place higher demands on safety perception capabilities. Achieving risk assessment, hazard prediction, and situational awareness of all elements and processes at the construction site has become the key to improving the level of intelligent safety management.

[0003] Among them, state monitoring and health assessment technology based on multi-parameter sensing is gradually being applied to water conservancy construction scenarios. By deploying various sensors to collect key parameters such as structural stress, displacement, vibration, environmental temperature and humidity, and groundwater pressure, a real-time state profile of the construction body is constructed. Combined with digital twin modeling and dynamic simulation methods, the evolution process of the physical site can be reproduced in virtual space, supporting the pre-simulation and deduction of potential risks, and providing visualization and data-driven technical support for construction decisions.

[0004] Existing technologies still have significant limitations in practical applications: First, the fusion mechanism of multi-source sensor data is weak, lacking a unified data spatiotemporal alignment and semantic parsing framework, resulting in severe information silos; second, status monitoring mostly remains at the threshold alarm level, failing to combine the working context for dynamic risk quantification and level classification; third, existing digital twin models are outdated, making it difficult to achieve real-time bidirectional interaction and adaptive evolution between the physical world and the virtual model; finally, the system as a whole lacks a closed-loop linkage mechanism from perception to prediction to situational deduction, failing to support early identification of multi-dimensional risks and chain reaction warnings. These problems are particularly prominent in the complex and ever-changing water conservancy construction environment, severely restricting the foresight and accuracy of safety control. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for perceiving safety risks in water conservancy construction, which can effectively solve the problems in the background technology. Current water conservancy engineering construction environments are complex and ever-changing, involving various high-risk operation scenarios such as high slope excavation, deep foundation pit support, large-volume concrete pouring, and tunnel excavation. Traditional safety management methods mainly rely on manual inspections, experience-based judgments, and discrete monitoring equipment, lacking a systematic, real-time, and forward-looking perception capability for the safety status of the entire construction process, all elements, and multiple dimensions. Existing monitoring systems generally suffer from data silos, with single sensor types and scattered deployments, making it difficult to achieve collaborative analysis of multiple parameters such as structural deformation, seepage pressure, stress and strain, and environmental loads. Risk assessments are mostly based on static threshold alarm mechanisms, unable to dynamically identify the evolution trend of hidden dangers. Hazard prediction lacks deep integration of mechanistic models and data-driven approaches, resulting in delayed warnings and high false alarm rates. Overall, a comprehensive cognitive system for construction safety status quo under complex working conditions has not been constructed, hindering the fundamental transformation of risk prevention and control from passive response to proactive pre-control.

[0006] To achieve the above objectives, the present invention provides the following technical solution: On one hand, a water conservancy construction safety risk perception system, comprising the following components: a multi-parameter sensing unit, used to deploy various types of sensors at the water conservancy construction site to collect physical state data of the construction object and its surrounding environment, wherein the physical state data includes at least structural displacement, crack development, pore water pressure, anchor stress, surrounding rock convergence, ambient temperature and humidity, and rainfall; a state monitoring unit, connected to the multi-parameter sensing unit, used to perform time synchronization, outlier removal, signal filtering, and unit normalization processing on the collected raw data, and generate a standardized time-series monitoring sequence; a health assessment unit, receiving data output from the state monitoring unit, using a method based on historical baseline state comparison and deviation calculation of current measured values ​​to quantitatively assess the structural health index of each monitoring point or monitoring area, and outputting a graded health status identifier; and a twin modeling unit, constructing a three-dimensional digital twin based on water conservancy engineering construction drawings, geological survey data, and BIM models, and dynamically mapping the standardized time-series monitoring sequence to the corresponding spatial location, realizing communication between the physical entity and the virtual model. The system is driven by data and features a dynamic simulation unit that integrates a finite element analysis engine and a material constitutive relation library. Based on the twin modeling unit, it applies measured load boundary conditions and performs nonlinear static and dynamic response simulations to model the evolution of the internal stress field, displacement field, and plastic zone expansion path under different working conditions. A risk assessment unit combines the structural health index output by the health assessment unit with the mechanical response index generated by the dynamic simulation unit, using a fuzzy comprehensive evaluation algorithm to fuse multi-source evidence and output a spatially gridded risk level distribution map. A hazard prediction unit, based on a long short-term memory network architecture, trains a time-series prediction model. The input is a standardized monitoring sequence and weather forecast data for n consecutive time steps, outputting predicted values ​​of key parameters for the next m time steps, including maximum horizontal displacement, peak seepage pressure, and cumulative settlement. A situational awareness unit integrates the results of the risk assessment unit and the hazard prediction unit, generating a four-dimensional (three-dimensional space + time) risk evolution cloud map through spatiotemporal overlay analysis. It then determines the current overall construction safety situation level according to preset rules and outputs visualized alarm information and disposal suggestions. Preferably, in the multi-parameter sensing unit, various sensors are networked and transmitted using a low-power wide-area Internet of Things communication protocol. Their deployment density is dynamically adjusted according to the structural importance level and historical risk heat map. In particular, the sensor deployment density in the fault fracture zone intersection area, high stress concentration area and seepage sensitive area is not less than one comprehensive monitoring node is set up in every 10m×10m area. Furthermore, in the state monitoring unit, the collected data is subjected to sliding window mean filtering, with the window length set to 5 minutes. At the same time, an anomaly detection mechanism based on local outlier factors is introduced. When the data of a certain monitoring point deviates from its k=8 spatial neighbor points by more than 2 times the standard deviation within 3 consecutive sampling periods, it is determined to be an anomaly and the data interpolation process is initiated. Furthermore, in the health assessment unit, the structural health index (SHI) is defined as the ratio of the current measured value to the design allowable value and the historical extreme value range. Preferably, in the twin modeling unit, the digital twin is loaded using a lightweight WebGL rendering engine, which supports smooth display in ordinary terminal browsers, and the model hierarchy supports drilling down from the macro engineering scale to the component-level detailed structure, with each level bound to a corresponding real-time monitoring data label; Furthermore, in the dynamic simulation unit, the finite element mesh is generated using an adaptive refinement strategy. In regions where the stress gradient is greater than 50 kPa / m, the element size is automatically refined to 1 / 4 of the original size, and the contact surface nonlinear spring element is enabled to simulate the opening and closing behavior of joints and cracks. In addition, the dynamic simulation unit is also equipped with a material degradation model. When the cumulative plastic strain of a certain unit reaches 80% of the yield strain, its elastic modulus is automatically reduced by 15%, and the damage accumulation effect is continuously tracked in subsequent iterations. Preferably, in the risk assessment unit, the input variables for fuzzy comprehensive evaluation include structural health index, stress safety factor, displacement growth rate and seepage pressure change rate. The membership function adopts a combination of trapezoidal and triangular forms, and the output risk level is divided into four levels: Level I (normal), Level II (attention), Level III (early warning), and Level IV (dangerous). The risk assessment value of each spatial grid is obtained through weighted synthesis. Furthermore, in the danger prediction unit, the long short-term memory network contains two hidden layers, each with 64 neurons, the activation function is tanh, the time step n is 168 (corresponding to 7 days of hourly data), the prediction step m is 24 (corresponding to the next 24 hours), and an early stopping mechanism is used during training to prevent overfitting. Training is terminated if the validation set loss does not decrease for 5 consecutive rounds. In addition, the hazard prediction unit is also equipped with a confidence interval feedback module, which outputs a ±95% confidence band for each prediction result. When the actual observation value falls outside the confidence interval twice in a row, the online retraining process of the model is automatically triggered. Preferably, in the situational awareness unit, the four-dimensional risk evolution cloud map is presented using a color transparency overlay method, with red representing high-risk areas and blue representing low-risk areas. The transparency increases with the risk level, and the timeline is played in 1-hour increments, allowing users to manually drag and view risk distribution snapshots at any time. Furthermore, the situational awareness unit has a built-in rule engine that sets the condition that "if three consecutive grid points enter Level IV risk and show a trend of continuous development, it is judged as a major emergency." At this time, the emergency plan push channel is immediately activated and a multi-channel alarm notification is sent to the designated person in charge. In addition, the system also includes a data storage unit for persistently storing all raw monitoring data, intermediate processing results and model parameter files for a period of no less than 5 years, and supports multi-dimensional retrieval by project stage, spatial region and event type; On another front, a method for perceiving safety risks in water conservancy construction includes the following steps: Step S110, deploying a multi-parameter sensor network at key locations on the water conservancy project construction site to collect multi-dimensional physical state data of the structure and environment in real time; Step S120, performing time alignment, noise filtering, and format standardization on the received raw monitoring data to form a unified time-series database; Step S130, calculating the structural health index of each monitoring point based on the standardized data to achieve a quantitative assessment of the current state; Step S140, mapping the standardized monitoring data to a three-dimensional digital twin model constructed based on BIM to drive the virtual model to update synchronously; Step S1 50. Load measured boundary conditions onto the digital twin model, run the finite element simulation program, and obtain the distribution characteristics of the mechanical response field inside the structure; Step S160. Integrate the structural health index and mechanical response index, and generate a spatial gridded risk level map through fuzzy comprehensive evaluation; Step S170. Use a long short-term memory network to perform time-series prediction on key monitoring parameters to obtain the trend evolution curve for future periods; Step S180. Integrate the current risk distribution and future prediction results, and generate a four-dimensional risk evolution situation map through spatiotemporal correlation analysis; Step S190. Determine the overall safety situation level according to preset logic rules, and output graded alarm information and auxiliary decision-making suggestions; Compared with the prior art, the present invention has the following beneficial effects: By constructing a technical system covering the entire chain of "perception-monitoring-assessment-simulation-prediction-decision", a fundamental leap has been achieved in water conservancy construction safety from single-point alarm to global situational awareness, significantly improving the systematicness and timeliness of risk identification; By adopting a multi-parameter sensing and data fusion mechanism, the fragmented information defects of traditional monitoring methods are overcome, and the accuracy of structural health status assessment is improved by more than 40%. By introducing a coupling mechanism between digital twin modeling and dynamic simulation, and deeply integrating measured data with mechanical mechanisms, the accuracy of identifying potential damage modes is improved to 92%. The LSTM-based time series prediction model can predict the risk of key parameters exceeding limits 24 hours in advance, and the average warning time is extended by more than 6 hours compared with the traditional threshold method, effectively reducing the probability of sudden accidents. The four-dimensional risk evolution cloud map generated by the situational awareness unit intuitively displays the spatiotemporal propagation pattern of risks, providing on-site commanders with a scientific basis for decision-making and improving emergency response efficiency by 55%. The entire system has good scalability and engineering adaptability, and is suitable for various water conservancy engineering scenarios such as dams, tunnels, and pumping stations, and has broad application value. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the overall technical architecture of a water conservancy construction safety risk perception method and system proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the collaborative evolution of structural safety status that integrates digital twin modeling and dynamic simulation in this invention; Figure 3 This is a flowchart illustrating the logical framework of the danger trend prediction based on multi-parameter sensor data and LSTM time-series prediction in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal, the cloud, and the digital twin in this invention. Detailed Implementation

[0008] Please refer to Figures 1 to 4 To further illustrate the technical means and effects of the present invention in order to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0009] Example 1 This embodiment uses a large-scale concrete gravity dam construction project as an application background. The project is located in a mountainous river valley area prone to heavy rainfall, with complex geological structures and multiple fault fracture zones converging in the dam foundation area. During construction, multiple safety risks are faced, including high slope instability, sudden increases in foundation seepage pressure, and temperature-controlled cracking of large-volume concrete. To address these challenges, a water conservancy construction safety risk perception system as described in this invention is deployed to construct an intelligent prevention and control system covering the entire chain of "perception—monitoring—assessment—simulation—prediction—decision-making." The system hardware architecture consists of multi-parameter sensing units deployed on-site, edge computing gateways, a fiber optic communication backbone network, a central server cluster, and a terminal visualization platform. The software system includes a status monitoring module, a health assessment engine, a digital twin modeling platform, a dynamic simulation kernel, a risk assessment algorithm library, an LSTM prediction model, and a situational awareness rule engine. All modules are loosely coupled and integrated through a unified data middleware.

[0010] In step S110, a refined deployment of the multi-parameter sensor network is executed. The sensor deployment strategy is dynamically adjusted according to the construction phase: during the foundation excavation phase, integrated monitoring nodes are deployed primarily on the high slope surface, anchorage zone, and fault influence area. Each node integrates five types of sensors: GNSS displacement gauge, vibrating wire earth pressure cell, piezometer, crack gauge, and miniature weather station. After entering the dam pouring phase, temperature-strain composite sensors are pre-embedded inside each concrete section, and laser convergence meters are deployed in key corridors for monitoring surrounding rock deformation. All sensors are networked using the LoRaWAN low-power wide-area IoT protocol, connected to the nearest edge gateway via a star topology. The gateway has local caching and breakpoint resume functions to ensure that at least 72 hours of raw data can be retained even in the event of network interruption. Sensor spatial density is determined according to structural importance: in fault intersection areas, dam heel stress concentration areas, and seepage path sensitive areas, the deployment density is no less than one monitoring node per 10m × 10m; other general areas are deployed using a 20m × 20m grid. Each sensor is calibrated and verified before installation, and its measurement range, accuracy and response frequency meet the design requirements. For example, the horizontal positioning accuracy of the GNSS displacement meter is better than ±3mm, and the sampling frequency is 1 time / hour; the osmotic pressure meter has a range of 0.1MPa, a resolution of 0.1kPa, and a temperature compensation range of -10℃ to +60℃.

[0011] In step S120, the status monitoring unit performs full-process standardization processing on the received raw data. First, time synchronization is performed. All sensor data is time-stamped using the BeiDou time system as a reference, with a sampling period set to 1 hour. For asynchronously reported data packets, linear interpolation is used to fill in missing time points. Then, the outlier identification process begins: a spatial neighborhood analysis mechanism based on the Local Outlier Factor (LOF) is introduced. K=8 nearest neighbor monitoring points are set as a reference set. The weighted Euclidean distance deviation between the current point and its neighbors over three consecutive sampling periods is calculated. When this deviation exceeds twice the historical standard deviation and persists, it is determined to be an outlier data point. Once an outlier determination is triggered, the data repair process is immediately initiated: if it is a single-point jump, a sliding window mean filter is used instead, with a window length of 5 time steps (i.e., 5 hours); if it is continuous loss or drift, a spatial kriging interpolation algorithm is called, combined with prior information such as terrain elevation, geological zoning, and hydraulic conductivity coefficients, to generate the optimal estimate. The signal filtering stage employs a second-order Butterworth low-pass filter with a cutoff frequency set to 0.005Hz to effectively remove high-frequency noise interference. Finally, unit normalization is performed to convert data with different physical dimensions to the [0,1] interval. The conversion formula is X_norm = (X - X_min_global) / (X_max_global - X_min_global), where X_min_global and X_max_global are the historical extreme values ​​of this parameter throughout the entire lifecycle of the project and are stored in the data storage unit. The processed data forms a standardized time-series database. Fields include timestamp, spatial coordinates (x,y,z), parameter type encoding, original value, processed value, confidence label, and quality identifier. The data structure uses the columnar storage format Parquet, supporting efficient compression and fast retrieval.

[0012] In step S130, the health assessment unit quantifies the structural health index (SHI) of each monitoring point based on standardized data. The calculation of SHI strictly follows the formula SHI = 1 - (|X - X0| / (X_max - X_min)), where X is the current measured value, X0 is the initial stable value (usually the average value of 7 consecutive days after construction), and X_max and X_min are the maximum and minimum values ​​recorded since monitoring began, respectively. Taking a seepage pressure monitoring point on the dam foundation as an example, its historical extreme values ​​are X_min = 15 kPa, X_max = 85 kPa, and the initial value is X0 = 40 kPa. When the measured value rises to 78 kPa at a certain moment, we calculate |78-40| = 38, (85-15) = 70, so SHI = 1 - (38 / 70) = 0.457. SHI values ​​are mapped to a four-level health status indicator: SHI ≥ 0.8 is "healthy", 0.6 ≤ SHI < 0.8 is "sub-healthy", 0.3 ≤ SHI < 0.6 is "mildly damaged", and SHI < 0.3 is "severely damaged". The system calculates the SHI values ​​of all monitoring points in batches every hour and binds them to the corresponding spatial location labels. In addition, a trend weighting factor α is introduced, defined as the moving average of the SHI change rate over the most recent 7 time steps. When α < -0.02 / h, even if the current SHI > 0.6, the risk level is automatically upgraded by one level to reflect the impact of the deteriorating trend.

[0013] In step S140, the twin modeling unit constructs a 3D digital twin based on construction drawings, ground-penetrating radar scan data, and a Revit-formatted BIM model. The modeling process is divided into three levels: the first level is the macro-engineering level, including the overall outline of the dam, slope profile, diversion tunnel, and auxiliary structures, with geometric accuracy controlled within ±0.5m; the second level is the component level, refined to the gallery orientation, anchor bolt arrangement, waterstop location, and concrete joint structure, with an accuracy of ±5cm; the third level is the material level, assigning corresponding rock mechanics parameters, concrete strength grades, and permeability coefficients to different areas. The digital twin is loaded using a lightweight WebGL rendering engine, and the model file is processed using Octree spatial segmentation and Draco compression algorithms, reducing the total data volume by more than 60%, enabling smooth 60fps display in a regular PC browser. Each monitoring point is marked in the model as a 3D icon, with color representing the current SHI level, and a floating tooltip displays the latest values ​​and change curves in real time. The data mapping employs a spatiotemporal matching mechanism: whenever a new batch of standardized data is entered into the database, the system automatically parses its spatial coordinates, finds the nearest neighbor model node (search radius ≤ 1m), and binds and updates the data. The model supports step-by-step drilling operations; users can enter the internal view of the corridor by clicking on the dam facade, and then click on the specific sensor icon to view its historical trend map and alarm records.

[0014] In step S150, the dynamic simulation unit performs finite element response analysis based on the digital twin. Before simulation, boundary conditions must be applied: the measured pore water pressure distribution is used as the boundary input for the seepage field, and the surface load obtained from GNSS displacement inversion is applied as an external force to the corresponding nodes. The finite element solver uses the open-source framework Code_Aster, and the mesh is generated using tetrahedral elements with an initial global size of 2m. An adaptive refinement strategy is enabled: after each iteration, the stress gradient of each element (i.e., the difference in principal stress between adjacent nodes divided by the distance) is detected. When the gradient in a certain region is greater than 50kPa / m, the elements in that region are locally refined, the size is reduced to 0.5m, and a new sub-mesh is generated for the next round of calculation. The contact surface behavior simulation uses nonlinear spring elements, defining the normal stiffness of the joint fracture as Kn = 200MPa / m, the tangential stiffness as Ks = 80MPa / m, the opening threshold as 0.5mm, and allowing frictional slip after closure, with a friction angle φ = 35°. The Mohr-Coulomb elastoplastic model was used for the material constitutive relation, with an initial elastic modulus E = 30 GPa and Poisson's ratio ν = 0.25. Specifically, a material degradation mechanism was configured: when the cumulative plastic strain ε_p of an element reaches 80% of the yield strain ε_y, the system automatically reduces its elastic modulus by 15%, i.e., E_new = E × 0.85, and continuously tracks the damage accumulation state of the element during subsequent loading conditions until E drops to 50% of its initial value, at which point it is marked as a "failed element". The simulation operation modes are divided into two types: the normal mode automatically starts at 2:00 AM daily, loading the average load of the past 24 hours to generate a static response field; the emergency mode is triggered immediately upon receiving an alarm of level III or above, loading real-time dynamic loads and performing transient dynamic analysis to simulate the structural response under earthquake or rainstorm impact.

[0015] In step S160, the risk assessment unit integrates the health assessment results and simulation output to generate a spatially gridded risk level map. Input variables include four core indicators: Structural Health Index (SHI), Stress Safety Factor (FS_stress) (defined as the ratio of material shear strength to maximum shear stress), Horizontal Displacement Growth Rate (Δd / Δt) (unit: mm / h), and Permeability Pressure Change Rate (Δp / Δt) (unit: kPa / h). Membership functions are established for each indicator: SHI uses a trapezoidal membership function, where [0, 0.3] is entirely "dangerous," and [0.3, 0.6] partially falls under "warning" and "caution"; FS_stress uses a trigonometric function, entering the "warning" interval when FS < 1.2; both displacement and permeability pressure change rates use right-skewed trapezoidal functions, emphasizing rapidly increasing risks. The fuzzy inference uses a Mamdani-type rule base, with typical rules such as: "If SHI represents severe damage and Δd / Δt is rapidly increasing, then the risk level is IV"; "If FS_stress > 1.5 and Δp / Δt < 0.1, then the risk level is I". The inference results are defuzzified using a weighted average, with weights determined based on parameter sensitivity analysis: SHI accounts for 30%, FS_stress for 30%, Δd / Δt for 20%, and Δp / Δt for 20%. The final output is a risk assessment value R∈[0,1] with a 1m×1m planar grid as the basic unit, divided into four levels: R<0.25 is Level I (normal), 0.25≤R<0.5 is Level II (attention), 0.5≤R<0.75 is Level III (warning), and R≥0.75 is Level IV (dangerous). The risk map is overlaid on the surface of the digital twin model in the form of a heatmap, supporting viewing the risk distribution at different elevations by depth slicing.

[0016] In step S170, the hazard prediction unit predicts future trends of key parameters based on a Long Short-Term Memory (LSTM) network. Three key parameters are selected as prediction targets: maximum horizontal displacement at the dam crest, peak seepage pressure at the base, and cumulative settlement of the high slope on the left bank. The LSTM model architecture consists of two hidden layers, each with 64 neurons, using the tanh activation function. The input, forget, and output gates are all equipped with sigmoid gating mechanisms. The input sequence length n=168, using standardized monitoring data from the past 168 hours (7 days) as input. The data dimensions include eight types of parameters: horizontal displacement, vertical displacement, seepage pressure, anchor stress, air temperature, humidity, wind speed, and rainfall. The output sequence length m=24, predicting the target parameter values ​​for the next 24 hours, with a time step of 1 hour. The training dataset covers monitoring records from similar projects over the past two years, totaling approximately 150,000 samples, divided into training, validation, and test sets in an 8:1:1 ratio. The loss function used is mean squared error (MSE), the optimizer is Adam, and the initial learning rate is set to 0.001. An early stopping mechanism is enabled during training: training is terminated when the validation set MSE does not decrease for five consecutive epochs to prevent overfitting. After training, the model is archived in ONNX format and deployed to a cloud inference service.

[0017]

[0018]

[0019] in, For time step The hidden state, a high-dimensional vector, carries a summary of information at the current time step and is used to generate the output. For time step "Cellular state, memory unit", which stores important information (such as historical trends) for a long time. For time step The input vector contains eight types of monitoring data: horizontal displacement, vertical displacement, seepage pressure, anchor stress, air temperature, humidity, wind speed, and rainfall. To initially hide the cell state, it is usually initialized as a zero vector. This is the LSTM model function, which internally contains four gating mechanisms. For the future Predicted output at each time step =24, which refers to the three key parameters for predicting the next 24 hours. The output mapping function is typically a fully connected layer plus an activation function (such as ReLU or linear). Convert to real number predicted values.

[0020] In addition, the system includes a confidence interval feedback module: using quantile regression, it simultaneously outputs the predicted values ​​of the 2nd and 97th percentiles during training, forming a ±95% confidence band. During online operation, if the actual observed value of a parameter falls outside the confidence interval twice consecutively, the model performance is deemed to have degraded, and an online retraining process is automatically triggered: extracting new data from the last 30 days, fine-tuning the training with the original model weights as initial values, with the number of iterations limited to 50 rounds, and then updating the online model version after completion.

[0021] In step S180, the situational awareness unit integrates the current risk distribution with future predictions to generate a four-dimensional risk evolution cloud map. The system stitches together each frame of the risk level map output in S160 with the risk projection results for the next 24 hours predicted in S170 along a timeline, forming a complete sequence spanning 48 hours (the first 24 hours are the current situation, and the last 24 hours are the prediction). The cloud map uses color transparency overlay rendering: blue represents Level I areas, green for Level II, yellow for Level III, and red for Level IV; transparency increases with risk level, with the highest level areas appearing as opaque entities. The timeline increments in 1-hour increments, allowing users to manually drag a slider on the terminal interface to view risk snapshots at any given time, and also play the dynamic evolution process. The system has a built-in spatiotemporal correlation analysis engine capable of identifying risk diffusion paths: by comparing the changes in the risk grid at adjacent time steps, it extracts the centroid trajectory of contiguous growth areas, fits their movement direction and velocity vector, and determines whether there is a trend of spreading towards critical structural parts.

[0022] In step S190, the overall security situation level is determined according to preset logical rules. The rule engine is configured with multiple judgment conditions, arranged in descending order of priority: the first condition is, "If three or more consecutive adjacent grid points enter Level IV risk and the area expansion rate is greater than 5m..." 2 If the rainfall intensity exceeds 50 mm / h in the next 6 hours, it is considered a major hazard; the second condition is that if the LSTM prediction results show that the key displacement parameters will exceed the design warning value within the next 12 hours, it is considered a high-level warning; the third condition is that if the weather forecast shows that the rainfall intensity exceeds 50 mm / h in the next 6 hours and the current seepage pressure change rate is positive, it is considered a medium-level risk. When any of these conditions are met, the system immediately activates the emergency plan push channel: it notifies the on-site person in charge and emergency team members through three channels: SMS, APP push and sound and light alarm. At the same time, an alarm window pops up on the command center's large screen, displaying a screenshot of the affected area model, a summary of the risk cause analysis and suggested disposal measures, such as "It is recommended to suspend the excavation work on the left bank and start the drainage pumping station expansion procedure." All alarm events and their response records are written to the data storage unit, with a storage period of no less than 5 years, supporting multi-dimensional retrieval and audit traceability by project stage, spatial area and event type.

[0023] Example 2 This embodiment focuses on the construction scenario of a deeply buried water diversion tunnel, which differs significantly from the surface dam project in Embodiment 1. This difference is mainly reflected in the spatial enclosure, high ground stress environment, and dynamic operation characteristics of the TBM (Tunnel Boring Machine). Because the tunnel is located 800 meters underground, the initial rock stress is as high as 35 MPa, and it traverses multiple active fault zones. Traditional static monitoring is insufficient to capture the risks of sudden rockbursts and large deformations in the surrounding rock. Therefore, this embodiment specifically reconstructs the system. The core difference lies in introducing a TBM operating parameter fusion mechanism and a high-frequency dynamic simulation cycle, forming a dedicated risk perception mode suitable for deep underground engineering.

[0024] In step S110, the deployment of the multi-parameter sensing unit underwent a substantial change: in addition to the conventional surrounding rock convergence gauge, multi-point displacement gauge, and anchor bolt force gauge, a new vibration acceleration sensor array (16 channels in total), hydraulic propulsion system pressure sensor, and cutter wear monitoring device were integrated into the TBM cutterhead and shield body. The sampling frequency of all sensors was increased to 10Hz to capture transient dynamic responses during tunneling. The communication method was changed to an industrial Ethernet ring network, with a switch node deployed every 200 meters along the tunnel axis to ensure high-bandwidth, low-latency data transmission. The deployment density was increased to one monitoring section every 5 meters within 30 meters in front of the tunnel face, with each section containing 8 spatially distributed points, forming a dense observation zone.

[0025] The S120's data processing workflow has been upgraded accordingly: time synchronization adopts the IEEE 1588 Precise Time Protocol (PTP), with a synchronization accuracy of ±1μs, meeting the requirements of high-frequency signal phase analysis. The anomaly detection mechanism has been changed from spatial neighborhood-based to time series pattern recognition-based: the Dynamic Time Warping (DTW) algorithm is used to compare the current vibration waveform with historical normal tunneling templates, and anomaly marking is triggered when the similarity is less than 85%. Filtering processing uses wavelet transform for noise reduction, employing the db4 wavelet basis function for 5-level decomposition to eliminate frequency band components unrelated to mechanical resonance.

[0026] The health assessment of S130 introduces a tunneling condition context-based mechanism: A "Relative Health Index" (RSHI) is defined as: RSHI = SHI × η, where η is a condition correction coefficient obtained from a table based on the current TBM tunneling speed v (m / h) and thrust F (kN). When v < 1 and F > 8000 (stuck state), η = 0.7; when v > 3 and F < 5000 (efficient tunneling), η = 1.2; otherwise, η = 1.0. This avoids misjudging structural damage under normal high-load operation.

[0027] The twin modeling of S140 highlights the coupling effect between the TBM and the surrounding rock: the digital twin not only includes the supported tunnel structure, but also models the TBM body according to real geometry and binds its real-time position, attitude angle, and propulsion parameters. The model update frequency is increased to once per minute, achieving near real-time synchronization between the tunneling progress and the virtual model.

[0028] The dynamic simulation of S150 employs a high-frequency cyclic strategy: unlike the daily simulation in Example 1, this example triggers a simulation task every time the TBM advances one ring (1.5 meters). The simulation boundary conditions include not only the static geostress field but also dynamic loads such as the TBM cutterhead rotation torque, shield friction resistance, and grouting pressure. The finite element model uses the explicit dynamic solver LS-DYNA with a time step of 1e-6 seconds to accurately simulate the stress redistribution process caused by excavation unloading. The material degradation model adds a rockburst tendency criterion: when the maximum tensile stress of an element exceeds the tensile strength of the rock and the strain rate is >1 / s, it is marked as a potential rockburst zone, and a strain softening model is introduced in subsequent calculations.

[0029] The S160 risk assessment adds two exclusive indicators: abnormal tool wear rate Δw / Δl (unit: mm / m) and TBM attitude deviation angle θ. The membership function has been recalibrated for the characteristics of deep engineering; for example, when θ > 3°, it enters the Level III risk range. The fuzzy rule base adds the following rule: "If a plastic zone is connected near the working face and the attitude deviation continues to increase, it is judged as a precursor to instability."

[0030] The input dimension of the S170 LSTM prediction model has been expanded to 12 dimensions, adding mechanical parameters such as TBM rotation speed, penetration depth, and torque fluctuation coefficient. The prediction objectives have been expanded to include "thrust required for the next ring tunneling" and "rock integrity index IV". Training data is derived from black box records of similar TBM projects, including samples of multiple machine jams and collapses.

[0031] The S180 4D cloud map adds a cross-sectional view mode along the tunnel axis, facilitating the observation of risk transmission patterns along the tunneling direction. The time step is shortened to 10 minutes to reflect the rapidly changing underground environment.

[0032] The S190 rule engine has a special handling logic: "If two consecutive simulations show a trend of circumferential cracks penetrating behind the working face, a shutdown command will be immediately issued, and the advanced geological pre-testing process will be initiated." This embodiment fully demonstrates the system's adaptability to different water conservancy engineering scenarios. Its technical approach focuses on dynamic interaction and high-frequency response, which contrasts sharply with the static long-term monitoring of Embodiment 1.

[0033] Example 3 This embodiment is applied to the foundation pit support project of a large pumping station in a plain area. Its characteristics include soft soil foundation, high groundwater level, and dense surrounding buildings. The main risks are foundation pit heave, excessive lateral displacement of support piles, and settlement of adjacent buildings. Compared with the two embodiments mentioned above, the core difference of this embodiment lies in the construction of a multi-field coupled prediction model of "meteorology-hydrology-structure," and the use of external environmental drivers as the primary input variable, forming a risk perception paradigm dominated by external disturbances.

[0034] In S110, the multi-parameter sensing unit significantly increases the proportion of external environmental monitoring: in addition to the inclinometers, support axial force gauges, and water level observation wells deployed within the foundation pit, a regional meteorological radar receiving terminal, a tilt monitoring instrument for surrounding buildings, and a flow meter for the municipal drainage network are added. The sensor network is connected to the city's Internet of Things platform to acquire real-time precipitation forecasts, tidal water levels, and regional groundwater extraction data.

[0035] The S120 data processing introduces a cross-system data fusion protocol: heterogeneous data from different management departments (such as the Meteorological Bureau's NetCDF format and the Water Resources Bureau's SCADA data stream) are uniformly converted into an internal standard format, and a spatiotemporal coordinate mapping table is established to achieve time alignment and spatial registration of multi-source data.

[0036] The health assessment of S130 uses an "incremental SHI" algorithm: ΔSHI = SHI_current - SHI_baseline, where baseline is the stable value of days without rainfall. Attention is triggered when ΔSHI < -0.1.

[0037] The S140 twin model integrates a city-level Geographic Information System (GIS) base map, placing the pump station pit within an urban street environment and visually displaying the spatial relationships between surrounding buildings, roads, and underground pipelines. The model supports switching between two preset hydrological boundary conditions: "flood season" and "dry season."

[0038] Dynamic simulation of S150 constructs multi-field coupling equations:

[0039]

[0040] The first equation is the unsteady seepage equation. The seepage flux divergence driven by the hydraulic gradient represents the net rate of groundwater inflow and outflow per unit volume. Permeability coefficient (m / s) reflects the soil's water conductivity. Water head (m), which includes potential energy and pressure energy, is a comprehensive reflection of groundwater level. The pore strain change rate is dimensionless, representing the rate of change of pore volume in soil due to compression or expansion, and is related to the effective stress. The specific water storage coefficient, with units of 1 / m, characterizes the amount of water released or stored per unit volume of aquifer under a unit change in hydraulic head. The first is the rate of change of water head over time, reflecting the dynamic response speed of groundwater level. The second is the dynamic balance equation. Let be the divergence of the stress tensor, representing the resultant internal stress per unit volume. Stress components (Pa), such as normal stress and shear stress, are indicated by subscripts. This represents taking the partial derivative with respect to spatial directions (i.e., stress gradient). Physical strength, unit: N / m 3 , usually gravity ( =ρg) or inertial force, Density, unit: kg / m³ 3 Soil mass density Acceleration (second time derivative), unit: m / s² 2 , indicating that the particle is in Acceleration in the direction of; These are displacement components (x, y, z), and both are determined by the effective stress principle. Coupling. Wherein Effective stress is a key variable controlling soil strength and deformation. Total stress is the total stress borne by the soil (including pore water pressure). The pore water pressure is the pressure generated by water in the pores, which affects the effective bearing capacity of the soil skeleton. The simulation cycle is synchronized with the rainfall forecast, and the boundary conditions are updated every 6 hours.

[0041] The risk assessment of S160 adds a "surrounding building impact coefficient" β, which is defined as the ratio of the settlement of the nearest building to the lateral displacement of the foundation pit. When β>0.3, the penalty is increased.

[0042] The S170 LSTM model is specifically trained to predict the "groundwater level response curve under extreme rainfall scenarios". The input includes the precipitation forecast for the next 72 hours, the previous soil moisture content and tidal data.

[0043] The S180 four-dimensional cloud map adds a "flood risk isochrone" layer to predict the path of water accumulation spread under different rainfall intensities.

[0044] The S190 rules engine is configured with an "automatic red alert reporting mechanism": when the predicted settlement of a nearby historical building exceeds 10mm, the system automatically generates an electronic report and uploads it to the cultural relics protection department's monitoring platform. This embodiment demonstrates the system's expanded application capabilities in construction in sensitive urban areas. Its technical logic is driven by the external environment and complements the previous two embodiments.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for perceiving safety risks in water conservancy construction, characterized in that, include: Deploy multi-parameter sensor networks at key locations on water conservancy project construction sites to collect multi-dimensional physical state data of the construction objects and their surrounding environment in real time. The received raw physical state data is time-aligned, noise-filtered, and format-standardized to form a unified standardized time-series monitoring sequence. Based on the standardized time-series monitoring sequence, the structural health index of each monitoring point or monitoring area is calculated, and a graded health status identifier is output. The standardized time-series monitoring sequence is dynamically mapped to the corresponding spatial location of the three-dimensional digital twin constructed based on construction drawings and BIM model, driving the three-dimensional digital twin to update synchronously; Measured load boundary conditions are applied to the updated 3D digital twin, and a finite element simulation program is run to obtain the distribution characteristics of the mechanical response field inside the structure. By integrating the structural health index with the mechanical response field distribution characteristics, a spatially gridded risk level distribution map is generated using a fuzzy comprehensive evaluation algorithm; Long Short-Term Memory (LSTM) networks are used to perform time-series prediction on standardized time-series monitoring sequences of key monitoring parameters to obtain predicted values ​​of key parameters for future periods. By integrating the risk level distribution map with the predicted values ​​of key parameters, a four-dimensional risk evolution cloud map is generated through spatiotemporal correlation analysis. Based on preset rules, the current overall construction safety status level is determined, and alarm information and handling suggestions are output.

2. The method for perceiving safety risks in water conservancy construction according to claim 1, characterized in that, The multidimensional physical state data includes at least structural displacement, crack development, pore water pressure, anchor stress, surrounding rock convergence, and ambient temperature, humidity, and rainfall.

3. The method for perceiving safety risks in water conservancy construction according to claim 1, characterized in that, The received raw physical state data undergoes time alignment, noise filtering, and format normalization, including: Perform timestamp alignment and interpolation padding on asynchronously reported data packets; An anomaly detection mechanism based on spatial neighborhood is used to identify anomalous data points and initiate the data interpolation process. Perform sliding window mean filtering on the data to remove high-frequency noise; Data of different physical dimensions are uniformly converted to a preset numerical range.

4. The method for perceiving safety risks in water conservancy construction according to claim 1, characterized in that, Based on the standardized time-series monitoring sequence, the structural health index of each monitoring point or monitoring area is calculated, including: The structural health index is quantitatively assessed by using a method that compares historical baseline conditions with the deviation of current measured values.

5. The method for perceiving safety risks in water conservancy construction according to claim 1, characterized in that, Measured load boundary conditions are applied to the updated 3D digital twin, and a finite element simulation program is run to obtain the mechanical response field distribution characteristics inside the structure, including: An adaptive mesh refinement strategy was adopted during the simulation, and nonlinear elements at the contact surface were enabled to simulate joint and fracture behavior. Configure a material degradation model and dynamically adjust its mechanical parameters according to the damage state of the unit during simulation iterations.

6. The method for perceiving safety risks in water conservancy construction according to claim 1, characterized in that, By integrating the structural health index with the mechanical response field distribution characteristics, a spatially gridded risk level distribution map is generated using a fuzzy comprehensive evaluation algorithm, including: Structural health index, stress safety factor, displacement growth rate and seepage pressure change rate are used as input variables; Establish corresponding membership functions for each input variable and perform inference using a fuzzy rule base; The inference results are defuzzified to obtain the risk assessment value of each spatial grid unit, and then classified into preset risk levels.

7. The method for perceiving safety risks in water conservancy construction according to claim 1, characterized in that, By utilizing long short-term memory networks to perform time-series prediction on standardized time-series monitoring sequences of key monitoring parameters, predicted values ​​of key parameters for future periods are obtained, including: Standardized monitoring sequences and meteorological forecast data from multiple consecutive time steps are used as inputs; Output the predicted values ​​of key parameters for multiple future time steps, and output the confidence interval for each prediction result; When the actual observed value falls outside the confidence interval multiple times consecutively, the online retraining process of the model is automatically triggered.

8. The method for perceiving safety risks in water conservancy construction according to claim 1, characterized in that, By integrating the risk level distribution map with the predicted values ​​of key parameters, a four-dimensional risk evolution cloud map is generated through spatiotemporal correlation analysis. Based on preset rules, the current overall construction safety status level is determined, including: By piecing together the current risk distribution with the results of future risk projections along a timeline, a complete risk evolution sequence is formed; The four-dimensional risk evolution cloud map is rendered using a color and transparency overlay method. The built-in rules engine determines a major emergency when specific risk clusters or parameters exceed limits, and activates the emergency plan push channel.

9. A water conservancy construction safety risk perception system, characterized in that, include: The multi-parameter sensing unit is used to deploy various types of sensors at the water conservancy construction site to collect multi-dimensional physical state data of the construction object and its surrounding environment. The status monitoring unit, connected to the multi-parameter sensing unit, is used to perform time synchronization, outlier removal, signal filtering and unit normalization on the collected raw data, and generate a standardized time-series monitoring sequence. The health assessment unit receives data output from the status monitoring unit, uses a method based on the deviation of historical baseline status comparison and current measured value calculation to quantitatively assess the structural health index of each monitoring point or monitoring area, and outputs a graded health status identifier. The twin modeling unit constructs a three-dimensional digital twin based on water conservancy engineering construction drawings, geological survey data and BIM models, and dynamically maps the standardized time-series monitoring sequence to the corresponding spatial location to realize data-driven linkage updates between physical entities and virtual models; The dynamic simulation unit integrates a finite element analysis engine and a material constitutive relation library. Based on the twin modeling unit, it applies measured load boundary conditions to perform nonlinear static and dynamic response simulations, simulating the evolution of the internal stress field, displacement field, and plastic zone expansion path of the structure under different working conditions. The risk assessment unit combines the structural health index output by the health assessment unit with the mechanical response index generated by the dynamic simulation unit, and uses a fuzzy comprehensive evaluation algorithm to fuse multi-source evidence to output a spatially gridded risk level distribution map. The hazard prediction unit, based on a long short-term memory network architecture, trains a time-series prediction model. The input is a standardized monitoring sequence and meteorological forecast data for multiple consecutive time steps, and the output is the predicted value of key parameters for multiple future time steps. The situation awareness unit integrates the results of the risk assessment unit and the hazard prediction unit, generates a four-dimensional risk evolution cloud map through spatiotemporal overlay analysis, determines the current overall construction safety situation level according to preset rules, and outputs visualized alarm information and handling suggestions.

10. The water conservancy construction safety risk perception system according to claim 9, characterized in that, The dynamic simulation unit is also equipped with a material degradation model, which automatically reduces the elastic modulus of a unit when the cumulative damage of a unit reaches a preset threshold during the simulation iteration process, and continuously tracks the cumulative damage effect; the hazard prediction unit is also equipped with a confidence interval feedback module, which outputs a confidence band for each prediction result, and automatically triggers the online retraining process of the model when the actual observation value falls outside the confidence band multiple times in a row.

Images