Hydraulic engineering intelligent monitoring method and system based on digital twinning
By using digital twin technology to collect multi-source heterogeneous data in water conservancy projects, establishing a dual-channel deep learning network model and performing feature fusion, the problems of low data utilization and insufficient risk assessment accuracy in water conservancy project monitoring have been solved. This has enabled full-coverage data collection and risk visualization, thereby improving construction safety and management level.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LEGER TECH SERVICES LTD
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-08
AI Technical Summary
Existing water conservancy project monitoring technologies suffer from problems such as low efficiency of manual inspections, low data utilization, insufficient accuracy of risk assessment, insufficient data fusion, inability to achieve all-weather monitoring, inability to capture subtle structural deformations and risk signs in a timely manner, and poor coordination between traditional monitoring and engineering design.
An intelligent monitoring method based on digital twins is adopted. By collecting multi-source heterogeneous data, a dual-channel deep learning network model is established. Data processing is carried out using an LSTM-Transformer hybrid network and a YOLOv7+ResNet50 dual-branch network. Feature fusion is performed by combining the DWA dynamic weighted attention mechanism, and hyperparameters are optimized by the GA genetic algorithm. Finally, a digital twin model is constructed based on BIM+GIS for simulation operation.
It has achieved full coverage of data collection and fusion, improved the accuracy and reliability of risk assessment, reduced prediction errors, provided a visualized risk display, and improved construction safety and management level.
Smart Images
Figure CN121998421A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water conservancy project monitoring technology, and in particular to an intelligent monitoring method and system for water conservancy projects based on digital twins. Background Technology
[0002] As a crucial infrastructure project for the national economy, the safety and stability of water conservancy projects directly impact project quality and the safety of lives and property in surrounding areas. Currently, manual inspections are inefficient, subjective, and lack comprehensive, round-the-clock monitoring. They pose safety hazards in dangerous areas such as deep foundation pits and high slopes, and cannot promptly capture subtle structural deformations and risk signs. Single-sensor monitoring data has limited dimensions and cannot comprehensively reflect the correlation between construction conditions, environmental changes, and the structural state of the project, resulting in low data utilization. Furthermore, existing monitoring technologies lack effective multi-source data fusion methods, making it difficult for time-series data and image data to complement each other, leading to insufficient accuracy in risk assessment. In addition, traditional monitoring suffers from poor coordination with engineering design and construction management, failing to achieve visualized risk presentation and dynamic simulation prediction, thus hindering accurate and real-time support for construction decisions. Summary of the Invention
[0003] The purpose of this invention is to solve the above problems by designing an intelligent monitoring method and system for water conservancy projects based on digital twins.
[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned intelligent monitoring method for water conservancy projects based on digital twins: Collect time-series sensor data, video image data, environmental data, and historical working condition data during the construction of water conservancy projects to obtain multi-source heterogeneous data. Preprocess the multi-source heterogeneous data to obtain initial multi-source heterogeneous data. A dual-channel deep learning network model was established. An improved LSTM-Transformer hybrid network was used to process time-series data, and a YOLOv7+ResNet50 dual-branch network was used to process image data. The dual-channel features were weighted and fused through the DWA dynamic weighted attention mechanism. The hyperparameters of the dual-channel deep learning network model are optimized using the GA genetic algorithm to obtain the target dual-channel deep learning network model. The initial multi-source heterogeneous data is then input into the target dual-channel deep learning network model to output the water conservancy project construction risk index. A digital twin model of a water conservancy project is constructed based on BIM+GIS. The construction risk index of the water conservancy project is input into the digital twin model for simulation operation, and a construction calibration report is output.
[0005] Furthermore, in the aforementioned intelligent monitoring method for water conservancy projects based on digital twins, the collection of time-series sensor data, video image data, environmental data, and historical operating condition data during the construction of the water conservancy project yields multi-source heterogeneous data. The multi-source heterogeneous data is then preprocessed to obtain initial multi-source heterogeneous data, including: Collect time-series sensor data, video image data, environmental data, and historical operating condition data during the construction of water conservancy projects to obtain multi-source heterogeneous data; Outliers in time-series sensor data are detected using the 3σ criterion. Data exceeding the mean ± 3 times the standard deviation are identified as outliers and replaced by linear interpolation of adjacent time-series data. Missing values are filled using the mean imputation method.
[0006] Furthermore, in the aforementioned intelligent monitoring method for water conservancy projects based on digital twins, the collection of time-series sensor data, video image data, environmental data, and historical operating condition data during the construction of the water conservancy project yields multi-source heterogeneous data. The multi-source heterogeneous data is then preprocessed to obtain initial multi-source heterogeneous data, including: The video image data is cleaned by image quality assessment algorithms to remove duplicate records of environmental data and historical working condition data, and erroneous data is corrected by logical verification. The Z-score standardization method is used to convert time series data of different dimensions into standard normal distribution data with a mean of 0 and a standard deviation of 1. The pixel values of video image data are normalized to the [0,1] interval. The historical working condition data of the classification class are numerically converted using one-hot encoding. Using the timestamps of time-series sensor data as a reference, the acquisition times of video image data and environmental data are aligned to obtain initial multi-source heterogeneous data.
[0007] Furthermore, in the aforementioned intelligent monitoring method for water conservancy projects based on digital twins, the establishment of a dual-channel deep learning network model, the use of an improved LSTM-Transformer hybrid network to process time-series data, the use of a YOLOv7+ResNet50 dual-branch network to process image data, and the weighted fusion of dual-channel features through a DWA dynamic weighted attention mechanism, includes: Based on the traditional LSTM forget gate, input gate, and output gate, a temporal attention gate is added. By calculating the attention weights of data at different time steps, the extraction of temporal features at key time nodes is enhanced. The temporal feature sequence output from the LSTM layer is input into the Transformer encoder, which contains multiple attention heads, each with a dimension of 64. The multi-head attention mechanism captures the long-distance dependencies between temporal features.
[0008] Furthermore, in the aforementioned intelligent monitoring method for water conservancy projects based on digital twins, the establishment of a dual-channel deep learning network model, the use of an improved LSTM-Transformer hybrid network to process time-series data, the use of a YOLOv7+ResNet50 dual-branch network to process image data, and the weighted fusion of dual-channel features through a DWA dynamic weighted attention mechanism, includes: The object detection branch is based on the YOLOv7 network, with an added small object detection head and optimized anchor box size. The feature enhancement branch uses the ResNet50 network and introduces residual connections to solve the gradient vanishing problem in deep networks. The temporal feature vector and image feature vector are input into the module through the DWA dynamic weighted attention mechanism. The initial weights of the channels are determined by calculating the information entropy and mutual information of the two features. Based on the initial weights, the weights are dynamically adjusted through an attention mechanism to calculate the temporal feature weights and image feature weights. The dual-channel features are then fused using a weighted summation method to output a fused feature vector.
[0009] Furthermore, in the aforementioned intelligent monitoring method for water conservancy projects based on digital twins, the hyperparameters of the dual-channel deep learning network model are optimized using the GA genetic algorithm to obtain the target dual-channel deep learning network model. The initial multi-source heterogeneous data is then input into the target dual-channel deep learning network model, and the water conservancy project construction risk index is output, including: Hyperparameters include the number of hidden units, dropout rate, and number of Transformer attention heads in the improved LSTM-Transformer network; the learning rate, batch size, and anchor box size parameters in the YOLOv7+ResNet50 dual-branch network; and the weight adjustment coefficients of the DWA attention mechanism. With the goal of minimizing the prediction error of construction risks in water conservancy projects, a fitness function is constructed, and selection, crossover, and mutation operations of a genetic algorithm are executed to iteratively output the optimal combination of hyperparameters.
[0010] Furthermore, in the aforementioned intelligent monitoring method for water conservancy projects based on digital twins, the step of constructing a digital twin model of the water conservancy project based on BIM+GIS, inputting the construction risk index of the water conservancy project into the digital twin model for simulation operation, and outputting a construction calibration report includes: The construction risk index of water conservancy projects is integrated into the digital twin model in real time. After establishing the correlation between the construction risk index of water conservancy projects and the corresponding monitoring areas and components in the model, simulation operation is carried out to obtain the operation results. Based on the simulation operation results, a construction calibration report is compiled.
[0011] Furthermore, in a digital twin-based intelligent monitoring system for water conservancy projects, the intelligent monitoring system includes the following modules: The multi-source data acquisition module is used to collect time-series sensor data, video image data, environmental data and historical working condition data during the construction of water conservancy projects to obtain multi-source heterogeneous data. The multi-source heterogeneous data is preprocessed to obtain initial multi-source heterogeneous data. The dual-channel model building module is used to build a dual-channel deep learning network model. It uses an improved LSTM-Transformer hybrid network to process time-series data and a YOLOv7+ResNet50 dual-branch network to process image data. It uses the DWA dynamic weighted attention mechanism to perform weighted fusion of dual-channel features. The construction risk prediction module is used to optimize the hyperparameters of the dual-channel deep learning network model using the GA genetic algorithm to obtain the target dual-channel deep learning network model. The initial multi-source heterogeneous data is input into the target dual-channel deep learning network model, and the construction risk index of water conservancy projects is output. The digital twin simulation module is used to build a digital twin model of a water conservancy project based on BIM+GIS. The construction risk index of the water conservancy project is input into the digital twin model of the water conservancy project for simulation operation, and a construction calibration report is output.
[0012] Furthermore, in a digital twin-based intelligent monitoring system for water conservancy projects, the construction risk prediction module includes the following sub-modules: The parameter submodule is used for hyperparameters including the number of hidden units, dropout rate, and number of Transformer attention heads for the improved LSTM-Transformer network; the learning rate, batch size, and anchor box size parameters for the YOLOv7+ResNet50 dual-branch network; and the weight adjustment coefficients for the DWA attention mechanism. The optimization submodule is used to construct a fitness function with the goal of minimizing the prediction error of construction risks in water conservancy projects, and to perform selection, crossover, and mutation operations of a genetic algorithm to iteratively output the optimal combination of hyperparameters.
[0013] Furthermore, in a digital twin-based intelligent monitoring system for water conservancy projects, the construction risk prediction module includes the following sub-modules: The simulation submodule is used to connect the construction risk index of water conservancy projects to the digital twin model in real time, establish the correlation between the construction risk index of water conservancy projects and the corresponding monitoring areas and components in the model, and then conduct simulation operation to obtain the operation results. Based on the simulation operation results, a construction calibration report is compiled.
[0014] Its beneficial effects lie in achieving full coverage of construction time-series data, image data, environmental data, and historical working condition data. Standardization and fusion processing eliminate data barriers, improve data quality, and provide comprehensive data support for risk assessment, solving the problems of single-dimensional and low-utilization traditional monitoring data. The construction of a dual-channel deep learning network model enables targeted feature extraction from time-series and image data. Combined with the DWA dynamic weighted attention mechanism fusion strategy, the weight of key features is strengthened, improving the rationality and effectiveness of feature fusion. Hyperparameter optimization using the GA genetic algorithm further improves the accuracy and reliability of risk index prediction, reducing prediction errors by more than 30% compared to traditional single-model predictions. The BIM+GIS integrated digital twin model enables visualized mapping and dynamic simulation of the risk index, transforming abstract risk data into an intuitive display of engineering status, helping construction personnel accurately locate high-risk areas and their causes. This effectively improves the safety and refined management level of water conservancy engineering construction and reduces the incidence of construction risk accidents. Attached Figure Description
[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.
[0016] Figure 1 This is a schematic diagram of the first embodiment of an intelligent monitoring method for water conservancy projects based on digital twins according to the present invention; Figure 2 This is a schematic diagram of a second embodiment of an intelligent monitoring method for water conservancy projects based on digital twins according to the present invention; Figure 3 This is a schematic diagram of the first embodiment of an intelligent monitoring system for water conservancy projects based on digital twins, as described in this invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.
[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, an intelligent monitoring method for water conservancy projects based on digital twins includes the following steps: Step 101: Collect time-series sensor data, video image data, environmental data and historical working condition data during the construction of water conservancy projects to obtain multi-source heterogeneous data. Preprocess the multi-source heterogeneous data to obtain initial multi-source heterogeneous data. Specifically, in this embodiment, time-series sensor data, video image data, environmental data, and historical operating condition data are collected during the construction of water conservancy projects to obtain multi-source heterogeneous data. The 3σ criterion is used to detect outliers in the time-series sensor data; data exceeding the mean ± 3 times the standard deviation are identified as outliers and replaced using linear interpolation of adjacent time-series data. Mean-filling is used to fill in missing values. Video image data is cleaned using an image quality assessment algorithm to remove duplicate records from environmental and historical operating condition data, and erroneous data is corrected through logical verification. The time-series data of different dimensions are converted to standard normal distribution data with a mean of 0 and a standard deviation of 1 using the Z-score standardization method. The pixel values of the video image data are normalized to the [0,1] interval, and the historical operating condition data of the classification category is numerically converted using one-hot encoding. The acquisition times of the video image data and environmental data are aligned using the timestamps of the time-series sensor data to obtain the initial multi-source heterogeneous data.
[0020] For key components of water conservancy engineering construction, such as dam foundation pits, spillways, diversion tunnels, and pile foundations, various types of time-series sensors, including vibrating wire strain gauges, rebar gauges, piezometers, displacement sensors, and tilt sensors, are deployed. Sensor deployment follows the principle of full coverage of key risk points and increased density in key areas. For example, one displacement sensor is deployed every 5 meters along the slope of the foundation pit, and strain gauges and piezometers are deployed at key sections of the dam body. Data acquisition units (DTUs) enable real-time acquisition of sensor data. The acquisition frequency is dynamically adjusted according to the construction stage: once every 10 minutes during foundation construction, and increased to once every 2 minutes during critical processes such as concrete pouring. The acquired data includes core time-series indicators such as stress, strain, displacement, seepage flow, and vibration frequency. The data is transmitted to edge computing nodes for temporary storage via 5G / fiber optic networks.
[0021] High-definition network cameras and infrared cameras are deployed at construction site entrances and exits, core work areas such as concrete pouring areas and hoisting areas, and hazardous areas such as high slopes and deep foundation pits. Cameras are installed at a height of no less than 8 meters to ensure comprehensive monitoring. Infrared cameras are used for data collection at night or in low-visibility environments. Camera parameters are set as follows: resolution 1920×1080, frame rate 25fps, with motion detection enabled to automatically improve recording clarity when personnel or equipment move within the monitored area. Video image data is stored in H.265 encoding format, and keyframes are extracted in real time to reduce data storage pressure. The collected content includes the operating status of construction equipment, the standardization of personnel operations, and structural defects.
[0022] Automatic weather stations, water quality monitors, and soil moisture sensors were deployed around the construction area and on the work surface to collect key data on temperature, humidity, wind speed, rainfall, light intensity, water pH, turbidity, and soil moisture content. Meteorological data was collected every 30 minutes, water quality data every 2 hours, and soil moisture data every hour. After data collection, the time and location of collection were simultaneously labeled to ensure data traceability.
[0023] By reviewing construction drawings, logs, supervision reports, and historical monitoring data archives of water conservancy projects, data such as process parameters, equipment operating parameters, quality inspection data, and historical risk event records from historical construction stages were collected. Paper archive data was scanned and converted into electronic data using OCR, and the electronic archive data was then formatted to create a historical construction data database.
[0024] The data cleaning process employs outlier detection, missing value imputation, and duplicate value removal. For time-series sensor data, the 3σ criterion is used to detect outliers; data exceeding the mean ± 3 standard deviations are identified as outliers and replaced using linear interpolation of adjacent time-series data. For missing values, if the missing rate is less than 5%, mean imputation is used; for a missing rate of 5%-20%, linear interpolation is used; and for a missing rate greater than 20%, the corresponding data segment is removed and the reason is noted. For video image data, blurry, occluded, overexposed / underexposed images are screened using image quality assessment algorithms and then removed or enhanced. For environmental data and historical operating condition data, duplicate records are removed, and erroneous data is corrected through logical verification.
[0025] For time-series data of different dimensions, the Z-score standardization method is used to convert the data into standard normal distribution data with a mean of 0 and a standard deviation of 1, eliminating the difference in dimensions; for video image data, the pixel values are normalized to the [0,1] interval, and the size is unified and scaled to 640×640 pixels to facilitate model processing; for historical working condition data of classification, such as construction process type, one-hot encoding is used for numerical conversion.
[0026] Using the timestamps of time-series sensor data as a benchmark, the acquisition times of video image data and environmental data are aligned to ensure the correlation of different types of data within the same time dimension. Data fusion technology is employed to associate and label multiple types of data from the same monitoring point, forming structured data samples, ultimately resulting in initial multi-source heterogeneous data with a unified format and reliable quality.
[0027] Step 102: Establish a dual-channel deep learning network model, use an improved LSTM-Transformer hybrid network to process time series data, use a YOLOv7+ResNet50 dual-branch network to process image data, and use the DWA dynamic weighted attention mechanism to perform weighted fusion of dual-channel features. Specifically, in this embodiment, a temporal attention gate is added to the traditional LSTM forget gate, input gate, and output gate. By calculating the attention weights of data at different time steps, the extraction of temporal features at key time nodes is enhanced. The temporal feature sequence output from the LSTM layer is input into the Transformer encoder, which contains multiple attention heads, each with a dimension of 64. A multi-head attention mechanism is used to capture long-distance dependencies between temporal features. The object detection branch is based on the YOLOv7 network, with an added small object detection head and optimized anchor box size. The feature enhancement branch uses a ResNet50 network, introducing residual connections to address the vanishing gradient problem in deep networks. The temporal feature vector and image feature vector are input into the module through a DWA dynamic weighted attention mechanism. Initial channel weights are determined by calculating the information entropy and mutual information of the two types of features. Based on these initial weights, the attention mechanism dynamically adjusts the weights, calculating the temporal feature weights and image feature weights. A weighted summation method is then used to fuse the dual-channel features, outputting a fused feature vector.
[0028] This channel employs an LSTM feature extraction layer + Transformer attention enhancement layer + fully connected output layer structure. The LSTM layer is improved and optimized by adding a temporal attention gate to the traditional LSTM forget gate, input gate, and output gate. This gate calculates attention weights for data at different time steps, enhancing the extraction of temporal features for key time nodes, such as construction process transitions and extreme weather periods. The LSTM layer has two hidden layers, each with 256 hidden units, and a dropout rate of 0.2 to prevent overfitting.
[0029] The temporal feature sequence output from the LSTM layer is input into the Transformer encoder, which contains 6 attention heads, each with a dimension of 64. The multi-head attention mechanism captures the long-distance dependencies between temporal features. A fully connected layer is then connected, with ReLU activation function and layer normalization layer to further optimize feature representation.
[0030] The preprocessed time-series sensor data and environmental data are divided into sample sequences of length 60 according to the time series, and input into the improved LSTM-Transformer hybrid network. After feature extraction, the output is a time-series feature vector with a dimension of 512.
[0031] This channel is divided into an object detection branch and a feature enhancement branch. The object detection branch is based on an improvement of the YOLOv7 network. By adding a small object detection head and optimizing the anchor frame size, it improves the detection accuracy of various objects in the construction scene. The feature enhancement branch uses the ResNet50 network. By introducing residual connections, it solves the gradient vanishing problem in deep networks and further deepens and extracts features from the feature map output by the YOLOv7 network, thereby enhancing the discriminative power of the features.
[0032] A transfer learning strategy was employed, with pre-trained weights loaded onto the YOLOv7 branch and trained on the COCO dataset. Pre-trained weights were also loaded onto the ResNet50 branch, and then fine-tuned using pre-processed video keyframe data. Data augmentation techniques such as random cropping, flipping, and color gamut transformation were used during training. The SGD optimizer was selected, with an initial learning rate of 0.001. Cosine annealing was used to adjust the learning rate, and the training iterations totaled 100 epochs.
[0033] The trained YOLOv7+ResNet50 dual-branch network performs object detection and feature extraction on the input image data, concatenates and fuses the feature maps output by the two branches, and converts them into image feature vectors of dimension 512 through a single global average pooling layer.
[0034] A dynamic weighted attention module is constructed, inputting temporal feature vectors and image feature vectors. The initial weights of each channel are determined by calculating the information entropy and mutual information of the two types of features. For example, when the temporal data fluctuates drastically at a certain moment, such as a sudden stress change, the weight of the temporal channel is increased; when a significant structural defect is detected in the image, the weight of the image channel is increased. Based on the initial weights, dynamic adjustments are made through an attention mechanism to calculate the temporal feature weight ω1 and the image feature weight ω2, where ω1 + ω2 = 1. Then, a weighted summation method is used to fuse the dual-channel features: fused feature vector = ω1 × temporal feature vector + ω2 × image feature vector. The final output is a fused feature vector with a dimension of 512, achieving complementary enhancement of the two data features.
[0035] Step 103: Optimize the hyperparameters of the dual-channel deep learning network model using the GA genetic algorithm to obtain the target dual-channel deep learning network model. Input the initial multi-source heterogeneous data into the target dual-channel deep learning network model and output the water conservancy project construction risk index. Specifically, in this embodiment, the hyperparameters include the number of hidden layer units, dropout rate, and number of Transformer attention heads in the improved LSTM-Transformer network; the learning rate, batch size, and anchor box size parameters in the YOLOv7+ResNet50 dual-branch network; and the weight adjustment coefficients of the DWA attention mechanism. With the goal of minimizing the prediction error of water conservancy engineering construction risks, a fitness function is constructed, and the selection, crossover, and mutation operations of the genetic algorithm are executed to iteratively output the optimal combination of hyperparameters.
[0036] The set of hyperparameters to be optimized was clearly defined, including the number of hidden units, dropout rate, and number of Transformer attention heads for improving the LSTM-Transformer network; the learning rate, batch size, and anchor box size parameters for the YOLOv7+ResNet50 network; and the weight adjustment coefficients for the DWA attention mechanism. The hyperparameters were converted into chromosomes using binary encoding, with each chromosome corresponding to a set of hyperparameter combinations. The population size was set to 50.
[0037] With the goal of minimizing the prediction error of construction risks in water conservancy projects, a fitness function is constructed. 80% of the preprocessed initial multi-source heterogeneous data is selected as the training set and 20% as the validation set. After training the model corresponding to each set of hyperparameters on the training set, the predicted risk value and the actual risk value are calculated on the validation set. Based on the mean squared error (MSE) of historical risk events, the fitness value is calculated as 1 / (1+MSE). A larger fitness value indicates better model performance.
[0038] The genetic algorithm performs selection, crossover, and mutation operations. Selection uses a roulette wheel selection method, prioritizing chromosomes with high fitness values for the next generation. Crossover uses a single-point crossover method with a crossover probability of 0.8. Mutation uses random bit flipping with a mutation probability of 0.01 to avoid getting trapped in local optima. The iteration terminates when the number of iterations reaches 100 rounds, or the fitness value increases by less than 0.001 for 10 consecutive rounds; at this point, the optimal hyperparameter combination is output.
[0039] The optimized hyperparameters are configured into the dual-channel deep learning network model, and the model is retrained using the complete training set. A validation set and a test set are used, with 10% of the initial data allocated separately as a test set to validate the model's performance. When the model's prediction accuracy on the test set reaches over 90%, it is selected as the target dual-channel deep learning network model. Preprocessed initial multi-source heterogeneous data is then input into the target dual-channel deep learning network model in batches, with a batch size of 32.
[0040] The model extracts, fuses, and infers features from the input data to output a construction risk index for water conservancy projects corresponding to the monitoring period and area. The risk index ranges from 0 to 10, where 0-2 represents low risk, 2-5 represents medium risk, 5-8 represents high risk, and 8-10 represents extremely high risk. The model also outputs the constituent elements of the risk index, providing a basis for subsequent risk tracing.
[0041] Step 104: Construct a digital twin model of the water conservancy project based on BIM+GIS, input the construction risk index of the water conservancy project into the digital twin model of the water conservancy project for simulation operation, and output a construction calibration report.
[0042] Specifically, in this embodiment, the construction risk index of water conservancy projects is connected to the digital twin model in real time. After establishing the correlation between the construction risk index of water conservancy projects and the corresponding monitoring areas and components in the model, the simulation is carried out to obtain the operation results. Based on the simulation results, a construction calibration report is prepared.
[0043] Data interface technology is used to integrate BIM and GIS models. Component information from the BIM model is mapped to geospatial information from the GIS model; for example, BIM components of a dam are linked to the corresponding regional topographic GIS data, enabling multi-scale visualization from component level to region level. Simultaneously, a real-time data transmission interface is integrated, laying the foundation for subsequent integration of risk indices, ultimately forming a complete digital twin model of the water conservancy project.
[0044] Through the data integration module of the digital twin platform, the water conservancy engineering construction risk index output by the target dual-channel deep learning network model is connected to the digital twin model in real time, establishing the correlation between the risk index and the corresponding monitoring area and components in the model. For example, the risk index of a certain foundation pit slope is mapped to the foundation pit slope component in the model, realizing the visualization and labeling of risk information.
[0045] Based on the actual construction progress and working conditions of the water conservancy project, simulation operation parameters are set in the digital twin platform, including the simulation time step, which is synchronized with the actual construction time, with one time step per hour, boundary conditions such as environmental loads and construction loads, and constraint conditions such as component connection constraints and foundation bearing capacity constraints. During the simulation operation, component information, geographical information, and accessed risk indices from the digital twin model are called in real time to simulate the changes in the construction status of the project under different risk levels.
[0046] During the simulation, key indicators such as stress distribution, displacement change, and stability status of each component in the model are monitored in real time. When the risk index of a certain area reaches the medium-high risk level, the construction parameters of that area, such as the correlation between pouring speed, support strength and risk index, are analyzed in detail to identify the key factors that lead to increased risk.
[0047] Based on the simulation results, a construction calibration report is prepared, which mainly includes the following: ① Project overview and monitoring scope; ② Risk index distribution and level assessment of each monitoring area; ③ Specific location and risk cause analysis of high-risk areas, combined with the characteristic contributions of time series data and image data; ④ Construction parameter calibration suggestions, such as adjusting concrete pouring strength, strengthening slope support measures, and optimizing construction sequence; ⑤ Rectification time limit and verification plan, such as re-monitoring the risk index after rectification to ensure that it is reduced to a low-risk level.
[0048] The construction calibration report is output in PDF format and simultaneously visualized on a digital twin platform. Different colors are used to indicate the risk level of each area: low risk green, medium risk yellow, high risk red, and extremely high risk dark red. This visually presents the risk distribution and construction calibration recommendations, providing accurate decision support for construction managers.
[0049] Its beneficial effects lie in achieving full coverage of construction time-series data, image data, environmental data, and historical working condition data. Standardization and fusion processing eliminate data barriers, improve data quality, and provide comprehensive data support for risk assessment, solving the problems of single-dimensional and low-utilization traditional monitoring data. The construction of a dual-channel deep learning network model enables targeted feature extraction from time-series and image data. Combined with the DWA dynamic weighted attention mechanism fusion strategy, the weight of key features is strengthened, improving the rationality and effectiveness of feature fusion. Hyperparameter optimization using the GA genetic algorithm further improves the accuracy and reliability of risk index prediction, reducing prediction errors by more than 30% compared to traditional single-model predictions. The BIM+GIS integrated digital twin model enables visualized mapping and dynamic simulation of the risk index, transforming abstract risk data into an intuitive display of engineering status, helping construction personnel accurately locate high-risk areas and their causes. This effectively improves the safety and refined management level of water conservancy engineering construction and reduces the incidence of construction risk accidents.
[0050] Please see Figure 2 In a digital twin-based intelligent monitoring method for water conservancy projects, time-series sensor data, video image data, environmental data, and historical operating condition data are collected during the construction of the water conservancy project to obtain multi-source heterogeneous data. Preprocessing this multi-source heterogeneous data to obtain initial multi-source heterogeneous data includes the following steps. Step 201: Clean the video image data using an image quality assessment algorithm to remove duplicate records of environmental data and historical working condition data, and correct erroneous data through logical verification. Step 202: Use the Z-score standardization method to convert the time series data of different dimensions into standard normal distribution data with a mean of 0 and a standard deviation of 1. Normalize the pixel values of video image data to the [0,1] interval. Use one-hot encoding to convert the historical working condition data of the classification class into numerical values. Step 203: Using the timestamps of the time-series sensor data as a reference, align the acquisition times of the video image data and environmental data to obtain initial multi-source heterogeneous data.
[0051] The above describes an embodiment of the intelligent monitoring method for water conservancy projects based on digital twins according to the present invention. Please refer to [link / reference]. Figure 3 In a digital twin-based intelligent monitoring system for water conservancy projects, the system includes the following modules: The multi-source data acquisition module is used to collect time-series sensor data, video image data, environmental data and historical working condition data during the construction of water conservancy projects, to obtain multi-source heterogeneous data. The multi-source heterogeneous data is preprocessed to obtain initial multi-source heterogeneous data. The dual-channel model building module is used to build a dual-channel deep learning network model. It uses an improved LSTM-Transformer hybrid network to process time-series data and a YOLOv7+ResNet50 dual-branch network to process image data. It uses the DWA dynamic weighted attention mechanism to perform weighted fusion of dual-channel features. The construction risk prediction module is used to optimize the hyperparameters of the dual-channel deep learning network model using the GA genetic algorithm to obtain the target dual-channel deep learning network model. The initial multi-source heterogeneous data is input into the target dual-channel deep learning network model, and the construction risk index of water conservancy projects is output. The digital twin simulation module is used to build a digital twin model of water conservancy projects based on BIM+GIS. The construction risk index of water conservancy projects is input into the digital twin model of water conservancy projects for simulation operation, and a construction calibration report is output.
[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and equivalents.
Claims
1. A method for intelligent monitoring of water conservancy projects based on digital twins, characterized in that, The intelligent monitoring method for water conservancy projects includes the following steps: Collect time-series sensor data, video image data, environmental data, and historical working condition data during the construction of water conservancy projects to obtain multi-source heterogeneous data. Preprocess the multi-source heterogeneous data to obtain initial multi-source heterogeneous data. A dual-channel deep learning network model was established. An improved LSTM-Transformer hybrid network was used to process time-series data, and a YOLOv7+ResNet50 dual-branch network was used to process image data. The dual-channel features were weighted and fused through the DWA dynamic weighted attention mechanism. The hyperparameters of the dual-channel deep learning network model are optimized using the GA genetic algorithm to obtain the target dual-channel deep learning network model. The initial multi-source heterogeneous data is then input into the target dual-channel deep learning network model to output the water conservancy project construction risk index. A digital twin model of a water conservancy project is constructed based on BIM+GIS. The construction risk index of the water conservancy project is input into the digital twin model for simulation operation, and a construction calibration report is output.
2. The intelligent monitoring method for water conservancy projects based on digital twins as described in claim 1, characterized in that, The process involves collecting time-series sensor data, video image data, environmental data, and historical operating condition data during the construction of the water conservancy project to obtain multi-source heterogeneous data. This multi-source heterogeneous data is then preprocessed to obtain initial multi-source heterogeneous data, including: Collect time-series sensor data, video image data, environmental data, and historical operating condition data during the construction of water conservancy projects to obtain multi-source heterogeneous data; Outliers in time-series sensor data are detected using the 3σ criterion. Data exceeding the mean ± 3 times the standard deviation are identified as outliers and replaced by linear interpolation of adjacent time-series data. Missing values are filled using the mean imputation method.
3. The intelligent monitoring method for water conservancy projects based on digital twins as described in claim 1, characterized in that, The process involves collecting time-series sensor data, video image data, environmental data, and historical operating condition data during the construction of the water conservancy project to obtain multi-source heterogeneous data. This multi-source heterogeneous data is then preprocessed to obtain initial multi-source heterogeneous data, including: The video image data is cleaned by image quality assessment algorithms to remove duplicate records of environmental data and historical working condition data, and erroneous data is corrected by logical verification. The Z-score standardization method is used to convert time series data of different dimensions into standard normal distribution data with a mean of 0 and a standard deviation of 1. The pixel values of video image data are normalized to the [0,1] interval. The historical working condition data of the classification class are numerically converted using one-hot encoding. Using the timestamps of time-series sensor data as a reference, the acquisition times of video image data and environmental data are aligned to obtain initial multi-source heterogeneous data.
4. The intelligent monitoring method for water conservancy projects based on digital twins as described in claim 1, characterized in that, The establishment of a dual-channel deep learning network model utilizes an improved LSTM-Transformer hybrid network to process temporal data and a YOLOv7+ResNet50 dual-branch network to process image data. It employs a DWA dynamic weighted attention mechanism to weightedly fuse dual-channel features, including: Based on the traditional LSTM forget gate, input gate, and output gate, a temporal attention gate is added. By calculating the attention weights of data at different time steps, the extraction of temporal features at key time nodes is enhanced. The temporal feature sequence output from the LSTM layer is input into the Transformer encoder, which contains multiple attention heads, each with a dimension of 64. The multi-head attention mechanism captures the long-distance dependencies between temporal features.
5. The intelligent monitoring method for water conservancy projects based on digital twins as described in claim 1, characterized in that, The establishment of a dual-channel deep learning network model utilizes an improved LSTM-Transformer hybrid network to process temporal data and a YOLOv7+ResNet50 dual-branch network to process image data. It employs a DWA dynamic weighted attention mechanism to weightedly fuse dual-channel features, including: The object detection branch is based on the YOLOv7 network, with an added small object detection head and optimized anchor box size. The feature enhancement branch uses the ResNet50 network and introduces residual connections to solve the gradient vanishing problem in deep networks. The temporal feature vector and image feature vector are input into the module through the DWA dynamic weighted attention mechanism. The initial weights of the channels are determined by calculating the information entropy and mutual information of the two features. Based on the initial weights, the weights are dynamically adjusted through an attention mechanism to calculate the temporal feature weights and image feature weights. The dual-channel features are then fused using a weighted summation method to output a fused feature vector.
6. The intelligent monitoring method for water conservancy projects based on digital twins as described in claim 1, characterized in that, The hyperparameters of the dual-channel deep learning network model are optimized using the GA genetic algorithm to obtain the target dual-channel deep learning network model. The initial multi-source heterogeneous data is input into the target dual-channel deep learning network model, and the water conservancy project construction risk index is output, including: Hyperparameters include the number of hidden units, dropout rate, and number of Transformer attention heads in the improved LSTM-Transformer network; the learning rate, batch size, and anchor box size parameters in the YOLOv7+ResNet50 dual-branch network; and the weight adjustment coefficients of the DWA attention mechanism. With the goal of minimizing the prediction error of construction risks in water conservancy projects, a fitness function is constructed, and selection, crossover, and mutation operations of a genetic algorithm are executed to iteratively output the optimal combination of hyperparameters.
7. The intelligent monitoring method for water conservancy projects based on digital twins as described in claim 1, characterized in that, The digital twin model of the water conservancy project, constructed based on BIM+GIS, is used to simulate the construction risk index of the water conservancy project and outputs a construction calibration report, including: The construction risk index of water conservancy projects is integrated into the digital twin model in real time. After establishing the correlation between the construction risk index of water conservancy projects and the corresponding monitoring areas and components in the model, simulation operation is carried out to obtain the operation results. Based on the simulation operation results, a construction calibration report is compiled.
8. A digital twin-based intelligent monitoring system for water conservancy projects, characterized in that, The intelligent monitoring system for water conservancy projects includes the following modules: The multi-source data acquisition module is used to collect time-series sensor data, video image data, environmental data and historical working condition data during the construction of water conservancy projects to obtain multi-source heterogeneous data. The multi-source heterogeneous data is preprocessed to obtain initial multi-source heterogeneous data. The dual-channel model building module is used to build a dual-channel deep learning network model. It uses an improved LSTM-Transformer hybrid network to process time-series data and a YOLOv7+ResNet50 dual-branch network to process image data. It uses the DWA dynamic weighted attention mechanism to perform weighted fusion of dual-channel features. The construction risk prediction module is used to optimize the hyperparameters of the dual-channel deep learning network model using the GA genetic algorithm to obtain the target dual-channel deep learning network model. The initial multi-source heterogeneous data is input into the target dual-channel deep learning network model, and the construction risk index of water conservancy projects is output. The digital twin simulation module is used to build a digital twin model of a water conservancy project based on BIM+GIS. The construction risk index of the water conservancy project is input into the digital twin model of the water conservancy project for simulation operation, and a construction calibration report is output.
9. The intelligent monitoring system for water conservancy projects based on digital twins as described in claim 8, characterized in that, The construction risk prediction module Includes the following sub-modules: The parameter submodule is used for hyperparameters including the number of hidden units, dropout rate, and number of Transformer attention heads for the improved LSTM-Transformer network; the learning rate, batch size, and anchor box size parameters for the YOLOv7+ResNet50 dual-branch network; and the weight adjustment coefficients for the DWA attention mechanism. The optimization submodule is used to construct a fitness function with the goal of minimizing the prediction error of construction risks in water conservancy projects, and to perform selection, crossover, and mutation operations of a genetic algorithm to iteratively output the optimal combination of hyperparameters.
10. The intelligent monitoring system for water conservancy projects based on digital twins as described in claim 8, characterized in that, The construction risk prediction module Includes the following sub-modules: The simulation submodule is used to connect the construction risk index of water conservancy projects to the digital twin model in real time, establish the correlation between the construction risk index of water conservancy projects and the corresponding monitoring areas and components in the model, and then conduct simulation operation to obtain the operation results. Based on the simulation operation results, a construction calibration report is compiled.