Water conservancy monitoring system and method based on digital twinning

Abstract

The application discloses a kind of based on digital twinning water conservancy monitoring system and method, it is related to digital twinning water conservancy monitoring field, including in the bottom of monitoring section Deployment integrated pressure sensing and thermal pulse probe composite sensor, release heat according to cycle and record temperature decay time curve and pressure value;Temperature decay time curve is input into heat dissipation characteristic matching model output medium category probability;When category instruction is silt burying mode and there is pressure increment, increment is input into silt thickness conversion template library to obtain corresponding silt thickness;Finally, in digital twinning body, maintain water surface elevation boundary unchanged, and corresponding twinning body riverbed elevation boundary is deduced the silt thickness upwards.This application solves the problem that the water level calculation of digital twinning body is wrong and the riverbed terrain boundary is seriously deviated from the physical actual environment due to the bottom sensor being buried by silt in traditional water conservancy monitoring.

Description

Technical Field

[0001] This invention relates to the field of digital twin water conservancy monitoring, and more specifically, to a water conservancy monitoring system and method based on digital twins. Background Technology

[0002] Digital twin water conservancy monitoring maps the spatial form and operational status of real water conservancy projects into a virtual space. It continuously refines the virtual model using on-site sensor data, enabling managers to monitor project operations digitally and aiding in flood forecasting and risk assessment. For open water body monitoring scenarios, the digital twin needs to reflect not only the water surface elevation but also the riverbed morphology and internal flow field. Water surface elevation boundaries typically correspond to the current water level, while riverbed elevation boundaries typically correspond to changes in the riverbed topography. These two concepts have different physical meanings and will have different impacts on hydrodynamic calculation results.

[0003] Patent document CN117371337A discloses a method and system for constructing a water conservancy model based on digital twins. This scheme involves collecting data from monitoring points within a detection area, establishing a water conservancy digital twin model using a convolutional neural network model, performing data quality analysis on the monitoring data obtained from the monitoring points, filtering target data, inputting it into the water conservancy digital twin model to obtain predicted data, and selecting a correction scheme based on the degree of deviation between the predicted data and the actual data to correct the water conservancy digital twin model. This scheme mainly focuses on the construction of the water conservancy digital twin model, data quality analysis, prediction deviation assessment, and model correction.

[0004] Patent document CN118709389A discloses a dual-engine driven method for a water conservancy digital twin platform. This scheme constructs a GIS-driven visualization engine and a UE-driven visualization engine through standardized coding, eight-directional spatiotemporal correlation at the same scale, and a static visualization scene of the digital twin. This forms a three-dimensional driving relationship for water conservancy objects and a multi-dimensional data-driven processing flow. It then calls model files for calculations and performs visualization simulation based on the calculation results to obtain flood forecast analysis results. This scheme mainly focuses on the three-dimensional scene driving, model file calling, and flood forecast display of the water conservancy digital twin platform.

[0005] Existing digital twin water conservancy monitoring systems typically emphasize the integration of on-site operational data and the updating of virtual models. However, in scenarios where siltation occurs at the bottom monitoring section, relying solely on pressure sensors can easily lead to misjudgments. An increase in bottom pressure could be due to rising water levels causing increased water column pressure, or it could be caused by the sensor being gradually buried by silt, resulting in pressure increments due to the weight of the silt and pore water pressure. For the digital twin, the corresponding boundary update directions for these two scenarios are completely different. The former corresponds to changes in water surface elevation boundaries, while the latter corresponds to changes in riverbed elevation boundaries.

[0006] In actual water conservancy projects, flood discharge, heavy rainfall inflow, and changes in the operation of gates or pumping stations can all alter local flow conditions, causing rapid sediment deposition near the monitoring section. If the monitoring system cannot identify the physical causes behind the pressure increase and instead directly converts the pressure increase into a water level change, it will mistakenly interpret sedimentation as a rise in water level, leading to incorrect water depth states, cross-sectional dimensions, and bottom boundaries within the digital twin. This will not only affect the hydrodynamic calculation results but may also cause deviations in the assessment of flood peak propagation, flow velocity distribution, local backwater, and scheduling risks.

[0007] Furthermore, flow velocity data from hydraulic sites are also susceptible to interference from environmental factors. Surface velocity data acquired by surface current radar may contain surface drift components caused by wind shear stress, and Doppler shift signals acquired by underwater acoustic current meters may be affected by underwater biological activity and bubble uplift. If this interfering data is imported into the digital twin without being identified and corrected, the top dynamic boundary and internal three-dimensional flow field in the virtual model will deviate from the actual hydrodynamic state. Therefore, current technology still requires a hydraulic monitoring method that can distinguish the physical sources of bottom pressure increments and further improve the reliability of digital twin boundary updates. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a water conservancy monitoring system and method based on digital twins, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A water conservancy monitoring method based on digital twins includes: A composite sensor integrating a pressure sensing probe and a thermal pulse probe is deployed at the bottom of the monitoring section. The thermal pulse probe periodically releases heat and records the temperature decay time-series curve of the medium surrounding the composite sensor, and the pressure value of the surrounding medium is acquired by the pressure sensing probe. The temperature decay time series curve is input into the heat loss feature matching model, and the heat loss feature matching model outputs the medium category probability representing the surrounding medium category, which includes pure water mode and silt burial mode. When the medium category probability indicates a sediment burial mode, and the pressure value has an increase compared to the historical reference pressure, the pressure increase is input into the sediment thickness conversion template library, and the sediment thickness conversion template library outputs the corresponding sediment thickness. The water surface elevation boundary is kept constant in the digital twin, and the sedimentation thickness is extrapolated upward from the twin riverbed elevation boundary corresponding to the location of the composite sensor in the digital twin.

[0010] Specifically, the training process of the heat dissipation feature matching model includes: Obtain the first temperature decay time series curve under a pure water environment and the second temperature decay time series curve under a sediment burial environment. The first temperature decay time series curve is combined with the first medium category label to form the first training sample, and the second temperature decay time series curve is combined with the second medium category label to form the second training sample. The initial time series classification model is trained in a supervised manner using the first training sample and the second training sample, and the parameters are iteratively updated until convergence, thus obtaining the trained heat dissipation feature matching model.

[0011] Specifically, the step of determining the historical reference pressure includes, when the composite sensor detects that the surrounding medium has changed from pure water mode to sediment burial mode, extracting the pressure value collected in pure water mode within a preset time window before the change occurs as the historical reference pressure; the step of inputting the pressure increment into the sediment thickness conversion template library, and having the sediment thickness conversion template library output the corresponding sediment thickness, includes: A mapping table between pressure increment and corresponding sediment thickness value is pre-established in the sediment thickness conversion template library; The target thickness value that matches the input pressure increment is queried in the mapping table, and the found target thickness value is determined as the final output sediment thickness.

[0012] Specifically, when multiple composite sensors are deployed at adjacent positions at the bottom of the monitoring section, after the step of extrapolating the sedimentation thickness upwards from the twin riverbed elevation boundary corresponding to the composite sensor position in the digital twin, the method further includes: Obtain the deposition thickness output by each of the multiple composite sensors; A spatial interpolation algorithm is used to fit the surface of the sedimentation thickness output by each of the multiple composite sensors to generate a spatial sedimentation thickness distribution matrix covering the monitoring section area. The elevation values ​​of the three-dimensional mesh nodes at the bottom of the digital twin are synchronously raised and updated using the spatial accumulation thickness distribution matrix.

[0013] Specifically, the method further includes: When the medium type probability indicator is pure water mode, the pressure value collected by the pressure sensing probe is converted into water pressure head height; The water pressure head height is superimposed on the reference elevation of the twin riverbed elevation boundary to update the water surface elevation boundary inside the digital twin.

[0014] Specifically, the method further includes: Wind shear stress characteristic data were collected using an anemometer and Doppler surface current radar was used to collect the original water surface flow velocity time series signal. Frequency domain signal decomposition is performed on the original water surface flow velocity time series signal to separate the low-frequency background flow velocity sequence and the high-frequency fluctuating flow velocity sequence; The wind shear stress characteristic data and the high-frequency fluctuating velocity sequence are input together into the multimodal cross-attention mechanism model, and the multimodal cross-attention mechanism model outputs the surface wind-driven drift disturbance velocity sequence. The gravity velocity sequence is obtained by subtracting the surface wind-driven drift disturbance velocity sequence from the original water surface velocity time sequence signal. The gravity velocity sequence is used as the top dynamic boundary condition and input into the three-dimensional hydrodynamic evolution model of the digital twin for flow field calculation.

[0015] Specifically, the training process of the multimodal cross-attention mechanism model includes: Historical wind shear stress characteristic data and synchronously occurring historical water surface original flow velocity time series signals are collected, and frequency domain signal decomposition is performed on the historical water surface original flow velocity time series signals to obtain historical high-frequency fluctuation flow velocity sequences. The actual background velocity sequence of the stable flow layer below the water surface is obtained by using an acoustic Doppler current profiler, and the difference sequence between the historical original surface velocity time series signal and the actual background velocity sequence is set as the true surface wind-driven drift disturbance amount label. Using the historical wind shear stress characteristic data and the historical high-frequency fluctuating velocity sequence as input features, and the real surface wind-driven drift disturbance amount label as supervision label, the initial multimodal cross-attention mechanism model is trained in a supervised manner to obtain the trained multimodal cross-attention mechanism model.

[0016] Specifically, the method further includes: Acoustic current meters deployed underwater were used to collect Doppler frequency shift signals of current velocity at various depths. Perform a Fourier transform on the flow velocity Doppler frequency shift signal to obtain the corresponding flow velocity fluctuation spectrum; The flow velocity fluctuation spectrum is input into the fluid spectrum morphology anomaly detection model, and the fluid spectrum morphology anomaly detection model outputs anomaly detection classification results to indicate whether biological noise pollution exists. When the anomaly detection classification result indicates the presence of biological noise pollution, the spatial grid corresponding to the depth layer where the acoustic flowmeter is located in the digital twin is marked as missing. The three-dimensional spatial kriging interpolation algorithm is invoked to complete the missing spatial grid using the flow field data of the uncontaminated adjacent spatial grids in the digital twin, and the completed flow field data is assimilated into the internal three-dimensional flow field of the digital twin.

[0017] Specifically, the fluid spectral morphology anomaly detection model is constructed based on a one-dimensional convolutional autoencoder structure, and the training process of the fluid spectral morphology anomaly detection model includes: The first historical velocity fluctuation spectrum that conforms to the energy decay pattern of natural fluid dynamics is obtained as a positive sample. The second historical velocity fluctuation spectrum with independent biological rhythm high-energy frequency peaks was obtained as a negative sample. The positive samples are used to perform feature reconstruction training on the one-dimensional convolutional autoencoder structure to minimize the reconstruction error; The negative sample is input into a one-dimensional convolutional autoencoder structure trained by the feature reconstruction to calculate the negative sample reconstruction error, and the negative sample reconstruction error is set as the reconstruction error threshold. In actual detection, when the reconstruction error of the input flow velocity fluctuation spectrum is greater than the reconstruction error threshold, the abnormal detection classification result indicating the presence of biological noise pollution is output.

[0018] This invention also discloses a water conservancy monitoring system based on digital twins, used to implement the above-mentioned water conservancy monitoring method based on digital twins, the system comprising: The bottom-level data acquisition module is used to periodically release heat at the bottom of the monitoring section through a composite sensor that integrates a pressure sensing probe and a thermal pulse probe, and record the temperature decay time series curve of the medium around the composite sensor, as well as to acquire the pressure value of the surrounding medium through the pressure sensing probe. The state matching and determination module is used to input the temperature decay time series curve into the heat dissipation feature matching model, and the heat dissipation feature matching model outputs the medium category probability representing the surrounding medium category, which includes pure water mode and silt burial mode. The solid phase burial decoupling module is used to input the pressure increment into the sediment thickness conversion template library when the medium type probability indicates a sediment burial mode and the pressure value has a pressure increment compared with the historical reference pressure, and the sediment thickness conversion template library outputs the corresponding sediment thickness. The digital twin reconstruction module is used to maintain the water surface elevation boundary unchanged in the digital twin and extrapolate the sedimentation thickness upward from the twin riverbed elevation boundary corresponding to the position of the composite sensor in the digital twin.

[0019] The advantage of this invention over existing technologies lies in its use of a composite sensor integrating a pressure sensing probe and a thermal pulse probe, installed at the bottom of the monitoring section. This sensor identifies the surrounding medium by utilizing the differences in heat dissipation between water and sediment. In aquatic environments, heat is easily affected by convection and diffusion, while in sediment-buried environments, heat is primarily conducted slowly through particulate matter and pore water. Therefore, the temperature decay time-series curve can reflect whether the environment around the sensor is pure water or sediment-buried. Furthermore, this invention expands the interpretation of pressure increments from simple water level calculations to a determination of the physical source under the constraints of the medium's state, thus avoiding misinterpreting pressure increases caused by sediment accumulation as water level increases.

[0020] When the area around the sensor is determined to be buried by sediment and the pressure reading shows an increase, this invention converts the pressure increment into sediment thickness and keeps the water surface elevation boundary unchanged in the digital twin, while extrapolating the corresponding riverbed elevation boundary upwards. This allows the bottom topography in the digital twin to dynamically change with the actual sedimentation process, avoiding distortions in the calculation of cross-section, water depth distribution, and flow field caused by the virtual model still using the old riverbed elevation. This processing method utilizes the unique technical characteristic of bottom pressure readings in hydraulic monitoring, which is simultaneously affected by water column pressure and solid burial pressure, achieving decoupling between water surface boundary updates and riverbed boundary updates, and improving the ability of the digital twin hydraulic model to express sudden sedimentation processes.

[0021] Furthermore, this invention extracts the sedimentation thickness output from multiple composite sensors on the same monitoring section and uses a spatial interpolation algorithm to generate a spatial sedimentation thickness distribution matrix, thereby synchronously raising and updating the three-dimensional mesh nodes at the bottom of the digital twin. Compared to methods that only obtain single-point sedimentation information, this processing can reflect the actual situation of uneven sedimentation at different locations within the section, making the virtual riverbed no longer a simple overall translation, but rather forming a bottom boundary that more closely resembles the real section shape based on the on-site sedimentation distribution.

[0022] Furthermore, regarding the issue of wind-induced surface flow velocity, this invention utilizes an anemometer to collect wind shear stress characteristic data and performs frequency domain decomposition on the original surface flow velocity time-series signal. The wind shear stress characteristic data and the high-frequency fluctuating flow velocity sequence are then input into a multimodal cross-attention mechanism model to obtain the surface wind-driven drift disturbance, which is then subtracted from the original surface flow velocity. This method leverages the technical characteristics of open water surface flow velocity being easily driven by wind shear stress, while digital twin hydrodynamic calculations need to reflect the gravity flow component. This reduces the contamination of top dynamic boundary conditions by wind and wave disturbances, making the flow velocity input into the three-dimensional hydrodynamic evolution model closer to the actual hydrodynamic flow velocity.

[0023] Furthermore, to address the issue of biological noise contamination in underwater acoustic velocity data, this invention performs a Fourier transform on the velocity Doppler frequency shift signal to obtain the velocity fluctuation spectrum. A fluid spectrum anomaly detection model is then used to determine the presence of independent high-energy frequency peaks associated with biological rhythms within the spectrum. When biological noise contamination is detected, the corresponding spatial grid is marked as missing, and a three-dimensional spatial kriging interpolation algorithm is invoked to complete the data using flow field data from adjacent uncontaminated spatial grids. This approach leverages the characteristic that natural fluid spectra typically exhibit continuous energy decay, while biological activity reflection signals easily form localized high-energy anomaly peaks. This prevents contaminated velocity data from directly entering the digital twin, improving the reliability of the internal three-dimensional flow field assimilation results. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall architecture of the water conservancy monitoring system based on digital twins according to the present invention; Figure 2 This is a physical deployment scene diagram of the underlying data acquisition module of this invention; Figure 3 This is a schematic diagram of the composite sensor for medium measurement according to the present invention; Figure 4 This is a schematic diagram illustrating the digital twin water surface and riverbed elevation boundary derivation of this invention. Detailed Implementation

[0025] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0026] This invention provides a water conservancy monitoring method based on digital twins. This method is used for monitoring sections of reservoirs, rivers, canals, sluice gate forebays, and pumping station inlet and outlet channels that are at risk of bottom siltation. The overall system can be composed of a field perception layer, an edge computing layer, a digital twin model layer, and a scheduling and display layer. The field perception layer collects data on bottom pressure, thermal attenuation, water surface velocity, wind speed and direction, and underwater acoustic velocity. The edge computing layer performs medium identification, siltation thickness calculation, wind-driven disturbance deduction, and abnormal acoustic data removal. The digital twin model layer updates the water surface boundary, riverbed boundary, and internal three-dimensional flow field accordingly. Figure 1 As shown.

[0027] A composite sensor is deployed at the bottom of the monitoring section, integrating a pressure sensing probe and a thermal pulse probe. The composite sensor can be installed inside an anti-scouring sleeve on the riverbed bottom, or on a pre-embedded bracket, a recessed area in a concrete base, or an anti-scouring base. To ensure that pressure and heat can be transferred to the surrounding medium, pressure-conducting windows and thermal contact surfaces can be provided on the anti-scouring sleeve or base. This allows the pressure sensing probe to detect the pressure generated by the surrounding water or sediment, and the thermal pulse probe to release heat to the surrounding medium and detect the heat loss process. Multiple composite sensors can be arranged laterally along the monitoring section, or separately in the main channel, side banks, and near-shore areas to adapt to uneven siltation distribution. Figure 2 As shown.

[0028] The thermal pulse probe can be implemented using a miniature resistance heating element, a ceramic heating core, or a metal film heating element. The temperature sensing element can be a thermistor, platinum resistance thermometer, digital temperature chip, or fiber optic temperature sensing unit. Within a monitoring cycle, the thermal pulse probe first releases heat for a short time at a preset power, then stops heating and continuously records the temperature decrease process of the medium surrounding the composite sensor, forming a temperature decay time-series curve. The monitoring cycle can be set from 1 minute to 60 minutes, the thermal pulse release duration can be set from 2 seconds to 60 seconds, and the temperature sampling frequency can be set from 1 Hz to 20 Hz, depending on the rate of change in the hydraulic engineering project. A shorter cycle is used when the water flow is rapid and the sediment changes quickly, while a longer cycle can be used in reservoir areas or slow-flowing river sections to reduce power consumption and extend equipment life.

[0029] Pressure sensing probes can be implemented using piezoresistive pressure chips, capacitive pressure chips, vibrating wire pressure sensors, or fiber optic pressure sensors. The pressure sensing probe acquires the pressure values ​​of the surrounding medium. These pressure values ​​undergo zero-point drift correction, temperature compensation, and abnormal peak filtering before being used for further analysis. Historical reference pressures are dynamically updated reference values ​​based on the hydrological environment, and are not fixed initial pressure values. Two methods can be used to determine historical reference pressures. In one embodiment, when the heat dissipation feature matching model detects that the medium surrounding the composite sensor has changed from pure water mode to sediment burial mode, the pressure values ​​collected in pure water mode within a preset time window before the change occurs are extracted. The preset time window can be set to 5 minutes to 60 minutes. In some embodiments, to avoid the influence of instantaneous fluctuations, abnormal peak data within this time window can be removed first, and then a segment of pressure data with a pressure change amplitude less than a preset fluctuation threshold can be selected as stable pressure data. The average or median value of this stable pressure data is then used as the historical reference pressure. Since this pressure value is collected in pure water mode before the sensor is covered by sediment, it mainly reflects the water column pressure under the water level conditions at that time and can be used as a reference benchmark for subsequent judgment of sediment burial pressure increments.

[0030] In another embodiment, the historical baseline pressure can be calculated in real time based on the current water surface elevation in the digital twin. Specifically, the current water surface elevation in the digital twin is first obtained, and then the initial absolute elevation of the composite sensor's location is obtained. When the current water surface elevation is higher than the initial absolute elevation of the composite sensor, the elevation difference between the two is the current water depth above the composite sensor. The current water depth is multiplied by the pure water density and the gravitational acceleration to obtain the theoretical pure water pressure that the composite sensor location should bear in the absence of sediment burial, and this theoretical pure water pressure is used as the historical baseline pressure. The pure water density can be taken as 1000 kg / m³, or it can be adjusted according to the on-site water temperature and sediment content; the gravitational acceleration can be taken as 9.8 m / s². If the current water surface elevation is not higher than the initial absolute elevation of the composite sensor, it means that a valid water column has not formed at this location theoretically, and the theoretical pure water pressure can be taken as 0 or the moment can be marked as an invalid calculation moment.

[0031] When actually determining the pressure increment, the current pressure value collected by the pressure sensing probe is compared with the dynamically determined historical reference pressure. The portion of the current pressure value that exceeds the historical reference pressure is considered as the pressure increment mainly caused by sediment burial.

[0032] This invention does not directly interpret an increase in bottom pressure as an increase in water level. Instead, it first determines the surrounding medium of the composite sensor by analyzing heat dissipation behavior. Heat dissipation in water is affected by convection, turbulent diffusion, and water exchange, resulting in a typically rapid temperature drop and potential fluctuations related to flow velocity. In sediment-buried environments, heat is primarily conducted slowly between the particle skeleton and pore water, leading to a more gradual temperature drop and a longer tail of the decay curve. Utilizing this difference in thermal decay, this invention distinguishes the physical source of bottom pressure into changes in water column pressure and changes in solid-phase burial pressure, such as… Figure 3 As shown.

[0033] After the temperature decay time series curve is input into the heat dissipation feature matching model, the model outputs the probability of the medium category. The medium category includes at least the pure water mode and the sediment-buried mode. Here, the pure water mode indicates that the composite sensor is not covered by sediment and mainly exchanges heat and is in direct pressure contact with the water body, without requiring that the water body on site is absolutely free of suspended sediment. The model output can be two probability values, one corresponding to the pure water mode and the other to the sediment-buried mode. In engineering implementation, the category with the higher probability can be selected as the judgment result, or a judgment threshold of 0.5 to 0.9 can be set; for example, when the probability of the sediment-buried mode reaches 0.7 or higher, the sensor is determined to be in a sediment-buried state. If the two probabilities are close, the state of the previous cycle can be maintained, or the current cycle can be marked as a pending confirmation state to avoid frequent switching during the transient process at the water-sand interface.

[0034] When the medium type probability indicates a sediment burial mode, and the pressure value shows an increase compared to the historical baseline pressure, it suggests that the increased bottom pressure is more likely due to the weight of the sediment, pore water pressure, and the sediment's encapsulation effect on the sensor, rather than simply from water level rise. In this case, the pressure increment is input into the sediment thickness conversion template library to obtain the corresponding sediment thickness. The pressure increment can first be judged by a noise band, which can be set to 0.1% to 2% of the sensor's full scale. Only pressure changes exceeding the noise band are included in the sediment thickness conversion, avoiding frequent fluctuations in the digital twin riverbed caused by minor drifts. If the monitoring section is also equipped with independent water level gauges, radar water level gauges, or other composite sensors operating in pure water mode, the pressure value can be corrected using synchronous water level changes before calculating the pressure increment mainly caused by solid burial. This avoids incorrectly including the pressure portion corresponding to water level changes in the sediment thickness when both rising water and sedimentation occur simultaneously.

[0035] The sediment thickness conversion template library can be established through indoor flue calibration, field test calibration, or historical dredging measurement data. During calibration, typical sediment of different thicknesses is laid on top of the composite sensor in stages, and the pressure increment after stabilization is recorded, forming a mapping relationship table with the measured sediment thickness. The mapping relationship table can be divided according to pressure increment, for example, with a interval of 0.1 kPa to 5 kPa, or more dense intervals can be set according to the required accuracy of the project. The sediment thickness value can cover a range of 0.01 meters to 5 meters, with a smaller range for shallow channels and a larger range for large rivers or reservoirs. If the input pressure increment matches the pressure interval in the mapping table, the target thickness value is directly output; if it does not match precisely, two adjacent pressure intervals can be looked up and interpolated to obtain continuous sediment thickness results. This preserves the stability of the lookup table while avoiding abrupt changes in the thickness result.

[0036] After obtaining the sedimentation thickness, the digital twin maintains the current water surface elevation boundary unchanged due to the pressure increment, while simultaneously extrapolating the same thickness upwards from the twin riverbed elevation boundary corresponding to the installation location of the composite sensor. The unchanged water surface elevation boundary here means that the pressure increment generated by the bottom composite sensor is no longer converted into water surface rise. If the digital twin is simultaneously connected to independent water level observation data, the water surface elevation can still be updated based on the independent water level observation data. Extrapolation upwards can be understood as raising the virtual riverbed at that location from its original elevation, causing the effective water depth and cross-section in the digital twin to decrease synchronously with the actual sedimentation, rather than mistakenly converting the pressure increment into water surface rise. The digital twin can be implemented using a 3D mesh, a 2D cross-sectional mesh, or 1D river network cross-section nodes. In a 3D mesh, the updated object is usually the bottom boundary node; in a 2D cross-sectional mesh, the updated object is usually the riverbed line node; in a 1D river network model, the updated object can be the bottom elevation of the cross-section and the cross-sectional morphology parameters. After the update, the model recalculates water depth, wetted perimeter, cross-sectional area, roughness influence zone, and connectivity between adjacent elements to ensure that subsequent hydrodynamic calculations use the new bottom boundary. Figure 4 As shown.

[0037] The heat dissipation feature matching model can be implemented using a time series classification model. In one embodiment, the model consists of a one-dimensional convolutional layer, a temporal feature extraction layer, and a binary classification output layer. The one-dimensional convolutional layer extracts the local decay slope, tail cooling features, and short-term perturbation features of the temperature curve; the temporal feature extraction layer can use a long short-term memory network, a gated recurrent unit, a temporal convolutional network, or a lightweight Transformer structure; the output layer uses binary probability output, enabling the model to provide probabilities for pure water mode and sediment burial mode. The input sequence length can be set to 30 to 2000 sampling points, the hidden feature dimension can be set to 32 to 256, and the number of model layers can be set to 1 to 6.

[0038] During training, the first temperature decay time-series curve under historical pure water conditions and the second temperature decay time-series curve under historical silt-covered conditions are acquired. Pure water environment samples can come from the initial sensor installation phase, the water tank calibration phase, or the post-dredging verification phase; silt-covered environment samples can come from artificial sand covering experiments, the on-site siltation verification phase, or the pre- and post-dredging comparison phase. The first temperature decay time-series curve is combined with the first medium category label to form the first training sample, and the second temperature decay time-series curve is combined with the second medium category label to form the second training sample. Before training, the curves can be time-aligned, initial temperature normalized, environmental water temperature compensated, and smoothed using filtering to ensure the model focuses on learning the thermal decay pattern rather than the absolute temperature differences caused by different weather or seasons.

[0039] The initial time series classification model employs supervised training. The loss function can be cross-entropy loss, the optimizer can be Adam, RMSProp, or stochastic gradient descent, the learning rate can be set to 0.0001 to 0.001, the batch size can be set to 16 to 128, and the number of training epochs can be set to 50 to 300. The convergence condition can be that the validation set loss no longer decreases within 5 to 20 consecutive epochs, or the validation set classification accuracy reaches a set requirement. After training, the model is deployed on edge computing devices or monitoring center servers, and the medium category probability can be obtained by inputting the on-site temperature decay time series curve into the model.

[0040] When multiple composite sensors are deployed at adjacent positions at the bottom of the same monitoring section, each composite sensor performs medium identification and sediment thickness calculation, obtaining the sediment thickness at multiple spatial points. After obtaining these sediment thicknesses, the edge computing layer uses a spatial interpolation algorithm to perform surface fitting, generating a spatial sediment thickness distribution matrix covering the monitoring section area. Spatial interpolation algorithms can employ inverse distance weighted interpolation, radial basis function interpolation, ordinary kriging interpolation, or triangular mesh interpolation. Inverse distance weighted interpolation can be used when the number of sensors is small, as it has low computational cost and is convenient for real-time operation at the edge; kriging interpolation can be used when the number of sensors is large and spatial correlation is desired.

[0041] The grid resolution of the spatial sedimentation thickness distribution matrix can be consistent with the 3D grid at the bottom of the digital twin, or a coarser matrix can be generated first and then mapped to the 3D grid nodes. For each bottom grid node, the system reads the corresponding thickness increment from the spatial sedimentation thickness distribution matrix based on its planar position and simultaneously raises the elevation value of that node. In this way, the bottom of the digital twin no longer shifts as a whole, but can form a morphology closer to the real cross-section, such as mid-channel sedimentation, side-shoal sedimentation, and locally scoured but not sedimentated areas.

[0042] When the medium type probability indication is in pure water mode, the pressure values ​​collected by the pressure sensing probe mainly reflect the water column pressure. In this case, instead of riverbed uplift, the pressure values ​​are converted into hydrostatic head height. This conversion process can be completed based on hydrostatic pressure conversion relationships, and corrected by incorporating on-site water density, sensor installation elevation, atmospheric pressure compensation, and temperature compensation. The hydrostatic head height is then superimposed on the twin riverbed elevation benchmark value at the sensor location to update the water surface elevation boundary within the digital twin. Through this processing, the same bottom composite sensor serves to update the water surface boundary in pure water mode and the riverbed boundary in sediment-buried mode, thus decoupling changes in water surface elevation and changes in riverbed siltation.

[0043] In a further embodiment, the present invention also identifies and subtracts wind-driven drift disturbances in the water surface velocity. Wind shear stress characteristic data is collected on-site using an anemometer, which can consist of wind speed, wind direction, the angle between the wind direction and the mainstream direction, gust intensity, and short-term wind speed change rate. A Doppler surface current radar collects the original water surface velocity time-series signal, with a sampling frequency set to 1 Hz to 20 Hz and a time window set to 30 seconds to 300 seconds. Since the water surface velocity includes both gravity flow components driven by water level difference, gravitational potential energy, and riverbed slope, and may also include drift components formed by wind blowing across the water surface, frequency domain signal decomposition is required first.

[0044] Frequency domain signal decomposition can employ Fast Fourier Transform (FFT), Short-Time Fourier Transform (SFT), or wavelet decomposition. After decomposition, the low-frequency background velocity sequence is used to describe relatively stable mainstream changes, while the high-frequency fluctuating velocity sequence is used to describe wind waves, surface disturbances, and short-period fluctuations. The frequency boundary can be set between 0.02 Hz and 0.5 Hz based on the wave period and cross-sectional flow pattern. For open reservoirs and scenarios with significant wind and waves, a lower boundary frequency can be used to retain more wind and wave disturbances; for narrow channels or sluice gate forebays, the boundary frequency can be appropriately adjusted based on the low-frequency fluctuations caused by pump and gate start-up and shutdown.

[0045] Wind shear stress feature data and high-frequency fluctuating velocity sequences are jointly input into a multimodal cross-attention mechanism model. This model can consist of a wind field feature encoder, a velocity fluctuation feature encoder, a cross-attention layer, and a perturbation sequence output layer. The wind field feature encoder can employ a multilayer perceptron or a one-dimensional convolutional network to encode wind speed and direction features into wind-driven feature vectors. The velocity fluctuation feature encoder can employ a one-dimensional convolutional network, a gated recurrent unit, or a Transformer encoding layer to encode the high-frequency fluctuating velocity sequence into velocity fluctuation features. The cross-attention layer aligns the wind field features with the velocity fluctuation features, learning which velocity fluctuation segment is more likely to be caused by wind shear stress. The number of attention heads can be set to 2 to 8, the hidden dimension can be set to 64 to 256, and the number of cross-attention layers can be set to 1 to 4. The model outputs a surface wind-driven drifting perturbation velocity sequence, with the output length consistent with the original water surface velocity time series signal.

[0046] After obtaining the surface wind-driven drift disturbance velocity sequence, this disturbance sequence is subtracted from the original surface velocity time series signal to obtain the gravity velocity sequence. Before subtraction, the timestamps of the two sequences must be aligned. If the sampling frequencies of the anemometer and surface velocity radar are different, linear interpolation, spline interpolation, or resampling can be used to unify the time step. The gravity velocity sequence serves as the input to the three-dimensional hydrodynamic evolution model of the digital twin, representing the top dynamic boundary conditions. The three-dimensional hydrodynamic evolution model can be solved using the finite volume method, finite element method, or finite difference method. The model can include variables such as water level, velocity, pressure, turbulent viscosity, and boundary friction. By subtracting the wind-driven drift disturbance, the top boundary more closely approximates the actual hydrodynamic velocity, rather than being directly contaminated by short-term wind and waves.

[0047] The training process of the multimodal cross-attention mechanism model can be completed based on historical data. First, historical wind shear stress characteristic data and synchronously occurring historical surface velocity time-series signals are collected. Frequency domain decomposition is then performed on the historical surface velocity time-series signals to obtain historical high-frequency fluctuating velocity sequences. Next, an acoustic Doppler current profiler is used to obtain the actual background velocity sequence of the stable flow layer below the water surface. The stable flow layer can be selected within a depth range of 0.3 to 2 meters above the water surface, unaffected by direct bottom friction; the specific depth is determined based on water depth and flow regime. The difference sequence between the historical surface velocity time-series signal and the actual background velocity sequence serves as a label for the true surface wind-driven drift disturbance. If the historical surface velocity and the actual background velocity have different sampling depths or frequencies, time synchronization and depth representativeness correction can be performed first before obtaining the difference sequence, avoiding measurement bias introduced by the label itself.

[0048] The training samples use historical wind shear stress characteristics and historical high-frequency fluctuating velocity sequences as input features, and the actual surface wind-driven drift disturbance labels as supervision labels. The training loss can be mean squared error loss, mean absolute error loss, or Huber loss. The optimizer can be Adam, with a learning rate of 0.00005 to 0.001, a batch size of 8 to 64, and 50 to 500 training epochs. After training, the model can estimate wind-driven disturbances based on the on-site wind field conditions and high-frequency water surface fluctuation characteristics.

[0049] In a further embodiment, the present invention also identifies and corrects for biological noise pollution in underwater acoustic current velocity data. An underwater acoustic current meter is used to collect current Doppler frequency shift signals at each depth layer. The acoustic current meter can be a point-type acoustic Doppler current meter or an acoustic Doppler current profiler. Each depth layer can be divided into intervals of 0.1 meters to 1 meter, with smaller intervals used in shallower water and larger intervals used in deep reservoirs or deep channel river sections.

[0050] Perform a Fourier transform on the flow velocity Doppler frequency shift signal to obtain the corresponding flow velocity fluctuation spectrum. The spectrum of natural fluid dynamics processes typically exhibits a continuous energy decay pattern, with higher energy at low frequencies that gradually decreases as the frequency increases. Underwater fish, plankton colonies, bubble buoyancy, or the activity of local reflectors can easily form independent high-energy peaks in certain frequency bands. These peaks do not represent actual water flow velocity changes, and if directly entered into a digital twin, they will cause local anomalies in the internal three-dimensional flow field.

[0051] After the velocity fluctuation spectrum is input into the fluid spectral anomaly detection model, the model outputs anomaly detection classification results to indicate the presence of biological noise pollution. When the anomaly detection classification results indicate the presence of biological noise pollution, the spatial grid corresponding to the depth layer where the acoustic current meter is located in the digital twin is marked as missing. A missing state does not mean there is no water flow at that location, but rather that the acoustic observation data for that grid at the current moment is unreliable and needs to be completed by neighboring data. Subsequently, a three-dimensional spatial kriging interpolation algorithm is called to complete the missing spatial grid using flow field data from uncontaminated neighboring spatial grids in the digital twin. The neighboring grids involved in the interpolation can come from adjacent lateral positions within the same depth layer, or from adjacent positions in the upper / lower depth layers and the forward / backward flow directions. The completed flow field data is then assimilated into the internal three-dimensional flow field of the digital twin, ensuring that the three-dimensional flow field remains continuous, smooth, and consistent with the surrounding hydrodynamic conditions.

[0052] A fluid spectral anomaly detection model can be constructed based on a one-dimensional convolutional autoencoder structure. The input is a normalized sequence of flow velocity fluctuation spectra. The encoder consists of multiple one-dimensional convolutional layers, activation layers, and downsampling layers, used to extract the continuous decay pattern of the natural fluid spectrum. The decoder consists of upsampling layers, deconvolutional layers, or one-dimensional convolutional reconstruction layers, used to reconstruct the input spectrum. The encoder kernel size can be set to 3 to 9, the number of channels can be set to 16 to 128, and the latent feature dimension can be set to 8 to 128. The model output is the reconstructed spectrum sequence, and the reconstruction error is used to measure whether the input spectrum conforms to the spectral morphology of natural fluids.

[0053] During training, the first historical flow velocity fluctuation spectrum, conforming to the energy decay pattern of natural fluid dynamics, is acquired as a positive sample. Positive samples can come from periods with no significant biological activity, no bubble interference, and normal conditions confirmed by on-site manual verification. Then, the second historical flow velocity fluctuation spectrum, exhibiting high-energy frequency peaks with independent biological rhythms, is acquired as a negative sample. Negative samples can come from periods with significant nighttime fish activity, abnormally enhanced acoustic echoes, and biological disturbances confirmed by on-site video or manual inspection. The training phase primarily utilizes positive samples to perform feature reconstruction training on a one-dimensional convolutional autoencoder, enabling the model to effectively reconstruct normal fluid spectra. Reconstruction errors can be calculated using mean squared error, mean absolute error, or frequency-weighted error. The learning rate can be set to 0.0001 to 0.001, the batch size to 16 to 128, and the number of training epochs to 50 to 300.

[0054] After training to reconstruct positive samples, negative samples are input into the trained one-dimensional convolutional autoencoder to calculate the reconstruction error. Because negative samples contain independent high-energy peaks, the model struggles to accurately reconstruct these anomalous peaks, resulting in a higher reconstruction error for negative samples compared to positive samples. The minimum, low quantile, or empirically validated value from the negative sample reconstruction error set can be selected as the reconstruction error threshold. When the spectrum is normalized to the 0-1 range, the reconstruction error threshold typically falls between 0.01 and 0.5, with the specific value adjusted based on the spectrum length, sampling noise, and model structure. In actual detection, when the reconstruction error of the input flow velocity fluctuation spectrum exceeds the reconstruction error threshold, the model outputs an anomaly detection classification result indicating biological noise contamination; otherwise, it outputs an uncontaminated or reliable result.

[0055] This invention also provides a water conservancy monitoring system based on digital twins for executing the aforementioned methods. The underlying data acquisition module of the system is responsible for the heating control, temperature sampling, pressure sampling, and raw data preprocessing of the composite sensor. The state matching and determination module is responsible for calling the heat dissipation feature matching model and outputting the probability of the pure water mode and the probability of the sediment burial mode based on the temperature decay time series curve. The solid phase burial decoupling module is responsible for calculating the pressure increment in the sediment burial mode and calling the sediment thickness conversion template library to output the sediment thickness. The digital twin reconstruction module is responsible for maintaining the water surface elevation boundary unchanged due to the pressure increment in the sediment burial mode and raising the riverbed elevation boundary; in the pure water mode, the digital twin reconstruction module updates the water surface elevation boundary based on the water pressure and head height. The system can also be configured with a wind-driven disturbance correction submodule and an acoustic anomaly completion submodule. The wind-driven disturbance correction submodule completes the acquisition of wind shear stress characteristic data, decomposition of the original water surface velocity time series signal, calculation of surface wind-driven drift disturbance velocity sequence, and output of gravity velocity sequence. The acoustic anomaly completion submodule completes the spectral analysis of velocity Doppler frequency shift signal, identification of biological noise pollution, marking of missing pollution grids, and three-dimensional spatial kriging interpolation completion.

[0056] Through the above implementation methods, bottom pressure data is no longer interpreted solely as water level changes, but is instead allocated to either surface boundary updates or riverbed boundary updates under the constraints of heat dissipation medium identification results. For surface wind-driven disturbances and underwater biological noise, this invention further completes identification, subtraction, and completion before entering the digital twin, enabling the digital twin's surface elevation, riverbed elevation, top dynamic boundary, and internal three-dimensional flow field to more closely approximate the actual state of a hydraulic engineering project.

[0057] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A water conservancy monitoring method based on digital twins, characterized in that, include: A composite sensor integrating a pressure sensing probe and a thermal pulse probe is deployed at the bottom of the monitoring section. The thermal pulse probe periodically releases heat and records the temperature decay time-series curve of the medium surrounding the composite sensor, and the pressure value of the surrounding medium is acquired by the pressure sensing probe. The temperature decay time series curve is input into the heat loss feature matching model, and the heat loss feature matching model outputs the medium category probability representing the surrounding medium category, which includes pure water mode and silt burial mode. When the medium category probability indicates a sediment burial mode, and the pressure value has an increase compared to the historical reference pressure, the pressure increase is input into the sediment thickness conversion template library, and the sediment thickness conversion template library outputs the corresponding sediment thickness. The water surface elevation boundary is kept constant in the digital twin, and the sedimentation thickness is extrapolated upward from the twin riverbed elevation boundary corresponding to the location of the composite sensor in the digital twin.

2. The water conservancy monitoring method based on digital twin according to claim 1, characterized in that, The training process of the heat dissipation feature matching model includes: Obtain the first temperature decay time series curve under a pure water environment and the second temperature decay time series curve under a sediment burial environment. The first temperature decay time series curve is combined with the first medium category label to form the first training sample, and the second temperature decay time series curve is combined with the second medium category label to form the second training sample. The initial time series classification model is trained in a supervised manner using the first training sample and the second training sample, and the parameters are iteratively updated until convergence, thus obtaining the trained heat dissipation feature matching model.

3. The water conservancy monitoring method based on digital twin according to claim 1, characterized in that, The step of determining the historical reference pressure includes extracting the pressure value collected in pure water mode within a preset time window before the change occurs as the historical reference pressure when the composite sensor detects that the surrounding medium has changed from pure water mode to sediment burial mode. The step of inputting the pressure increment into the sediment thickness conversion template library and having the sediment thickness conversion template library output the corresponding sediment thickness includes: A mapping table between pressure increment and corresponding sediment thickness value is pre-established in the sediment thickness conversion template library; The target thickness value that matches the input pressure increment is queried in the mapping table, and the found target thickness value is determined as the final output sediment thickness.

4. The water conservancy monitoring method based on digital twin according to claim 1, characterized in that, When multiple composite sensors are deployed at adjacent positions at the bottom of the monitoring section, after the step of extrapolating the sedimentation thickness upwards from the twin riverbed elevation boundary corresponding to the composite sensor positions in the digital twin, the method further includes: Obtain the deposition thickness output by each of the multiple composite sensors; A spatial interpolation algorithm is used to fit the surface of the sedimentation thickness output by each of the multiple composite sensors to generate a spatial sedimentation thickness distribution matrix covering the monitoring section area. The elevation values ​​of the three-dimensional mesh nodes at the bottom of the digital twin are synchronously raised and updated using the spatial accumulation thickness distribution matrix.

5. The water conservancy monitoring method based on digital twin according to claim 1, characterized in that, The method further includes: When the medium type probability indicator is pure water mode, the pressure value collected by the pressure sensing probe is converted into water pressure head height; The water pressure head height is superimposed on the reference elevation of the twin riverbed elevation boundary to update the water surface elevation boundary inside the digital twin.

6. The water conservancy monitoring method based on digital twin according to claim 1, characterized in that, The method further includes: Wind shear stress characteristic data were collected using an anemometer and Doppler surface current radar was used to collect the original water surface flow velocity time series signal. Frequency domain signal decomposition is performed on the original water surface flow velocity time series signal to separate the low-frequency background flow velocity sequence and the high-frequency fluctuating flow velocity sequence; The wind shear stress characteristic data and the high-frequency fluctuating velocity sequence are input together into the multimodal cross-attention mechanism model, and the multimodal cross-attention mechanism model outputs the surface wind-driven drift disturbance velocity sequence. The gravity velocity sequence is obtained by subtracting the surface wind-driven drift disturbance velocity sequence from the original water surface velocity time sequence signal. The gravity velocity sequence is used as the top dynamic boundary condition and input into the three-dimensional hydrodynamic evolution model of the digital twin for flow field calculation.

7. The water conservancy monitoring method based on digital twin according to claim 6, characterized in that, The training process of the multimodal cross-attention mechanism model includes: Historical wind shear stress characteristic data and synchronously occurring historical water surface original flow velocity time series signals are collected, and frequency domain signal decomposition is performed on the historical water surface original flow velocity time series signals to obtain historical high-frequency fluctuation flow velocity sequences. The actual background velocity sequence of the stable flow layer below the water surface is obtained by using an acoustic Doppler current profiler, and the difference sequence between the historical original surface velocity time series signal and the actual background velocity sequence is set as the true surface wind-driven drift disturbance amount label. Using the historical wind shear stress characteristic data and the historical high-frequency fluctuating velocity sequence as input features, and the real surface wind-driven drift disturbance amount label as supervision label, the initial multimodal cross-attention mechanism model is trained in a supervised manner to obtain the trained multimodal cross-attention mechanism model.

8. The water conservancy monitoring method based on digital twin according to claim 1, characterized in that, The method further includes: Acoustic current meters deployed underwater were used to collect Doppler frequency shift signals of current velocity at various depths. Perform a Fourier transform on the flow velocity Doppler frequency shift signal to obtain the corresponding flow velocity fluctuation spectrum; The flow velocity fluctuation spectrum is input into the fluid spectrum morphology anomaly detection model, and the fluid spectrum morphology anomaly detection model outputs anomaly detection classification results to indicate whether biological noise pollution exists. When the anomaly detection classification result indicates the presence of biological noise pollution, the spatial grid corresponding to the depth layer where the acoustic flowmeter is located in the digital twin is marked as missing. The three-dimensional spatial kriging interpolation algorithm is invoked to complete the missing spatial grid using the flow field data of the uncontaminated adjacent spatial grids in the digital twin, and the completed flow field data is assimilated into the internal three-dimensional flow field of the digital twin.

9. The water conservancy monitoring method based on digital twin according to claim 8, characterized in that, The fluid spectral morphology anomaly detection model is constructed based on a one-dimensional convolutional autoencoder structure. The training process of the fluid spectral morphology anomaly detection model includes: The first historical velocity fluctuation spectrum that conforms to the energy decay pattern of natural fluid dynamics is obtained as a positive sample. The second historical velocity fluctuation spectrum with independent biological rhythm high-energy frequency peaks was obtained as a negative sample. The positive samples are used to perform feature reconstruction training on the one-dimensional convolutional autoencoder structure to minimize the reconstruction error; The negative sample is input into a one-dimensional convolutional autoencoder structure trained by the feature reconstruction to calculate the negative sample reconstruction error, and the negative sample reconstruction error is set as the reconstruction error threshold. In actual detection, when the reconstruction error of the input flow velocity fluctuation spectrum is greater than the reconstruction error threshold, the abnormal detection classification result indicating the presence of biological noise pollution is output.

10. A water conservancy monitoring system based on digital twins, characterized in that, For implementing the water conservancy monitoring method based on digital twins as described in any one of claims 1 to 9, the system comprises: The bottom-level data acquisition module is used to periodically release heat at the bottom of the monitoring section through a composite sensor that integrates a pressure sensing probe and a thermal pulse probe, and record the temperature decay time series curve of the medium around the composite sensor, as well as to acquire the pressure value of the surrounding medium through the pressure sensing probe. The state matching and determination module is used to input the temperature decay time series curve into the heat dissipation feature matching model, and the heat dissipation feature matching model outputs the medium category probability representing the surrounding medium category, which includes pure water mode and silt burial mode. The solid phase burial decoupling module is used to input the pressure increment into the sediment thickness conversion template library when the medium type probability indicates a sediment burial mode and the pressure value has a pressure increment compared with the historical reference pressure, and the sediment thickness conversion template library outputs the corresponding sediment thickness. The digital twin reconstruction module is used to maintain the water surface elevation boundary unchanged in the digital twin and extrapolate the sedimentation thickness upward from the twin riverbed elevation boundary corresponding to the position of the composite sensor in the digital twin.

Images