Water conservancy project structure leakage monitoring and positioning system based on multi-sensor cooperation

Abstract

The application discloses a water conservancy project structure leakage monitoring and positioning system based on multi-sensor cooperation and relates to the technical field of water conservancy project monitoring. A heterogeneous graph structure is constructed by fusing real sensor nodes and virtual physical field nodes; a space-time attention and physical constraint graph convolution network is deployed, a time asynchronous alignment module is used to adaptively compensate for the response delay of the heterogeneous sensor to the leakage event, a physical constraint graph attention layer is used to modulate the information transmission between nodes by using a dynamic physical coupling coefficient, a continuous leakage probability distribution field with sparse constraints is generated, and a joint loss function containing a data fitting term and a physical residual term is constructed to force the model to predict a seepage mechanics partial differential equation. The application can realize sub-sensor precision leakage positioning, has high robustness and interpretability in a data sparse and strong noise environment, and can effectively reduce the false positive rate and the whole life cycle cost of water conservancy project leakage monitoring.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering monitoring technology, and in particular to a water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration. Background Technology

[0002] As a crucial component of infrastructure, the safe operation of water conservancy projects directly impacts water resource security, energy supply, and the safety of life and property downstream. Leakage is one of the most common and dangerous forms of structural damage in water conservancy projects. It is highly insidious and sudden, often developing gradually from within and potentially leading to major safety accidents such as piping, landslides, or even dam failure. Therefore, accurate monitoring and location of leakage in water conservancy projects is of paramount practical importance for ensuring the long-term safe operation of these projects.

[0003] Currently, various technical methods have been developed in the field of leakage monitoring and detection in water conservancy projects, mainly including the following aspects: (a) Single sensor monitoring technology Traditional leakage monitoring primarily relies on single-type sensors for fixed-point measurements. For example, piezometers are deployed within the dam to monitor pore water pressure changes, distributed optical fibers are used to sense temperature field anomalies to infer seepage paths, or accelerometers are employed to capture micro-vibration signals induced by seepage. Furthermore, acoustic emission sensors, due to their high sensitivity to weak leakage signals, are widely used for early leakage detection in pipelines and valves. In non-contact detection, geophysical methods such as direct current resistivity methods, ground-penetrating radar, and resistivity CT are applied to detect seepage channels in reservoirs and dams due to their non-destructive testing advantages. In recent years, technologies such as UAV-borne thermal infrared imaging, lidar, and underwater robots equipped with sonar and high-definition cameras have also been increasingly applied to surface and underwater inspections in hydraulic engineering projects.

[0004] However, all of the aforementioned single-sensor technologies have inherent limitations. Pressure sensors are insensitive to slow-developing, minute leaks and struggle to capture early, weak signals; vibration sensors, while highly sensitive, are extremely susceptible to environmental disturbances (such as construction, rainfall, and traffic), leading to a persistently high false alarm rate; acoustic emission technology generates highly non-stationary signals in complex environments, with severe background noise interference, making it difficult for traditional time-frequency analysis methods to effectively separate noise from features. Geophysical methods, while capable of detecting internal structures, are significantly affected by factors such as water depth, silt layer thickness, and environmental resistivity, and single geophysical methods suffer from multiple solutions. Underwater detection equipment faces challenges in terms of visual limitations and positioning accuracy in turbid water environments. These studies demonstrate that monitoring methods based on single sensors are limited by a single signal source and cannot comprehensively and accurately describe the operational status of hydraulic engineering structures.

[0005] (II) Multi-sensor fusion and data-driven methods To overcome the limitations of single sensors, researchers have begun exploring multi-sensor fusion techniques. In the field of water supply networks, some studies have proposed a parallel multi-layer sensor fusion method combining hydrophones, acoustic emission sensors, and vibration sensors, achieving high leakage detection accuracy using convolutional neural networks and few-shot learning. Other studies have fused vibration and pressure signals into a color-coded image using an improved Gram angle field method and constructed a dual-channel multi-scale feature fusion network for leakage level identification. In the area of ​​locating internal leaks in marine pipeline valves, researchers have proposed an intelligent graph construction method based on graph neural networks, converting acoustic emission signals into a graph structure for end-to-end localization, achieving good robustness in high-noise environments.

[0006] However, existing data-driven methods still have the following shortcomings: First, there is the problem of time asynchrony. Different sensors have significantly different response speeds to leakage events (e.g., acoustic emission response is the fastest, followed by vibration, and then pressure osmosis is the slowest). Existing fusion methods often assume that the data are aligned in time, failing to effectively compensate for this physical response delay. Second, there is insufficient spatial correlation modeling. Traditional methods either directly stitch together multi-source data and input it into a neural network or rely solely on Euclidean spatial distance for information fusion, failing to fully utilize the inherent physical field continuity and conservation laws of the leakage process, resulting in weak generalization ability of the model in sparse data regions. Third, there is a lack of physical interpretability. Although pure data-driven black-box models can fit the observed data, their outputs are difficult to establish a direct correlation with the basic laws of seepage mechanics (such as Darcy's law and mass conservation), making it difficult for engineers to verify and trust the model's conclusions. Fourth, there is the bottleneck of sample imbalance and positioning accuracy. Leakage is spatially sparse, and traditional point-by-point loss functions tend to cause the model to be biased towards predicting "no leakage." Moreover, the positioning accuracy is limited by the sensor deployment density, making it difficult to achieve sub-sensor precision leakage boundary identification.

[0007] In summary, although significant progress has been made in the technology for monitoring leakage in hydraulic engineering structures, the following technical challenges still urgently need to be addressed: The problem of asynchronous response of heterogeneous sensors: How to mathematically model and compensate for the time response differences of different types of sensors to the same leakage event, so as to achieve accurate alignment of multi-source evidence on the time axis.

[0008] The problem of physical field continuity and information fusion: How to explicitly introduce physical laws (such as seepage continuity and mass conservation) into the data fusion process so that the model can still make inferences that conform to physical logic in the sparse sensor region, and use physical field discontinuity (parameter mutation caused by leakage) to assist in localization.

[0009] Subsensor precision positioning problem: How to reconstruct discrete sensor observations into a continuous physical field distribution, achieve fine characterization of leakage boundaries, and overcome the limitation of sensor deployment density on positioning accuracy.

[0010] To address the problems existing in the above-mentioned technologies, a water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration is proposed. Summary of the Invention

[0011] The main objective of this invention is to provide a system for monitoring and locating leakage in hydraulic engineering structures based on multi-sensor collaboration, which can effectively solve the problems in the background art.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A multi-sensor collaborative system for monitoring and locating leakage in hydraulic engineering structures includes: A multi-sensor perception layer is deployed in key parts of hydraulic engineering structures, including at least two types of heterogeneous sensors, for real-time acquisition of multi-source time-series monitoring data reflecting the structural status. The edge computing and data preprocessing layer is connected to the multi-sensor perception layer and is used to clean, normalize, and spatiotemporally align the multi-source time-series monitoring data, and extract initial features. The core layer of the digital twin incorporates a graph neural network model that couples physical information. This graph neural network model takes pre-processed multi-source time-series monitoring data as input and constructs a heterogeneous graph structure that integrates sensor nodes and virtual physical field nodes. In the time dimension, it asynchronously aligns the response characteristics of heterogeneous sensors using an attention mechanism and constrains the information transmission between nodes based on physical constitutive equations in the spatial dimension. The probability, precise location, and confidence interval of leakage are generated and output through a Poisson reconstruction layer. In addition, a visualization and early warning layer, connected to the core layer of the digital twin, is used to map the probability, precise location and confidence interval of the leakage to the three-dimensional digital model, and generate a graded early warning signal when the preset threshold is exceeded. The heterogeneous graph structure is defined as follows: , where the set of nodes Includes sensor nodes and virtual physics field nodes; edge set This includes sensor-sensor edges, sensor-physical field edges, and physical field-physical field edges.

[0013] Furthermore, in the heterogeneous graph structure, The sensor node corresponds to the location and real-time data stream of the physical sensor; The virtual physics field nodes represent intermediate physical state variables that cannot be directly measured but follow the laws of physical conservation. The sensor-sensor edge is constructed based on spatial distance and sensor type to capture signal propagation between homogeneous sensors; The sensor-physical field boundary establishes a mapping relationship based on the physical constitutive equation; The physical field-physical field boundary is topologically connected based on the partial differential equations of mass conservation and momentum conservation.

[0014] Furthermore, the graph neural network model includes a physical sensing encoder for embedding and encoding sensor data with different physical dimensions. The physical sensing encoder maps the original monitoring data to a unified feature space by combining a multilayer perceptron with learnable physical prior embedding vectors, so as to distinguish the physical attributes of different sensor data.

[0015] Furthermore, the graph neural network model includes a time-asynchronous alignment module, which employs a deformable temporal convolution kernel and uses an attention mechanism to dynamically learn a time offset for each sensor node at each moment, so as to adaptively align the time delay of different types of sensors in response to the same leakage event and generate node features aligned in the time dimension.

[0016] Furthermore, the graph neural network model includes a physically constrained graph attention propagation layer for updating node features, wherein the physically constrained graph attention propagation layer introduces a dynamic physical coupling coefficient as a weighting factor when calculating the attention coefficient between nodes. The dynamic physical coupling coefficient is determined by the observation residuals of the two nodes in the same physical field space. When the physical field residuals increase, the coupling coefficient decreases to block the invalid information transmission between physical discontinuities caused by leakage.

[0017] Furthermore, the update rule for node features in the physical constraint graph attention propagation layer is defined as follows: ,in, Based on physical coupling coefficient Modulated attention weights; The node features are those that have been asynchronously aligned over time. For information aggregation functions; Let i be the set of neighboring nodes of node i; i and j are the node numbers.

[0018] Furthermore, the graph neural network model includes a spatial leakage field reconstruction module. This module utilizes the final hidden states of all sensor nodes and solves a regularized inverse problem with sparse constraints to reconstruct discrete sensor signals into a leakage probability distribution field in continuous space, thereby achieving sub-sensor precision localization of leakage boundaries.

[0019] Furthermore, the training process of the graph neural network model is optimized by a joint loss function, which... Including data fitting loss term and physical residual loss term Defined as: , Weighting coefficients between (0,1); where the data fitting loss term Used to measure the difference between model predictions and actual sensor observations; the physical residual loss term is used to constrain the model's predictions of virtual physical field nodes to satisfy a preset seepage mechanics partial differential equation, so as to drive the model to make inferences based on physical laws in sparse data regions, and is defined as: , This represents the predicted volumetric water content. The predicted seepage flow rate; For water head; Permeability coefficient; For a single virtual physics node; It is the set of all virtual physical field nodes.

[0020] Furthermore, the multi-sensor sensing layer includes at least two of the following sensor types: distributed fiber optic acoustic wave sensor, embedded piezometer, acceleration vibration sensor, thermal infrared imager, hydrophone, and MEMS microelectromechanical sensor; wherein, the heterogeneous sensors work together to capture stress waves induced by leakage, changes in seepage pressure, structural vibration response, and abnormal temperature field signals, respectively.

[0021] A method for monitoring and locating leakage in hydraulic engineering structures based on multi-sensor collaboration includes the following steps: Step S1: Real-time collection of multi-source time-series monitoring data, including vibration, pressure, temperature and acoustic emission, is achieved through a heterogeneous sensor network deployed on the hydraulic engineering structure. Step S2: Construct a heterogeneous graph of the hydraulic engineering structure containing sensor nodes and virtual physics field nodes, where the virtual physics field nodes represent intermediate state variables that satisfy the seepage mechanics equations; Step S3: Input the multi-source temporal monitoring data into a pre-trained spatiotemporal attention and physical constraint graph convolutional network. The network adaptively aligns the response delay of heterogeneous sensors in the temporal dimension through deformable convolution, and dynamically adjusts the graph attention weights between nodes based on physical residuals in the spatial dimension to propagate and fuse multimodal information. Step S4: Generate a continuous spatial leakage probability distribution field with sparse constraints through the Poisson reconstruction module in the network output layer, and determine the specific coordinates and influence range of the leakage point; Step S5: Superimpose the leakage probability distribution field onto the three-dimensional digital model of the water conservancy project, and trigger alarms of different levels when the leakage probability or leakage flow exceeds the preset safety threshold.

[0022] The present invention has the following beneficial effects: Compared with existing technologies, this solution reconstructs discrete sensor signals into a continuous physical field and uses physical laws (partial differential equations) to perform optimal interpolation of the field distribution between two points in accordance with physical logic. This enables the leakage boundary to be located with sub-meter (<0.5 meters) or even centimeter accuracy, even when the sensor spacing is 5 meters, thus achieving leakage location with sub-sensor accuracy.

[0023] Compared to existing technologies, this solution can automatically learn and compensate for the physical response delays between heterogeneous sensors. During fusion localization, all evidence is precisely aligned on the timeline, ensuring temporal consistency of spatial localization and overcoming localization errors caused by multi-sensor response delays.

[0024] Compared with existing technologies, this solution introduces virtual physical field nodes and physical residual loss. Even if a sensor node fails, its neighboring virtual nodes can still maintain an accurate estimate of the physical state of the area based on the law of physical conservation and information transmitted from adjacent normal sensors. This achieves fault-tolerant monitoring, overcomes the impact of sensor failure or data loss, and enables the monitoring results to maintain high performance even in data-sparse and noisy environments.

[0025] Compared with existing technologies, this solution introduces a dynamic physical coupling coefficient, which can prevent the physical field residuals between nodes from being misjudged as leakage events when they do not conform to the physical field discontinuity characteristics caused by leakage, thereby significantly reducing the false alarm rate.

[0026] Compared with existing technologies, the intelligent model constructed in this solution has powerful interpolation and physical inference capabilities by setting virtual nodes. It can deploy a sparser sensor network while ensuring monitoring accuracy, using fewer sensors and powerful algorithms to achieve or even exceed the monitoring effect of dense deployment, significantly reducing the cost of hardware procurement, installation and long-term maintenance. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall structure of the water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration of the present invention; Figure 2 This is a flowchart illustrating the method for monitoring and locating leakage in hydraulic engineering structures based on multi-sensor collaboration according to the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. Example 1:

[0029] See Figures 1-2 The implementation process of the water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration proposed in this invention may specifically include the following stages and steps: Phase 1: System Deployment and Data Acquisition Step 1.1: Construct a multi-sensor sensing network The sensors are deployed at key parts of the target hydraulic engineering structure to collect multi-source time-series monitoring data reflecting the structural status in real time. These sensors include two of the following sensor types: distributed fiber optic acoustic wave sensor, embedded piezometer, acceleration vibration sensor, thermal infrared imager, hydrophone, and MEMS microelectromechanical sensor. The heterogeneous sensors work together to capture stress waves induced by leakage, changes in seepage pressure, structural vibration response, and abnormal temperature field signals.

[0030] Specifically, heterogeneous sensor clusters are deployed at key sections and potential risk areas of target hydraulic engineering structures (such as concrete dams, earth-rock dam core walls, and levees).

[0031] Internal sensors: Piezometers (to monitor pore water pressure), temperature sensors (to monitor seepage thermal anomalies), and MEMS accelerometers (to monitor structural vibration) are pre-embedded during the dam construction process.

[0032] Surface / external sensors: Install distributed fiber optic acoustic sensors (DAS) (to capture micro-vibration signals caused by seepage) and hydrophones (to capture acoustic signals of underwater leakage) on the dam face or in the gallery.

[0033] Spatiotemporal registration: Record the precise three-dimensional spatial coordinates (x, y, z) and sampling frequency of each sensor to establish a digital ledger of sensors.

[0034] Step 1.2: Define virtual physics nodes In a heterogeneous graph structure, the set of nodes Includes sensor nodes and virtual physics field nodes; edge set This includes sensor-sensor edges, sensor-physical field edges, and physical field-physical field edges.

[0035] The location of the physical sensor corresponding to the sensor node and the real-time data stream; Virtual physics nodes represent intermediate physical state variables that cannot be directly measured but follow the laws of physical conservation. Sensor-sensor edges are constructed based on spatial distance and sensor type to capture signal propagation between homogeneous sensors; The sensor-physical field boundary establishes a mapping relationship based on the physical constitutive equation; The physical field-physical field boundary is topologically connected based on the partial differential equations of mass conservation and momentum conservation.

[0036] Specifically: In the digital twin platform, a set of virtual physical field nodes is generated based on the structural mechanics model and the finite element mesh generation concept. .

[0037] It should be noted that nodes are not physical entities, but rather spatial interpolation points located between sensors (e.g., generated according to a grid density of 1 meter × 1 meter × 1 meter).

[0038] Each virtual node It carries the implicit variables to be solved, such as: seepage velocity q, volumetric water content θ, and total head h.

[0039] Phase Two: Data Preprocessing and Heterogeneous Graph Construction Step 2.1: Data Cleaning and Physical Sensing Encoding The graph neural network model includes a physical sensing encoder for embedding and encoding sensor data with different physical dimensions. The physical sensing encoder maps the raw monitoring data to a unified feature space by combining a multilayer perceptron with learnable physical prior embedding vectors, so as to distinguish the physical attributes of different sensor data.

[0040] Specifically: Edge computing units receive raw time-series data. Where N is the number of sensor nodes; T is the length of the time window; and d is the feature dimension.

[0041] Outlier removal: In one possible implementation, median filtering or the 3σ criterion can be used to remove sensor drift or spike noise.

[0042] Physical sensing embedding: for each sensor node The data is then input into a physical sensing encoder. This encoder maps data from different physical dimensions (Pa, °C, m / s²) to a unified high-dimensional feature vector. It also adds a learnable physical prior embedding to represent the sensor type (such as "pressure type" or "vibration type"). .

[0043] Step 2.2: Construct the heterogeneous graph structure Specifically: Based on the nodes defined in steps 1.1 and 1.2, construct the graph topology, defined as follows: , where the set of nodes Including all real sensor nodes and virtual physics nodes edge set This includes sensor-sensor edges, sensor-physical field edges, and physical field-physical field edges.

[0044] edge set include: Spatial proximity edge: Add an edge between sensor nodes whose spatial distance is less than a threshold R; Physical coupling edge: between each sensor node and its neighboring virtual physics nodes Add edges between them (for example, the vibration sensor should be connected to the stress field nodes around it).

[0045] Physical conservation edge: in adjacent virtual physics nodes Add edges between them; these edges implicitly impose constraints on the conservation of mass and momentum.

[0046] Phase 3: Spatiotemporal Feature Fusion and Intelligent Reasoning Step 3.1: Time Asynchronous Alignment The graph neural network model includes a time-asynchronous alignment module. This module employs deformable temporal convolution kernels and uses an attention mechanism to dynamically learn a time offset for each sensor node at each time step. This adaptively aligns the time delays of different types of sensors' responses to the same leakage event, generating time-aligned node features. Specifically: The processed feature sequence is input into the time-asynchronous alignment module, and the following calculations are performed: Dynamic offset calculation: For the current time t, the model calculates the time offset of each sensor node i through an attention mechanism. .

[0047] Physical significance: For example, when leakage occurs, the acoustic emission sensor (high frequency) responds first, followed by the piezometer (low frequency). The model automatically adjusts so that when fusing information at time t, the acoustic emission node focuses on time t itself, while the piezometer node focuses on time t−Δ, thus mathematically aligning the evolution rates of different physical processes.

[0048] This leads to the initial temporal alignment features. .

[0049] Step 3.2: Spatial Propagation of Physical Constraint Graph The graph neural network model includes a physically constrained graph attention propagation layer for updating node features; this layer introduces a dynamic physical coupling coefficient as a weighting factor when calculating the attention coefficients between nodes. The update rule for node features in the physical constraint graph attention propagation layer is defined as follows: ,in, Based on physical coupling coefficient Modulated attention weights; The node features are those that have been asynchronously aligned over time. For information aggregation functions; Let i be the set of neighboring nodes of node i; i and j are the node numbers.

[0050] Specifically: The aligned features are input into the physical constraint graph attention layer for multiple spatial information transfers.

[0051] Calculate the dynamic physical coupling coefficient The calculation formula is: ,in, It is a function that maps sensor observations to the same physical field space (e.g., water head height), ensuring that when two sensor readings are physically discontinuous (e.g., due to the presence of a water barrier or leak causing drastic changes in water head difference), They will become smaller, thereby reducing the weight of information transmission between them, corresponding to the physical anomalies caused by leakage; This is a function for a multilayer perceptron.

[0052] It should be noted that: dynamic physical coupling coefficient The coupling coefficient is determined by the observation residuals of the two nodes in the same physical field space. As the physical field residual increases, the coupling coefficient decreases to block the transmission of invalid information between physically discontinuous regions caused by leakage. If the head difference between nodes i and j is large (physical field discontinuity, usually a characteristic of leakage channels), then... It will get smaller.

[0053] Modulating attention weights: In the graph attention mechanism of this scheme, dynamic physical coupling coefficients are utilized. Modulate the original attention score. Even if two sensors are close in Euclidean space, their information will not merge if the physical field between them is disrupted (there are cracks or seepage barriers), thus avoiding physically unreasonable "crosstalk".

[0054] Node status update: Aggregate neighborhood information and update all nodes (including virtual physics field nodes). The hidden state of ) Through multiple rounds of propagation, the virtual nodes gradually "sense" the data from the surrounding real sensors, forming an understanding of the entire physical field.

[0055] Phase Four: Physical Consistency Verification and Loss Calculation Step 4.1: Apply physical residual loss The training process of a graph neural network model is optimized by a joint loss function. Including data fitting loss term and physical residual loss term Defined as: , The weight coefficients are in the range (0,1); Among them, the data fitting loss term It is used to measure the difference between model predictions and actual sensor observations, and can be constructed using statistical methods; The physical residual loss term is used to constrain the model's predictions of virtual physical field nodes to satisfy a pre-defined partial differential equation of seepage mechanics, thereby driving the model to make inferences based on physical laws in sparse data regions. It is defined as follows: , This represents the predicted volumetric water content. The predicted seepage flow rate; For water head; Permeability coefficient; For a single virtual physics node; The set of all virtual physics nodes Specifically: During the training and inference process, for each virtual physics node Calculate whether it conforms to the laws of physics.

[0056] In one possible implementation, the following verification method can be used: Darcy's Law Verification: Calculate the predicted seepage flow rate at this point. With head gradient The residual between, i.e. .

[0057] Mass conservation check: Calculate the residual between the rate of change of water volume and the flux divergence at that point, i.e. .

[0058] If the predicted permeation flow rate It is divergent (e.g., water volume is not conserved at source-free points), or the flow velocity is not aligned with the head gradient direction, resulting in physical residual loss. This will increase, forcing the model to revise its predictions until they conform to physical laws. This ensures that even in areas without sensors, the model's inferences are consistent with physical principles.

[0059] Phase 5: Result Decoding and Early Warning Step 5.1: Reconstruction and Location of Spatial Leakage Field The graph neural network model includes a spatial leakage field reconstruction module. This module utilizes the final hidden states of all sensor nodes to reconstruct the discrete sensor signals into a leakage probability distribution field in continuous space by solving a regularized inverse problem with sparse constraints, thereby achieving sub-sensor precision localization of leakage boundaries.

[0060] Specifically: Utilize the updated characteristics of all nodes (especially virtual physics nodes) (characteristics), a continuous leakage probability distribution field is generated through the Poisson reconstruction layer.

[0061] By solving the optimization problem: This yields the leakage probability value at any point within the structure (not just the sensor location). Because... The sparse constraint results in a clear and sharp boundary of the leak area in the final output, which can achieve sub-sensor accuracy in positioning (for example, with a sensor spacing of 5 meters, the leak can be located within 0.5 meters).

[0062] Step 5.2: Digital Twin Visualization and Early Warning Specifically: The reconstructed leakage probability distribution field is overlaid on a BIM or 3D GIS model and displayed in the form of a heat map.

[0063] Threshold determination: Set a yellow warning threshold (leakage probability > 60%) and a red warning threshold (leakage probability > 90%).

[0064] Once a red zone appears or the estimated leakage rate exceeds the design standard, the system will automatically trigger an audible and visual alarm and send a precise diagnostic report containing "three-dimensional coordinates + leakage intensity + development trend" to the management personnel.

[0065] 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 illustrative of the principles of 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 this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-sensor collaborative system for monitoring and locating leakage in hydraulic engineering structures, characterized in that, include: A multi-sensor perception layer is deployed in key parts of hydraulic engineering structures, including at least two types of heterogeneous sensors, for real-time acquisition of multi-source time-series monitoring data reflecting the structural status. The edge computing and data preprocessing layer is connected to the multi-sensor perception layer and is used to clean, normalize, and spatiotemporally align the multi-source time-series monitoring data, and extract initial features. The core layer of the digital twin incorporates a graph neural network model that couples physical information. This graph neural network model takes pre-processed multi-source time-series monitoring data as input and constructs a heterogeneous graph structure that integrates sensor nodes and virtual physical field nodes. In the time dimension, it asynchronously aligns the response characteristics of heterogeneous sensors using an attention mechanism and constrains the information transmission between nodes based on physical constitutive equations in the spatial dimension. The probability, precise location, and confidence interval of leakage are generated and output through a Poisson reconstruction layer. In addition, a visualization and early warning layer, connected to the core layer of the digital twin, is used to map the probability, precise location and confidence interval of the leakage to the three-dimensional digital model, and generate a graded early warning signal when the preset threshold is exceeded. The heterogeneous graph structure is defined as follows: Among them, the node set Includes sensor nodes and virtual physics field nodes; edge set This includes sensor-sensor edges, sensor-physical field edges, and physical field-physical field edges.

2. The water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration according to claim 1, characterized in that, In the heterogeneous graph structure The sensor node corresponds to the location and real-time data stream of the physical sensor; The virtual physics field nodes represent intermediate physical state variables that cannot be directly measured but follow the laws of physical conservation. The sensor-sensor edge is constructed based on spatial distance and sensor type to capture signal propagation between homogeneous sensors; The sensor-physical field boundary establishes a mapping relationship based on the physical constitutive equation; The physical field-physical field boundary is topologically connected based on the partial differential equations of mass conservation and momentum conservation.

3. The water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration according to claim 1, characterized in that, The graph neural network model includes a physical sensing encoder for embedding and encoding sensor data with different physical dimensions. The physical sensing encoder maps the original monitoring data to a unified feature space by combining a multilayer perceptron with learnable physical prior embedding vectors, so as to distinguish the physical attributes of different sensor data.

4. The water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration according to claim 1, characterized in that, The graph neural network model includes a time-asynchronous alignment module, which uses deformable temporal convolution kernels and an attention mechanism to dynamically learn a time offset for each sensor node at each moment. This adaptively aligns the time delays of different types of sensors in response to the same leakage event, generating aligned node features in the time dimension.

5. The water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration according to claim 1, characterized in that, The graph neural network model includes a physically constrained graph attention propagation layer for updating node features, wherein the physically constrained graph attention propagation layer introduces a dynamic physical coupling coefficient as a weighting factor when calculating the attention coefficient between nodes. The dynamic physical coupling coefficient is determined by the observation residuals of the two nodes in the same physical field space. When the physical field residuals increase, the coupling coefficient decreases to block the invalid information transmission between physical discontinuities caused by leakage.

6. The water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration according to claim 5, characterized in that, The update rule for node features in the physical constraint graph attention propagation layer is defined as follows: ,in, Based on physical coupling coefficient Modulated attention weights; The node features are those that have been asynchronously aligned over time. For information aggregation functions; Let i be the set of neighboring nodes of node i; i and j are the node numbers.

7. The water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration according to claim 1, characterized in that, The graph neural network model includes a spatial leakage field reconstruction module. The spatial leakage field reconstruction module utilizes the final hidden states of all sensor nodes and solves a regularized inverse problem with sparse constraints to reconstruct discrete sensor signals into a leakage probability distribution field in continuous space, thereby achieving sub-sensor precision positioning of leakage boundaries.

8. The water conservancy engineering structure leakage monitoring and location system based on multi-sensor collaboration according to claim 1, characterized in that, The training process of the graph neural network model is optimized by a joint loss function, which... Including data fitting loss term and physical residual loss term Defined as: , The weighting coefficients are between (0,1); where the data fitting loss term is... Used to measure the difference between model predictions and actual sensor observations; the physical residual loss term is used to constrain the model's predictions of virtual physical field nodes to satisfy a preset seepage mechanics partial differential equation, so as to drive the model to make inferences based on physical laws in sparse data regions, and is defined as: , This represents the predicted volumetric water content. The predicted seepage flow rate; For water head; Permeability coefficient; For a single virtual physics node; It is the set of all virtual physical field nodes.

9. The system for monitoring and locating leakage in hydraulic engineering structures based on multi-sensor collaboration according to any one of claims 1-8, characterized in that, The multi-sensor sensing layer includes at least two of the following sensor types: distributed fiber optic acoustic wave sensor, embedded piezometer, acceleration vibration sensor, thermal infrared imager, hydrophone, and MEMS microelectromechanical sensor; wherein, the heterogeneous sensors work together to capture stress waves induced by leakage, changes in seepage pressure, structural vibration response, and abnormal temperature field signals, respectively.

10. A method for monitoring and locating leakage in hydraulic engineering structures based on multi-sensor collaboration, applied to the system described in any one of claims 1 to 9, characterized in that, Includes the following steps: Step S1: Real-time collection of multi-source time-series monitoring data, including vibration, pressure, temperature and acoustic emission, is achieved through a heterogeneous sensor network deployed on the hydraulic engineering structure. Step S2: Construct a heterogeneous graph of the hydraulic engineering structure containing sensor nodes and virtual physics field nodes, where the virtual physics field nodes represent intermediate state variables that satisfy the seepage mechanics equations; Step S3: Input the multi-source temporal monitoring data into a pre-trained spatiotemporal attention and physical constraint graph convolutional network. The network adaptively aligns the response delay of heterogeneous sensors in the temporal dimension through deformable convolution, and dynamically adjusts the graph attention weights between nodes based on physical residuals in the spatial dimension to propagate and fuse multimodal information. Step S4: Generate a continuous spatial leakage probability distribution field with sparse constraints through the Poisson reconstruction module in the network output layer, and determine the specific coordinates and influence range of the leakage point; Step S5: Superimpose the leakage probability distribution field onto the three-dimensional digital model of the water conservancy project, and trigger alarms of different levels when the leakage probability or leakage flow exceeds the preset safety threshold.

Images