A city water asset dynamic risk assessment method based on a physically enhanced graph network

Abstract

The application relates to the technical field of urban water system asset management, and discloses a kind of urban water asset dynamic risk assessment method based on physical reinforcement graph network, first acquire urban water system asset multi-source data, carry out spatio-temporal alignment processing;The spatio-temporal alignment data is fused with static attribute, spatial geographic information to construct dynamic heterogeneous graph;The dynamic heterogeneous graph is input into the pre-trained physical reinforcement graph neural network model, and the failure probability index and the failure consequence index are output through the graph attention mechanism with physical causal mask constraint, the real-time risk level is determined and closed-loop management and control are executed;Model evolution is realized through Bayesian update and incremental learning.The application deeply integrates the physical laws of hydraulics with graph neural networks, solving the problems of spatio-temporal misalignment of monitoring data, insufficient utilization of topological correlation, lack of physical constraints in models, and inability to adaptively evolve, and achieving precise, interpretable, and adaptive dynamic risk assessment of urban water system assets.

Description

Technical Field

[0001] This invention relates to the field of urban water system asset management technology, and in particular to a dynamic risk assessment method for urban water assets based on physical augmented graph networks. Background Technology

[0002] Urban water systems (including water supply networks, drainage networks, sewage treatment plants, pumping stations, and other facilities) are a core component of urban infrastructure. Their safe and stable operation is directly related to the quality of life of urban residents, urban flood control and drainage safety, and the ecological health of the water environment. With the acceleration of urbanization and the intensification of the impact of climate change, urban water systems are facing increasing operational pressure and a significant rise in the risk of asset failure.

[0003] Urban water system assets have the following significant characteristics: they are diverse and numerous, including various types such as pipelines, manholes, pumping stations, and sewage treatment equipment; they are widely distributed spatially and have complex topological relationships, with assets forming a complex network structure through hydraulic connections, where failure in one place can trigger a chain reaction; they are subject to significant dynamic changes, with asset status affected by external factors such as rainfall, hydrology, and seasons, resulting in time-varying risks; and the consequences of failure are multidimensional, not only causing losses to the assets themselves but also potentially leading to social impacts such as urban flooding, environmental pollution, and traffic disruption.

[0004] Existing methods for risk assessment of urban water system assets suffer from the following technical shortcomings: Traditional risk assessment methods directly align upstream and downstream monitoring data by timestamp, ignoring the propagation delay of hydraulic signals in the pipeline network, resulting in physical misalignment of input data and an inability to accurately identify the causal relationship of risk propagation; Most methods treat assets as independent entities, failing to fully utilize the pipeline network topology and making it difficult to quantify the impact of a single asset failure on the overall system; Purely data-driven models may output results that violate physical laws, resulting in poor interpretability; Purely physical models are computationally complex, have difficult-to-calibrate parameters, and are difficult to apply in real time; They only output risk level values, lacking confidence information and unable to support highly reliable decision-making; Once established, assessment models remain unchanged for a long time, unable to adapt to long-term evolution such as system modification and climate change, and also unable to cope with small sample problems under extreme conditions.

[0005] Therefore, there is an urgent need for a dynamic risk assessment method that can integrate physical laws with data-driven approaches, make full use of topological structures, and possess continuous evolution capabilities. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a dynamic risk assessment method for urban water assets based on physical augmented graph networks, which overcomes the technical problems of spatiotemporal misalignment of monitoring data, insufficient utilization of topological associations, lack of physical constraints in the model, single assessment output, and inability of the model to adaptively evolve.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, the main technical solutions adopted in this application include:

[0010] This application provides a method for dynamic risk assessment of urban water assets based on physical augmented graph networks, the specific steps of which include:

[0011] S1. Acquire real-time hydraulic monitoring data, static physical attribute data, spatial geographic information data, current hydrological conditions factors, and pipeline topology connections of each asset in the target city's water system; wherein, the real-time hydraulic monitoring data is obtained from monitoring equipment installed at each asset.

[0012] The location of the monitoring equipment constitutes a monitoring node;

[0013] S2. Based on the real-time hydraulic monitoring data and the pipeline network topology, calculate the propagation delay of the hydraulic signal reflected in the real-time hydraulic monitoring data in the pipeline network; perform spatiotemporal alignment processing on the real-time hydraulic monitoring data based on the propagation delay to obtain a spatiotemporally aligned feature dataset.

[0014] S3. Perform multi-source feature encoding and fusion on the spatiotemporally aligned feature dataset, the static physical attribute data, and the spatial geographic information data to obtain a dynamic heterogeneous map;

[0015] S4. Input the dynamic heterogeneous graph into a pre-trained physical augmented graph neural network model, and output the failure probability index and failure consequence index of each asset through a graph attention mechanism with physical causal mask constraints.

[0016] The physical augmentation graph neural network model adopts a multi-task learning architecture, which includes a shared spatiotemporal feature extraction backbone network, as well as a parallel failure probability prediction head and a failure consequence assessment head; wherein the failure probability prediction head is used to output the failure probability index, and the failure consequence assessment head is used to output the failure consequence index.

[0017] S5. Determine the risk classification threshold based on the current hydrological conditions; determine the level of the failure probability index and the failure consequence index based on the risk classification threshold to obtain the real-time risk level of each asset; based on the real-time risk level, match control instructions from the preset multi-objective optimization strategy library and execute closed-loop control operations.

[0018] Optionally, in some embodiments of this application, the dynamic heterogeneous graph includes:

[0019] The system includes various types of graph nodes, such as nodes for monitoring inspection wells, nodes for monitoring pump stations, and nodes for monitoring pipe sections. Each graph node corresponds one-to-one with the monitoring equipment. Each graph node contains the static attribute characteristics, spatial location characteristics, and real-time status characteristics of its corresponding asset.

[0020] Multiple types of edges are included, such as hydraulic connection edges and control edges; each edge contains its corresponding asset topology characteristics and hydraulic parameter characteristics.

[0021] Optionally, in some embodiments of this application, the method further includes the following step before S1:

[0022] S01. Obtain historical monitoring data and corresponding historical fault labels. Based on the pipeline network topology connection relationship, use the hydraulic propagation model to perform physical alignment and physical constraint repair on the historical monitoring data to generate a historical physical causal alignment sample set.

[0023] Simultaneously, extreme operating condition samples are extracted from the historical physical causal alignment sample set; an adversarial generative network is trained based on the extreme operating condition samples to generate synthetic data under extreme operating conditions; the synthetic data is merged into the historical physical causal alignment sample set to obtain an expanded historical physical causal alignment sample set.

[0024] S02. Iteratively train the initial physics-enhanced graph neural network model based on the expanded historical physics causal alignment sample set;

[0025] The initial physical augmentation graph neural network model includes a graph attention layer and a temporal convolutional layer; the graph attention layer is used to learn the hydraulic dependence between upstream and downstream nodes in the pipeline network and assign attention weights to the feature aggregation between different nodes; the temporal convolutional layer is used to extract the temporal change features of each monitoring node within the time window;

[0026] During training, the hybrid loss function is minimized through backpropagation until it converges, thus obtaining the pre-trained physical augmentation graph neural network model. The hybrid loss function is composed of a weighted average of a data-driven loss term and a physical constraint loss term from two task heads. The two task heads are a failure probability prediction head corresponding to the failure probability index and a failure consequence assessment head corresponding to the failure consequence index.

[0027] The data-driven loss term is used to minimize the error between the failure probability index and failure consequence index predicted by the initial physical augmentation graph neural network model and the historical failure labels; the physical constraint loss term is used to calculate the residual value of the output result of the initial physical augmentation graph neural network model that violates the hydraulic conservation equation, and to penalize the residual value to constrain the output of the initial physical augmentation graph neural network model to conform to the physical mechanism; wherein, the physical constraint loss term includes a mass conservation residual term and an energy conservation residual term.

[0028] Optionally, in some embodiments of this application, step S2, performing spatiotemporal alignment processing on the real-time hydraulic monitoring data based on the propagation delay, includes:

[0029] Based on the pipeline network topology and the physical attributes of each pipeline extracted from static physical attribute data, the location and connection relationship of each monitoring node are determined.

[0030] The theoretical propagation delay of the hydraulic state signal between any two monitoring nodes is calculated using a hydraulic propagation model. The theoretical propagation delay is determined based on the ratio of the pipe section length to the average flow velocity under the current operating conditions. The average flow velocity is calculated using the Manning formula based on real-time hydraulic monitoring data.

[0031] Using the current sampling time of the downstream monitoring node as a reference, the real-time hydraulic monitoring data of the upstream monitoring node is shifted backward on the time axis using the theoretical propagation delay, so that the upstream and downstream monitoring data are aligned in physical time.

[0032] Optionally, in some embodiments of this application, S3 includes:

[0033] The time window length is selected based on the current hydrological conditions factors, including rainfall intensity level and river level.

[0034] When the rainfall intensity level is higher than the preset threshold or the river water level is higher than the warning level, the time window length is shortened; when the rainfall intensity level is lower than the preset threshold and the river water level is lower than the warning level, the time window length is extended.

[0035] Within the time window, the real-time status feature sequence of each monitoring node is extracted and used as the real-time status feature of the corresponding node in the dynamic heterogeneous graph.

[0036] Optionally, in some embodiments of this application, in step S4, the graph attention mechanism with physical causal mask constraints includes:

[0037] The physical causal mask constraint, based on the water flow direction determined by the pipeline topology connection relationship and the physical reachability determined by the comparison result of the propagation delay and the preset delay threshold, masks non-physically related node connections when calculating attention weights.

[0038] Optionally, in some embodiments of this application, S5 includes:

[0039] A dynamic adjustment coefficient is generated based on the current hydrological conditions factors.

[0040] Using the dynamic adjustment coefficient, the preset risk classification threshold is corrected in real time to obtain the currently effective first classification threshold corresponding to the failure probability index and the second classification threshold corresponding to the failure consequence index.

[0041] The failure probability index is compared with the first grading threshold and converted into a probability level; the failure consequence index is compared with the second grading threshold and converted into a consequence level.

[0042] The real-time risk level of each asset is determined based on the probability level and consequence level.

[0043] Optionally, in some embodiments of this application, step S5, based on the real-time risk level, matches control instructions from a preset multi-objective optimization strategy library and executes closed-loop control operations, including:

[0044] Based on the real-time risk level, the data sampling frequency of the sensors associated with each asset is dynamically adjusted: when the real-time risk level is higher than a preset threshold, the sampling frequency is increased; when the real-time risk level is lower than the preset threshold, the sampling frequency is decreased or the sensor is switched to sleep mode.

[0045] And / or, based on the real-time risk level, dynamically activate or put into sleep mode the backup sensors: when the real-time risk level is higher than a preset threshold, activate the surrounding associated backup sensors to form a local encrypted monitoring network; when the real-time risk level is lower than the preset threshold, switch the backup sensors to sleep mode.

[0046] Optionally, in some embodiments of this application, the method further includes the following step before S2:

[0047] The preset risk red line conditions include at least one of the following: pipe diameter does not meet the design standard, pipe section structure fails, pipe section blockage causes water flow cross-sectional area loss to exceed 50%, and pipe section overflows within the design standard range.

[0048] When an asset in the target city's water system triggers the aforementioned risk red line conditions, the asset is directly deemed invalid, and an emergency control order is executed.

[0049] Optionally, in some embodiments of this application, the method further includes the following after S5:

[0050] Collect feedback data after the execution of the differentiated closed-loop control instructions, including actual fault occurrence and actual fault loss data;

[0051] Using the Bayesian update algorithm, the weight parameters in the pre-trained physical augmentation graph neural network model are corrected in reverse based on the feedback data;

[0052] When the accumulated amount of feedback data reaches a preset threshold, an incremental learning mechanism is triggered, which updates only the network layer parameters near the output end in the pre-trained physical augmentation graph neural network model, while keeping the network layer parameters near the input end unchanged, in order to prevent catastrophic forgetting.

[0053] (III) Beneficial Effects

[0054] This application presents a dynamic risk assessment method for urban water assets based on physically augmented graph networks. First, by calculating the propagation delay of hydraulic signals in the pipeline network and performing spatiotemporal alignment based on this delay, it fundamentally solves the problem of spatiotemporal misalignment of upstream and downstream data caused by simple timestamp alignment in traditional methods, ensuring the accuracy of the model input in terms of physical causality. Second, it constructs a dynamic heterogeneous graph from multi-source heterogeneous data and introduces a physical causal mask based on water flow direction and propagation delay into the graph attention mechanism. This effectively shields non-physically related feature aggregation, ensuring that the model output strictly follows the laws of hydraulic propagation and significantly enhancing the interpretability of the assessment results. It achieves consistency with physical data; secondly, it adopts a multi-task learning architecture that includes a shared backbone network and parallel prediction heads, enabling synchronous and high-precision output of failure probability index and fault consequence index, avoiding information fragmentation in single-task models; finally, by introducing real-time hydrological condition factors to dynamically adjust the risk classification threshold, it overcomes the shortcomings of fixed thresholds in terms of insufficient sensitivity during the flood season or high false alarm rate during the dry season, and performs closed-loop operations by matching corresponding control instructions from a multi-objective optimization strategy library based on real-time risk levels, realizing full-process automation from accurate assessment to intelligent decision-making, significantly improving the risk response capability and management efficiency of urban water systems in complex dynamic environments. Attached Figure Description

[0055] Figure 1 This is a flowchart of a method for dynamic risk assessment of urban water assets based on physical augmented graph networks, as proposed in this application. Detailed Implementation

[0056] To better explain and facilitate understanding of this application, the following detailed description of the application is provided in conjunction with the accompanying drawings and specific embodiments.

[0057] In existing technologies, methods for dynamic risk assessment of urban water assets based on physical augmented graph networks can be mainly categorized into the following two types:

[0058] The first category is assessment methods based on physical or rule-based models. These methods rely on hydraulic equations (such as the Saint-Venant equation and Manning's formula) or manually defined rules (such as risk matrices and indicator scoring tables) to quantify asset risk. While these methods have clear physical meaning and interpretability, they suffer from the following drawbacks: First, model parameters (such as roughness coefficients and propagation delays) are difficult to calibrate precisely, and the computational complexity is high, making it difficult to meet the needs of real-time dynamic assessment. Second, rule-based models rely on expert experience, and fixed thresholds cannot adapt to the dynamic changes in hydrological conditions. Third, they assess individual assets in isolation, ignoring the impact of pipeline topology on risk propagation, making it difficult to accurately quantify the consequences of failure.

[0059] The second category is machine learning evaluation methods based on pure data-driven approaches. These methods use historical monitoring data to train neural network models and predict asset risks through data fitting. While these methods can automatically learn complex patterns in the data, they have the following drawbacks: First, the model inputs typically use simple timestamp alignment, ignoring the propagation delay of hydraulic signals in the pipeline network, resulting in physically misaligned input data and an inability to accurately identify the causal relationships of risk propagation. Second, pure data-driven models lack physical constraints and may output predictions that violate hydraulic conservation laws, leading to poor interpretability. Third, the model structure does not fully utilize the pipeline network topology, treating assets as independent entities and making it difficult to learn the interactions between nodes. Fourth, once trained, the model becomes fixed and unchanging, unable to adapt to long-term evolution such as system modifications and climate change, and also unable to cope with small sample problems under extreme conditions.

[0060] To address this, this application provides a method for dynamic risk assessment of urban water assets based on physically augmented graph networks. This method deeply integrates the physical laws of hydraulics with graph neural networks, solves the data misalignment problem through physical-guided spatiotemporal alignment, uniformly represents multi-source heterogeneous information and topological relationships through dynamic heterogeneous graphs, constrains graph attention mechanisms to conform to the laws of hydraulic propagation through physical causal masking, synchronously outputs failure probabilities and fault consequences through a multi-task learning architecture, matches real-time hydrological conditions through adaptive thresholds, and achieves continuous model evolution through Bayesian updates and incremental learning. This overcomes the shortcomings of the two existing technologies and realizes accurate, interpretable, adaptive dynamic assessment and intelligent management of urban water system asset risks.

[0061] To better explain and facilitate understanding of this application, a detailed description of its embodiments is provided below in conjunction with the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a clearer and more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0062] Example 1

[0063] Figure 1 According to this application, a dynamic risk assessment method for urban water assets based on physical augmented graph networks is proposed, such as... Figure 1 As shown, this method for dynamic risk assessment of urban water assets based on physical augmented graph networks includes:

[0064] S1. Acquire real-time hydraulic monitoring data, static physical attribute data, spatial geographic information data, current hydrological conditions factors, and pipeline topology connections of each asset in the target city's water system; wherein, the real-time hydraulic monitoring data is obtained from monitoring equipment installed at each asset.

[0065] The location of the monitoring equipment constitutes a monitoring node;

[0066] S2. Based on the real-time hydraulic monitoring data and the pipeline network topology, calculate the propagation delay of the hydraulic signal reflected in the real-time hydraulic monitoring data in the pipeline network; perform spatiotemporal alignment processing on the real-time hydraulic monitoring data based on the propagation delay to obtain a spatiotemporally aligned feature dataset.

[0067] S3. Perform multi-source feature encoding and fusion on the spatiotemporally aligned feature dataset, the static physical attribute data, and the spatial geographic information data to obtain a dynamic heterogeneous map;

[0068] S4. Input the dynamic heterogeneous graph into a pre-trained physical augmented graph neural network model, and output the failure probability index and failure consequence index of each asset through a graph attention mechanism with physical causal mask constraints.

[0069] The physical augmentation graph neural network model adopts a multi-task learning architecture, which includes a shared spatiotemporal feature extraction backbone network, as well as a parallel failure probability prediction head and a failure consequence assessment head; wherein the failure probability prediction head is used to output the failure probability index, and the failure consequence assessment head is used to output the failure consequence index.

[0070] S5. Determine the risk classification threshold based on the current hydrological conditions; determine the level of the failure probability index and the failure consequence index based on the risk classification threshold to obtain the real-time risk level of each asset; based on the real-time risk level, match control instructions from the preset multi-objective optimization strategy library and execute closed-loop control operations.

[0071] This embodiment constructs a complete closed-loop technical solution: First, by acquiring real-time hydraulic data from monitoring devices located at each asset and combining it with propagation delay for physical guidance of spatiotemporal alignment, it solves the problem of data spatiotemporal misalignment caused by neglecting the propagation time of hydraulic signals in traditional methods, ensuring that the input data accurately corresponds in terms of physical causality, laying a reliable foundation for subsequent evaluation; Second, it integrates the aligned feature data with static physical attributes and spatial geographic information to construct a dynamic heterogeneous graph, uniformly representing the pipeline topology, asset type, spatial location, and real-time status, solving the technical challenge of collaborative modeling of multi-source heterogeneous information; On this basis, the dynamic heterogeneous graph is input into a pre-trained physical augmentation graph neural network model, which is then enhanced through a graph attention mechanism with physical causality mask constraints. By shielding non-physically related node connections, feature aggregation conforms to the laws of hydraulic propagation. A multi-task learning architecture synchronously outputs failure probability and fault consequence indices, achieving a deep integration of physical laws and data-driven approaches. This ensures both model interpretability and computational efficiency. Subsequently, risk classification thresholds are dynamically adjusted based on current hydrological conditions, enabling risk assessment criteria to match different conditions such as flood season and heavy rainfall in real time, addressing the insufficient adaptability of fixed thresholds. Finally, based on real-time risk levels, control instructions are matched from a multi-objective optimization strategy library and executed in a closed-loop manner, achieving an automated closed loop from assessment to decision-making. This supports multi-dimensional optimization of cost, energy consumption, and risk, significantly improving the intelligence level of urban water system asset risk management and resource utilization efficiency.

[0072] Example 2

[0073] Before using the model, this embodiment includes a training step for the physics-enhanced graph neural network model:

[0074] S01. Historical data processing and sample set construction;

[0075] Historical monitoring data and corresponding historical fault labels are obtained. Based on the pipeline network topology, the historical monitoring data are physically aligned and physically constrained and repaired using a hydraulic propagation model to generate a historical physical causal alignment sample set.

[0076] Specifically, obtain historical monitoring data of the city's drainage network for the past 5 years, including:

[0077] Water level data for each inspection well (sampling frequency 5 minutes / time);

[0078] Flow data for each pipe section; operating status data for each pumping station;

[0079] The corresponding historical fault tags (records of events such as pipe rupture, blockage, and overflow, totaling 237).

[0080] Based on the pipeline network topology, a hydraulic propagation model is used to physically align and physically constrain historical monitoring data, generating a physical causal alignment sample set.

[0081] Taking a typical pipe section as an example:

[0082] Upstream manhole A and downstream manhole B, with a pipe section length L=1200m, pipe diameter D=800mm, material is concrete (roughness coefficient n=0.013), slope i=0.003.

[0083] The average flow velocity v over a historical period is estimated using Manning's formula, and the propagation delay is calculated. t=L / v, shifting the historical data of point A upstream backward. t, to achieve physical alignment.

[0084] Meanwhile, for outliers or missing values ​​in historical monitoring data, the Saint-Venant equation is used to verify the outlier / missing data to ensure the physical consistency of the dataset.

[0085] For example, if the response relationship between the water level change at point A upstream and the water level change at point B downstream within the theoretical propagation time delay violates the continuity condition of the Saint-Venant equation, then the data set is determined to be abnormal, and the abnormal data is replaced with the simulated value of the Saint-Venant equation, thereby ensuring the physical consistency of the historical dataset.

[0086] Simultaneously, extreme condition samples (such as severe rainstorms and pipeline rupture events, totaling 23 incidents) were extracted from the historical physical causal alignment sample set. A generative adversarial network (GAN) was trained based on these extreme condition samples to generate 300 sets of synthetic data under extreme conditions. These synthetic data were then merged into the historical physical causal alignment sample set to obtain an expanded historical physical causal alignment sample set, totaling approximately 25,000 samples.

[0087] S02. Iteratively train the initial physics-enhanced graph neural network model based on the expanded historical physics causal alignment sample set;

[0088] The initial physically enhanced graph neural network model includes a graph attention layer and a temporal convolutional layer:

[0089] The graph attention layer consists of three layers, each with eight attention heads, used to learn the hydraulic dependencies between upstream and downstream nodes in the pipeline network and to assign attention weights to the feature aggregation of different nodes. The temporal convolutional layer consists of two layers with a kernel size of 5, used to extract the temporal variation features of each monitoring node within a time window. The model input consists of node features (static attributes, spatial location, real-time status) and edge features (topological relationships, hydraulic parameters), and the output consists of the predicted values ​​from the two task heads.

[0090] Because urban water systems exhibit multipath propagation (e.g., an upstream node may be connected to the same downstream node via multiple parallel pipe sections), calculating the propagation time solely based on the shortest path or simple average would result in significant distortion. Therefore, this embodiment introduces a flow-driven dynamic weighting mechanism, calculating the overall propagation delay using the following formula:

[0091]

[0092] Where K is the number of all connected physical paths between the upstream monitoring node and the downstream monitoring node;

[0093] Let L be the theoretical propagation delay of the j-th path. j v is the total length of the path. j The average flow velocity of this path under the current operating conditions (calculated using the Manning formula based on real-time hydraulic monitoring data).

[0094] w j The flow weight for the j-th path is the normalized value of the minimum pipe segment flow on that path, i.e. Q j This represents the real-time flow rate of that path. When a path experiences a flow rate decrease due to a valve closure or pipe blockage, Q... j When the value is 0, its weight is automatically set to zero and it is not included in the overall delay calculation.

[0095] The comprehensive propagation delay T calculated in this embodiment composite It is used for the physical alignment of upstream and downstream data, and also serves as the basic input for subsequent risk propagation path analysis; the longer the time delay of the path, the lower its contribution weight in risk propagation, and vice versa.

[0096] Through the aforementioned flow-driven dynamic weighting mechanism, the main paths with high flow rates and fast velocities are assigned high weights, and their propagation delays play a dominant role in the final result; while the side paths with low flow rates and slow velocities are assigned low weights, and their impact is relatively small. This mechanism allows the graph attention layer to consider the combined effects of multiple hydraulic propagation paths simultaneously, rather than relying solely on a single shortest path, thus more realistically reflecting the hydraulic propagation characteristics in complex pipe networks.

[0097] During training, the hybrid loss function is minimized through backpropagation until it converges (approximately 1500 iterations), resulting in the pre-trained physical augmented graph neural network model.

[0098] The hybrid loss function is composed of a weighted average of data-driven loss terms and physical constraint loss terms from two task heads (data-driven weight 0.6, physical constraint weight 0.4); the two task heads are a failure probability prediction head corresponding to the failure probability index and a failure consequence assessment head corresponding to the failure consequence index.

[0099] The data-driven loss term is used to minimize the error between the failure probability index and failure consequence index predicted by the initial physical augmentation graph neural network model and the historical failure labels, and the mean square error function is used.

[0100] The physical constraint loss term is used to calculate the residual value of the output result of the initial physical enhancement graph neural network model that violates the hydraulic conservation equation, and to penalize the residual value to constrain the output of the initial physical enhancement graph neural network model to conform to the physical mechanism; wherein, the physical constraint loss term includes a mass conservation residual term and an energy conservation residual term.

[0101] The mass conservation residual term is constructed based on the node flow continuity equation. It is used to calculate the difference between the total flow into a node and the total flow out of the node predicted by the model, and to penalize the difference to satisfy the flow balance at the node.

[0102] The energy conservation residual term is constructed based on the pipeline energy equation. It is used to calculate the difference between the head difference at the two ends of the pipeline predicted by the model and the sum of the friction head loss and local head loss of the pipeline under the predicted flow rate. This difference is penalized to satisfy the law of energy conservation.

[0103] By incorporating physical constraint loss into the total loss function, the model not only fits historical data during training but also learns to satisfy the fundamental laws of hydraulics, thereby improving the model's interpretability and generalization ability.

[0104] This embodiment constructs a complete training process for a physically enhanced graph neural network model, introducing physical constraints into multiple stages such as data preprocessing, sample augmentation, model structure, and loss function. This achieves a deep integration of data-driven approaches and physical mechanisms, resulting in the following beneficial effects:

[0105] By calculating the propagation delay using Manning's formula and physically aligning historical monitoring data, the problem of spatiotemporal misalignment caused by traditional timestamp alignment is solved, ensuring accurate physical causal correspondence between upstream and downstream monitoring data and providing physically consistent input features for the model. Simultaneously, the Saint-Venant equation is used for consistency verification and repair of outliers, effectively identifying and correcting abnormal records in historical data that violate hydraulic principles, thus improving the reliability of the training sample set.

[0106] To address the scarcity of extreme working condition samples in urban water systems, GANs are used to generate synthetic data based on a small number of real extreme samples, which can effectively expand the training set and expose the model to more diverse working conditions during training, thereby enhancing the model's generalization ability to rare extreme events.

[0107] A flow-driven dynamic weighting mechanism is introduced, which comprehensively considers all connectivity paths between upstream and downstream nodes, rather than relying solely on a single shortest path. By allocating flow weights, the propagation delay of the main path plays a dominant role, while the influence of side paths is relatively small. This more accurately reflects the hydraulic propagation characteristics in complex pipe networks and avoids information distortion caused by improper path selection.

[0108] By incorporating mass conservation and energy conservation residuals into the hybrid loss function, the model not only fits historical data during training but also learns to satisfy fundamental laws of hydraulics. This mechanism constrains the model output to conform to physical laws, reducing predictions that violate flow continuity or energy conservation, thereby improving the model's interpretability and physical consistency.

[0109] By employing a shared spatiotemporal feature extraction backbone network and parallel failure probability prediction and failure consequence assessment heads, the two tasks can share underlying features. This architecture leverages the correlations between multiple tasks to facilitate feature learning while simplifying the model structure and reducing training complexity.

[0110] Example 3

[0111] Based on the physical augmented graph neural network model trained in Example 2, the model is then used in the online application process.

[0112] When a city experiences a short period of heavy rainfall, the system activates a real-time risk assessment.

[0113] The flowchart of the urban water asset dynamic risk assessment method based on physical augmented graph networks is shown below. Figure 1 As shown, this method for dynamic risk assessment of urban water assets based on physical augmented graph networks includes:

[0114] S1. Obtain real-time hydraulic monitoring data, static physical attribute data, spatial geographic information data, current hydrological conditions factors, and pipeline topology connections of each asset in the target city's water system; wherein, the real-time hydraulic monitoring data is obtained from monitoring equipment installed at each asset.

[0115] In this embodiment, the real-time hydraulic monitoring data includes:

[0116] Upstream inspection well J01: water level 2.1m, flow rate 0.3m³ / s;

[0117] Downstream inspection well J05: water level 1.8m, flow rate 0.2m³ / s;

[0118] Rain gauge: 35mm of rainfall per hour;

[0119] River water level: 2.0m;

[0120] The default sampling frequency of the monitoring equipment is 5 minutes / time.

[0121] Static physical property data includes:

[0122] Pipe section P12 (connecting J01 and J05): pipe diameter 800mm, pipe length 1200m, material concrete, slope 0.003, design life 50 years, used for 15 years, roughness coefficient 0.013.

[0123] Spatial geographic information data includes:

[0124] Latitude and longitude coordinates and ground elevation of J01 and J05;

[0125] Affected area: Population density 12,000 people / km², road type: urban main road, 200m from the nearest primary school.

[0126] Current hydrological conditions include:

[0127] Real-time rainfall intensity: 35 mm / h (heavy rain);

[0128] Receiving water level in the river: 2.0m (2.5m below the warning level).

[0129] The pipeline network topology connections include:

[0130] The direction of water flow from J01 to P12 to J05, and the connection relationship between J01, J05 and other nodes.

[0131] The sequence following S1 and preceding S2 includes rapid response steps for risk red lines:

[0132] The preset risk red line conditions include at least one of the following: pipe diameter does not meet the design standard, pipe section structure fails, pipe section blockage causes water flow cross-sectional area loss to exceed 50%, and pipe section overflows within the design standard range.

[0133] When an asset in the target city's water system triggers the aforementioned risk red line conditions, the asset is directly deemed invalid, skipping the routine assessment process from S2 to S5, and directly executing emergency control instructions.

[0134] In this case, none of the assets triggered the risk threshold conditions, and the process proceeded directly to the standard procedure.

[0135] S2. Based on the real-time hydraulic monitoring data and the pipeline network topology, calculate the propagation delay of the hydraulic signal reflected in the real-time hydraulic monitoring data in the pipeline network; perform spatiotemporal alignment processing on the real-time hydraulic monitoring data based on the propagation delay to obtain a spatiotemporally aligned feature dataset.

[0136] S21. Based on the pipeline network topology and the physical attributes of each pipeline extracted from static physical attribute data, determine the location and connection relationship of each monitoring node; wherein, the monitoring node corresponds to the location of the manhole, pump station, or pipeline section where monitoring equipment is installed.

[0137] In this embodiment, J01 and J05 are both inspection wells equipped with water level gauges and are monitoring nodes.

[0138] S22. Calculate the theoretical propagation delay of the hydraulic state signal between any two monitoring nodes using a hydraulic propagation model; wherein the theoretical propagation delay is determined based on the ratio of the pipe section length to the average flow velocity under the current operating conditions, and the average flow velocity is calculated based on real-time hydraulic monitoring data using the Manning formula.

[0139] For pipe segment P12 between J01 and J05, the average flow velocity is calculated using Manning's formula: ;

[0140] Where v is the average flow velocity of the cross section, n is the Manning roughness coefficient, R is the hydraulic radius, and S is the hydraulic gradient (taken as the slope of the pipe section).

[0141] In this embodiment, the system uses the real-time monitoring of water level h=2.1m and pipe diameter D=800mm by J01 to calculate the water flow area A and wetted perimeter P in real time using the geometric relationship of the circular cross section, and obtains R=A / P≈0.2m.

[0142] Substituting n=0.013, R=0.2m, and S=0.003, the average flow velocity under the current operating condition is calculated to be 1.2m / s;

[0143] Calculate propagation delay t = L / v = 16.7 minutes; where L is the length of the pipe section.

[0144] This calculation step demonstrates the precise adaptability of this application to non-full-flow dynamic operating conditions, which is different from traditional simplified algorithms.

[0145] S23. Using the current sampling time (10:00) of the downstream monitoring node J05 as a reference, the real-time hydraulic monitoring data of the upstream monitoring node J01 is shifted backward by 16.7 minutes on the time axis using the theoretical propagation delay, so that the upstream and downstream monitoring data are aligned in physical time.

[0146] After alignment, the water level of J01 at 09:43:18 (2.1m) and J05 at 10:00 (1.8m) form a causal relationship. The Pearson correlation coefficient of the two sequences after translation is calculated to be 0.89, which is greater than the preset threshold of 0.7, thus confirming them as causal related samples.

[0147] S3. Perform multi-source feature encoding and fusion of the spatiotemporally aligned feature dataset with the static physical attribute data and spatial geographic information data to obtain a dynamic heterogeneous map.

[0148] The dynamic heterogeneous graph includes:

[0149] The system includes various types of graph nodes, such as nodes for monitoring inspection wells, nodes for monitoring pump stations, and nodes for monitoring pipe sections. Each graph node corresponds one-to-one with the monitoring equipment. Each graph node contains the static attribute characteristics, spatial location characteristics, and real-time status characteristics of its corresponding asset.

[0150] Multiple types of edges are included, such as hydraulic connection edges and control edges; each edge contains its corresponding asset topology characteristics and hydraulic parameter characteristics.

[0151] Hydraulic connection edges represent the connections where water flows naturally within a pipe network, reflecting passive hydraulic transmission. Control edges represent the connections where water flow is artificially intervened through active control equipment (such as pumping stations, gates, and valves), reflecting active control behavior.

[0152] S31. Select the time window length based on the current hydrological conditions factors; the hydrological conditions factors include rainfall intensity level and river level.

[0153] In this embodiment, the rainfall intensity level is heavy rain (>30mm / h), and the river water level is 2.0m, which is lower than the warning water level of 2.5m. Therefore, the time window length is shortened to 5 minutes.

[0154] S32. When the rainfall intensity level is higher than the preset threshold (30mm / h) or the river water level is higher than the warning level, shorten the time window length to capture rapid changes; when the rainfall intensity level is lower than the preset threshold and the river water level is lower than the warning level, lengthen the time window length to smooth out noise.

[0155] This embodiment satisfies the former condition, and the window length is set to 5 minutes.

[0156] S33. Extract the real-time status feature sequence of each monitoring node within the time window, and use it as the real-time status feature of the corresponding node in the dynamic heterogeneous graph.

[0157] Within a 5-minute time window, the real-time status feature sequences of each monitoring node (J01 water level: [2.1, 2.3, 2.8, 3.2, 3.5]; J05 water level: [1.8, 1.9, 2.1, 2.3, 2.6]) are extracted as the real-time status features of the corresponding nodes in the dynamic heterogeneous graph.

[0158] Simultaneously, static attributes (material, pipe diameter, design life) and spatial location (coordinates, elevation) are encoded as node features to construct a dynamic heterogeneous graph.

[0159] Example of a dynamic heterogeneous graph:

[0160] Diagram nodes: inspection wells, pumping stations;

[0161] Side: pipe section;

[0162] Time dimension: temporal characteristics;

[0163] Spatial dimension: location.

[0164] In this embodiment, monitoring nodes J01 and J05 are mapped to inspection well nodes in the dynamic heterogeneous graph, and pipe segment P12 is mapped to a hydraulic connection edge. The edge features include pipe length, pipe diameter, slope, roughness coefficient, current flow velocity, and propagation delay.

[0165] The structure of the dynamic heterogeneous graph is updated in real time according to the system's operating status. When a new monitoring device, pipeline section modification, or change in the start-up or shutdown status of a pumping station is detected, the system automatically triggers graph structure reconstruction: updating the set of nodes and edges according to the latest pipeline topology connections, updating node characteristics according to the latest static attributes, and updating the dynamic characteristics of nodes and edges according to the latest real-time monitoring data. The graph structure update frequency is synchronized with the risk assessment cycle (once every 5 minutes in this embodiment) to ensure that the graph model always reflects the current physical state of the system.

[0166] S34. Physical enhancement of edge features; The edge features in the dynamic heterogeneous graph include not only static pipe length, pipe diameter, slope, and roughness coefficient, but also real-time hydraulic impedance coefficient and signal attenuation factor.

[0167] The hydraulic impedance coefficient is used to characterize the resistance to water flow in the pipe. This coefficient is calculated based on the dynamic correction values ​​of real-time flow rate, flow velocity and pipe wall roughness, and can dynamically reflect the hydraulic transmission efficiency under different working conditions.

[0168] The signal attenuation factor is used to quantify the loss of risk signals (such as pressure waves and pollution plumes) due to friction and diffusion during long-distance transmission. This factor is determined based on the hydraulic impedance coefficient, pipe length, and number of intermediate nodes in the path. The longer the transmission distance and the greater the impedance, the smaller the attenuation factor, indicating greater signal loss.

[0169] When calculating the graph attention weights, the signal attenuation factor is used as the weight coefficient of the edge features and multiplied by the original attention weights. In this way, when aggregating neighbor node information, the model can automatically reduce the contribution of distant, high-impedance nodes, making the attention mechanism more focused on physically adjacent nodes that can effectively transmit signals.

[0170] Through the above mechanism, the graph neural network not only learns the connection relationship of the pipeline network topology, but also learns the strength of the connection and the loss of signal transmission, thereby improving the model's ability to capture long-distance hydraulic dependence.

[0171] S4. Input the dynamic heterogeneous graph into a pre-trained physical augmented graph neural network model, and output the failure probability index and failure consequence index of each asset through a graph attention mechanism with physical causal mask constraints.

[0172] The failure probability index P ranges from 0 to 5, with a higher value indicating a greater likelihood of asset failure in the near future. The failure consequence index C also ranges from 0 to 5, with a higher value indicating more severe consequences of failure. Subsequent steps will determine the risk level based on these indices.

[0173] The physical augmented graph neural network model adopts a multi-task learning architecture, which includes a shared spatiotemporal feature extraction backbone network, as well as parallel failure probability prediction head and failure consequence assessment head; wherein the failure probability prediction head is used to output the failure probability index, and the failure consequence assessment head is used to output the failure consequence index.

[0174] The physical causal mask constraint, based on the water flow direction determined by the pipeline topology and the physical reachability determined by comparing the propagation delay with a preset delay threshold, masks non-physically related node connections when calculating attention weights. The physical causal mask matrix possesses dynamic adaptive capabilities, automatically adjusting the feature aggregation time window according to real-time flow velocity. The time window is inversely proportional to the flow velocity: when heavy rain causes increased flow velocity, the threshold automatically shrinks, aggregating only rapidly arriving signals from nearby locations and excluding interference from lagging data from distant locations; when the flow velocity slows down, the threshold automatically expands, incorporating the slowly propagating hydraulic effects from distant locations into the calculation. Through this mechanism, the physical causal mask matrix can dynamically adjust the delay threshold according to real-time flow velocity, thereby enabling the attention mechanism to prioritize node connections that conform to the hydraulic propagation laws during feature aggregation, effectively suppressing interference from non-physically related nodes. This mechanism helps improve the physical consistency of the model under different operating conditions, ensuring good synchronization between attention weight allocation and the hydraulic propagation laws.

[0175] Based on the aforementioned basic physical causal mask, this embodiment further introduces a hierarchical constraint mechanism, dividing the mask into two layers: hard mask and soft mask, in order to achieve a balance between physical rigor and data-driven flexibility.

[0176] Hard Mask: Based on water flow direction constraints and extreme time delay constraints, it directly cuts off node connections that clearly violate physical laws. Specifically, when the water flow direction between two nodes is countercurrent (i.e., from downstream to upstream), or when the theoretical propagation delay exceeds a preset physical limit threshold, the attention coefficient of the corresponding node is directly set to 0, forcibly blocking the connection.

[0177] Soft Mask: For physically reachable (following the flow direction and with time delays within limits) but weakly correlated node pairs, instead of directly severing them, a penalty term related to the correlation strength is applied. The correlation strength is quantified based on real-time flow weights, historical correlation coefficients, or hydraulic impedance coefficients: the weaker the correlation, the larger the penalty term, causing the corresponding attention weight to approach 0 but not completely disappear. For example, when the historical water level changes between upstream and downstream nodes have low correlation, or when the flow rate along the path is extremely low under current conditions, the model will significantly reduce the weight of this connection during feature aggregation, but still retain a very weak correlation possibility.

[0178] Through the aforementioned hierarchical constraints, the hard mask ensures the model's rigor in terms of absolute physical correctness, preventing interference from non-physical connections; the soft mask, on the other hand, preserves the model's weak perception of potential unknown hydraulic connections (such as indirect connections caused by underground leakage, the slight influence of side pipes, and alternative paths when sensor data is abnormal), enabling the model to make reasonable inferences even under extreme conditions or when data is missing. This hierarchical constraint mechanism combines the advantages of physical laws and data-driven methods, improving the model's robustness and generalization ability under complex conditions.

[0179] In this embodiment, for any node pair (i,j), if the water flow direction does not allow the flow from i to j, or the propagation delay exceeds a preset threshold (e.g., 1800 seconds), the corresponding attention coefficient is set to zero before the graph attention weight is calculated, so that no feature aggregation is performed between the node pair.

[0180] Furthermore, based on the topology, the water flow direction is determined to be J01→J05. The propagation delay of 1000 seconds is less than the preset threshold, so there is a physical causal relationship between J01 and J05, allowing feature aggregation; while other nodes that are not directly hydraulically connected to J01 are shielded.

[0181] Output of the Physics Augmented Graph Neural Network Model:

[0182] J01 (upstream inspection well): failure probability index P=3.4, failure consequence index C=4.2;

[0183] J05 (downstream inspection well): failure probability index P=2.1, failure consequence index C=3.8.

[0184] S5. Determine the risk classification threshold based on the current hydrological conditions; determine the level of the failure probability index and the failure consequence index based on the risk classification threshold to obtain the real-time risk level of each asset; based on the real-time risk level, match control instructions from the preset multi-objective optimization strategy library and execute closed-loop control operations.

[0185] S51. Generate a dynamic adjustment coefficient based on the current hydrological conditions; use the dynamic adjustment coefficient to correct the preset risk classification threshold in real time to obtain the currently effective first classification threshold and second classification threshold.

[0186] In one specific embodiment of the present invention, the determination of the risk classification threshold based on the current hydrological conditions is achieved through the following nonlinear mapping function:

[0187] ;

[0188] Among them, Th baseα is the baseline threshold; α is the rainfall sensitivity coefficient, which controls the degree of influence of rainfall intensity on the threshold, and its value ranges from 0 to 1; β is the rainfall response rate coefficient, which controls the response speed of the threshold to changes in rainfall intensity; γ is the water level sensitivity coefficient, which controls the degree of influence of river water level on the threshold, and its value ranges from 0 to 1.

[0189] Given the current rainfall intensity R and the pipeline network design return period R... design The ratio, River water level H and warning water level H warn The difference.

[0190] The above parameters can be initially obtained by fitting offline historical data, or they can be optimized and updated online in conjunction with the subsequent threshold self-correction mechanism, so that the threshold function can adapt to the individual characteristics of different regions and different pipe networks, rather than a fixed and unchanging artificial rule.

[0191] This function maps hydrological condition factors R and H to dynamic thresholds. Its design principle is based on the physical response characteristics of urban water systems under extreme hydrological conditions.

[0192] (1) Saturation effect of rainfall intensity: When the rainfall intensity is close to the design return period of the pipeline network, the risk increases rapidly with the increase of rainfall intensity; after the design return period, the pipeline network is already in a full state, and the rate of risk increase slows down significantly, showing a saturation trend. The tanh function, due to its "fast at first and slow later" S-shaped characteristic, is naturally suitable for simulating this physical law.

[0193] (2) Sudden change characteristics of river water level: When the river water level is below the warning level, it has little impact on the drainage system; once it approaches or exceeds the warning level, the outflow is blocked and the risk of waterlogging will suddenly increase significantly. The Sigmoid function has a smooth "S"-shaped jump characteristic, which can accurately characterize the risk change when the water level crosses the critical value.

[0194] (3) Multi-factor synergistic amplification: The effects of rainfall and water level on risk are not linearly superimposed, but rather synergistically amplified—under high water levels, the same rainfall will cause a higher risk. The coupling structure of the two nonlinear terms in the formula realizes the quantitative expression of this synergistic effect through the multiplication factor.

[0195] In this embodiment, the current period is the flood season (June-September) and a yellow rainstorm warning is in effect. The system determines that the risk sensitivity needs to be increased, so the dynamic adjustment coefficient is set to 1.2.

[0196] This coefficient is used to adjust the preset risk classification benchmark threshold in real time. Specifically, the benchmark threshold is divided by a dynamic adjustment coefficient (i.e., the threshold is lowered), making it easier to be classified as a higher risk level under the same risk index, thereby enhancing risk sensitivity under flood season rainstorm conditions. After adjustment, the currently effective first classification threshold (probability level) and second classification threshold (consequence level) are obtained.

[0197] Since a dynamic adjustment coefficient greater than 1 indicates a need for more sensitive risk assessment, the baseline threshold is divided by this coefficient to lower the threshold, thereby obtaining a higher risk level under the same index. If the coefficient is less than 1 (e.g., during non-flood seasons), the threshold is multiplied by the coefficient to raise it and reduce sensitivity.

[0198] Probability level threshold correction:

[0199] Pre-correction baseline thresholds: P≥4.5→Rp=5, P≥3.5→Rp=4, P≥2.5→Rp=3, P≥1.5→Rp=2;

[0200] The revised effective threshold (the threshold is lowered to the base threshold ÷ 1.2, because a dynamic adjustment coefficient > 1 means more sensitive): P ≥ 3.5 → Rp = 5, P ≥ 2.5 → Rp = 4, P ≥ 1.5 → Rp = 3, P ≥ 0.5 → Rp = 2.

[0201] Consequence level threshold correction:

[0202] Pre-correction baseline thresholds: C≥4.0→Rc=5, C≥3.0→Rc=4, C≥2.0→Rc=3, C≥1.0→Rc=2;

[0203] The revised effective threshold (baseline threshold ÷ 1.2) is as follows: C≥3.0→Rc=5, C≥2.0→Rc=4, C≥1.0→Rc=3, C≥0.0→Rc=2.

[0204] Here, Rp represents the probability level and Rc represents the consequence level. Both are integers from 1 to 5. The larger the Rp value, the higher the probability of asset failure, and the larger the Rc value, the more severe the consequences of failure.

[0205] In this embodiment, the correction of the first and second level thresholds is not performed independently, but rather through a linked correction mechanism. When the probability level threshold is lowered due to the flood season, the consequence level threshold is lowered simultaneously, but the magnitude of the reduction is differentiated based on historical data analysis results: the probability threshold reduction coefficient is 1.2, and the consequence threshold reduction coefficient is 1.1. This is because historical data shows that the flood season has a greater impact on the failure probability than on the failure consequences. This linked mechanism, based on data analysis of over 2000 historical events, can more accurately match the risk characteristics under different operating conditions.

[0206] S52. Level Determination and Risk Value Calculation:

[0207] The failure probability index is compared with the currently effective first grading threshold and converted into a probability level Rp (an integer ranging from 1 to 5); the failure consequence index is compared with the second grading threshold level by level and converted into a consequence level Rc (an integer ranging from 1 to 5).

[0208] Calculate the risk value RV = Rp × Rc (an integer ranging from 1 to 25), and determine the real-time risk level based on the preset value range of RV.

[0209] The preset value range is shown below:

[0210] RV∈[1,2] indicates low risk;

[0211] RV∈[3,4] represents a lower risk;

[0212] RV∈[5,9] represents medium risk;

[0213] RV∈[10,16] indicates a higher risk;

[0214] RV∈[20,25] indicates high risk.

[0215] J01: Failure probability index P=3.4. Compared with the corrected probability threshold (P=3.4 is in the interval P≥2.5 and P<3.5), the corresponding probability level Rp=4; Failure consequence index C=4.2. Compared with the corrected consequence threshold (C=4.2 is in the interval C≥3.0 and C<4.0), the corresponding consequence level Rc=5; RV=4×5=20, which belongs to the high-risk interval, so it is judged as high risk.

[0216] J05: Failure probability index P=2.1, compared with the corrected probability threshold (P=2.1 is in the interval P≥1.5 and P<2.5), the corresponding probability level Rp=3; Failure consequence index C=3.8, compared with the corrected consequence threshold (C=3.8 is in the interval C≥3.0 and C<4.0), the corresponding consequence level Rc=4; RV=3×4=12, which belongs to the higher risk interval, so it is judged as higher risk.

[0217] S53. Based on the real-time risk level, dynamically adjust the data sampling frequency of the sensors associated with each asset.

[0218] When the risk level is high, the sampling frequency will be increased from 5 minutes / time to 1 minute / time, and the surrounding backup sensors (such as water level gauges J02 and J03) will be activated to form a localized encrypted monitoring network. For high-risk assets, the maintenance frequency and standards will be increased. When the risk level is relatively high, the sampling frequency will be increased from 5 minutes / time to 3 minutes / time. When the risk level is below the preset threshold, the sampling frequency will be reduced or the sensors will be switched to sleep mode.

[0219] Through the aforementioned dynamic threshold correction mechanism, the system lowers the risk level determination threshold under the flood season and heavy rainfall conditions, so that J01 is determined to be high risk when P=3.4, while if a fixed threshold is used, it can only be determined to be relatively high risk. This achieves adaptive adjustment of risk assessment to real-time hydrological conditions and improves the timeliness and accuracy of high risk identification.

[0220] J01 (High Risk): Increase the sampling frequency of associated sensors from 5 minutes / time to 1 minute / time; start the standby unit of the downstream pumping station; generate a preventive maintenance work order and push it to the operation and maintenance personnel (requiring on-site investigation within 2 hours); send high-level warnings to nearby primary schools and street offices; generate a risk propagation heat map based on the graph traversal results, dynamically highlight the high-risk propagation path (J01→P12→J05) on the map, mark the direction of pollution spread with arrows, and interactively display the cascading impact range (10 upstream flooded nodes and 25 downstream polluted nodes).

[0221] J05 (Higher Risk): Increase the sampling frequency from 5 minutes / time to 3 minutes / time and send a medium alert.

[0222] In addition, when the real-time risk level is lower than a preset threshold, the system automatically reduces the sampling frequency or switches the sensor to sleep mode to save energy and communication resources; when there are backup sensors around high-risk assets, the system automatically activates these backup sensors to form a local encrypted monitoring network, further increasing the monitoring density.

[0223] S54. Match control instructions from the preset multi-objective optimization strategy library.

[0224] Before matching control commands, the system first predicts the execution effect of each command combination. Taking J01 as an example, the system uses a performance prediction model (lightweight regression model) trained on historical data to estimate the impact trend of different command combinations on fault risk: simply increasing the sampling frequency can reduce the risk to a certain extent; simultaneously activating the backup pump can further reduce the risk; increasing maintenance work orders and early warnings can minimize the risk, but the corresponding cost is also higher. The system selects the matching command combination based on user preferences (such as prioritizing risk reduction or cost control) and generates an expected effect report to assist maintenance personnel in decision-making.

[0225] In this embodiment, the user preference is set to "prioritize risk reduction." Therefore, the system selects to execute all instructions on J01 (highest risk reduction, but also highest cost) and execute medium instructions on J05 (balancing cost and risk reduction) from the Pareto front of the strategy library. The strategy library pre-sets a variety of instruction combinations, each labeled with the expected risk reduction effect, cost, energy consumption, and social impact. The system dynamically selects the optimal combination based on the real-time risk level and user preference.

[0226] The strategy library pre-sets several instruction combinations: Strategy A (only increases sampling frequency, low cost, moderate risk reduction); Strategy B (increases sampling frequency + early warning, medium cost, moderate risk reduction); Strategy C (increases sampling frequency + starts backup pump + work order + early warning, high cost, high risk reduction). The system selects the optimal combination based on the real-time risk level and user preferences (prioritizing risk reduction or energy saving).

[0227] Optionally, after S5, there is also a step of continuously collecting feedback data and continuously optimizing the model based on the feedback data.

[0228] Collect feedback data after the execution of the differentiated closed-loop control instructions, including actual fault occurrence and actual fault loss data;

[0229] Using the Bayesian update algorithm, the weight parameters in the pre-trained physical augmentation graph neural network model are corrected in reverse based on the feedback data;

[0230] When the accumulated amount of feedback data reaches a preset threshold, an incremental learning mechanism is triggered, which updates only the network layer parameters near the output end in the pre-trained physical augmentation graph neural network model, while keeping the network layer parameters near the input end unchanged, in order to prevent catastrophic forgetting.

[0231] One week later, a minor leak occurred in pipe section J01, and the maintenance record was fed back into the system. Feedback data after the execution of the differentiated closed-loop control instructions was collected, including the actual fault occurrence and actual fault loss data (repair cost approximately 20,000 yuan, with no social impact).

[0232] (1) Using the Bayesian update algorithm, the weight parameters in the pre-trained physical augmentation graph neural network model are corrected in reverse based on the feedback data. With the Beta distribution Beta(2,2) as the prior, after observing a "fault", the posterior is Beta(3,2), and the weight parameters are corrected from (0.5,0.5) to (0.6,0.4), that is, the model believes that the contribution of the real-time state correction factor should be relatively increased.

[0233] (2) When the accumulated amount of feedback data reaches a preset threshold (e.g., 500 data points), the incremental learning mechanism is triggered:

[0234] Incremental learning only updates the parameters of the last two network layers closest to the output in the pre-trained physical augmentation graph neural network model (these layers learn task-specific mappings and need to adapt to new data), while keeping the parameters of the first three network layers closest to the input unchanged (these layers learn the general spatiotemporal characteristics of hydraulic signals and are relatively stable). This prevents catastrophic forgetting while adapting to new data and maintains the ability to remember historical operating conditions.

[0235] (3) Introduce long-term change adaptation mechanisms;

[0236] To address potential long-term changes in urban water systems (such as aging pipe networks, new water storage facilities, and climate change), this embodiment also introduces a model update triggering mechanism based on cumulative statistics of feedback data. The system continuously monitors the changing trends of risk distribution in the feedback data: when the proportion of high-risk assets deviates from the historical benchmark by more than a preset threshold (e.g., 20%) within three consecutive months, or when the average risk level deviates by more than 0.5 levels, incremental learning or model reconstruction is automatically triggered. This mechanism enables the model to automatically adapt to long-term evolution such as system upgrades, aging pipe networks, and climate change without human intervention.

[0237] This embodiment constructs a complete closed-loop technical solution through five interconnected technical steps: First, by acquiring real-time hydraulic data from monitoring devices located at each asset and combining it with propagation delay for physical guidance of spatiotemporal alignment, it solves the problem of data spatiotemporal misalignment caused by neglecting the propagation time of hydraulic signals in traditional methods, ensuring that the input data accurately corresponds in terms of physical causality, laying a reliable foundation for subsequent evaluation; Second, it integrates the aligned feature data with static physical attributes and spatial geographic information to construct a dynamic heterogeneous graph, uniformly representing the pipeline topology, asset type, spatial location, and real-time status, solving the technical challenge of collaborative modeling of multi-source heterogeneous information; On this basis, the dynamic heterogeneous graph is input into a pre-trained physical augmentation graph neural network model, and through a graph with physical causality mask constraints... The attention mechanism forcibly blocks non-physically related node connections, ensuring that feature aggregation conforms to the laws of hydraulic propagation. Through a multi-task learning architecture, it synchronously outputs failure probability and fault consequence indices, achieving a deep integration of physical laws and data-driven approaches. This ensures both model interpretability and computational efficiency. Subsequently, the risk classification threshold is dynamically adjusted based on current hydrological conditions, enabling risk assessment criteria to match different conditions such as flood season and heavy rain in real time, thus addressing the insufficient adaptability of fixed thresholds. Finally, based on real-time risk levels, control instructions are matched from a multi-objective optimization strategy library and executed in a closed-loop manner, achieving an automated closed loop from assessment to decision-making. This supports multi-dimensional optimization of cost, energy consumption, and risk, significantly improving the intelligence level of urban water system asset risk management and resource utilization efficiency.

[0238] Furthermore, this invention also ensures rapid response capabilities by directly triggering emergency response under extreme conditions through preset risk red line conditions; it achieves weight fine-tuning through Bayesian updates and updates only high-level network layers through incremental learning, retaining historical features while adapting to new data, avoiding catastrophic forgetting, and enabling the model to continuously adapt to system transformation and climate change, thus solving the technical defect of model solidification and inability to evolve.

[0239] Example 4

[0240] Finally, this application also proposes a computing device including a processor and a memory. The memory stores a computer program, and the processor executes the instructions stored in the memory so that the computer device performs the urban water asset dynamic risk assessment method based on physical augmented graph networks described in the above embodiments.

[0241] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0242] Furthermore, it should be noted that in the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0243] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0244] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for dynamic risk assessment of urban water assets based on physical augmented graph networks, characterized in that, include: S1. Acquire real-time hydraulic monitoring data, static physical attribute data, spatial geographic information data, current hydrological conditions factors, and pipeline topology connections of each asset in the target city's water system; wherein, the real-time hydraulic monitoring data is obtained from monitoring equipment installed at each asset, and the locations of the monitoring equipment constitute monitoring nodes; S2. Based on the real-time hydraulic monitoring data and the pipeline network topology, calculate the propagation delay of the hydraulic signal reflected in the real-time hydraulic monitoring data in the pipeline network; perform spatiotemporal alignment processing on the real-time hydraulic monitoring data based on the propagation delay to obtain a spatiotemporally aligned feature dataset. S3. Perform multi-source feature encoding and fusion on the spatiotemporally aligned feature dataset, the static physical attribute data, and the spatial geographic information data to obtain a dynamic heterogeneous map; S4. Input the dynamic heterogeneous graph into a pre-trained physical augmented graph neural network model, and output the failure probability index and failure consequence index of each asset through a graph attention mechanism with physical causal mask constraints. The physical augmentation graph neural network model adopts a multi-task learning architecture, which includes a shared spatiotemporal feature extraction backbone network, as well as a parallel failure probability prediction head and a failure consequence assessment head; wherein the failure probability prediction head is used to output the failure probability index, and the failure consequence assessment head is used to output the failure consequence index. S5. Determine the risk classification threshold based on the current hydrological conditions; determine the level of the failure probability index and the failure consequence index based on the risk classification threshold to obtain the real-time risk level of each asset; based on the real-time risk level, match control instructions from the preset multi-objective optimization strategy library and execute closed-loop control operations.

2. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, The dynamic heterogeneous graph includes: The system includes various types of graph nodes, such as nodes for monitoring inspection wells, nodes for monitoring pump stations, and nodes for monitoring pipe sections. Each graph node corresponds one-to-one with the monitoring equipment. Each graph node contains the static attribute characteristics, spatial location characteristics, and real-time status characteristics of its corresponding asset. Multiple types of edges are included, such as hydraulic connection edges and control edges; each edge contains its corresponding asset topology characteristics and hydraulic parameter characteristics.

3. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, Before S1, it also includes: S01. Obtain historical monitoring data and corresponding historical fault labels. Based on the pipeline network topology connection relationship, use the hydraulic propagation model to perform physical alignment and physical constraint repair on the historical monitoring data to generate a historical physical causal alignment sample set. Simultaneously, extreme operating condition samples are extracted from the historical physical causal alignment sample set; an adversarial generative network is trained based on the extreme operating condition samples to generate synthetic data under extreme operating conditions; the synthetic data is merged into the historical physical causal alignment sample set to obtain an expanded historical physical causal alignment sample set. S02. Iteratively train the initial physics-enhanced graph neural network model based on the expanded historical physics causal alignment sample set; The initial physical augmentation graph neural network model includes a graph attention layer and a temporal convolutional layer; the graph attention layer is used to learn the hydraulic dependence between upstream and downstream nodes in the pipeline network and assign attention weights to the feature aggregation between different nodes; the temporal convolutional layer is used to extract the temporal change features of each monitoring node within the time window; During training, the hybrid loss function is minimized through backpropagation until it converges, thus obtaining the pre-trained physical augmentation graph neural network model. The hybrid loss function is composed of a weighted average of a data-driven loss term and a physical constraint loss term from two task heads. The two task heads are a failure probability prediction head corresponding to the failure probability index and a failure consequence assessment head corresponding to the failure consequence index. The data-driven loss term is used to minimize the error between the failure probability index and failure consequence index predicted by the initial physical augmentation graph neural network model and the historical failure labels; the physical constraint loss term is used to calculate the residual value of the output result of the initial physical augmentation graph neural network model that violates the hydraulic conservation equation, and to penalize the residual value to constrain the output of the initial physical augmentation graph neural network model to conform to the physical mechanism; wherein, the physical constraint loss term includes a mass conservation residual term and an energy conservation residual term.

4. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, In step S2, the real-time hydraulic monitoring data is spatiotemporally aligned based on the propagation delay, including: Based on the pipeline network topology and the physical attributes of each pipeline extracted from static physical attribute data, the location and connection relationship of each monitoring node are determined. The theoretical propagation delay of the hydraulic state signal between any two monitoring nodes is calculated using a hydraulic propagation model. The theoretical propagation delay is determined based on the ratio of the pipe section length to the average flow velocity under the current operating conditions. The average flow velocity is calculated using the Manning formula based on real-time hydraulic monitoring data. Using the current sampling time of the downstream monitoring node as a reference, the real-time hydraulic monitoring data of the upstream monitoring node is shifted backward on the time axis using the theoretical propagation delay, so that the upstream and downstream monitoring data are aligned in physical time.

5. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, S3 includes: The time window length is selected based on the current hydrological conditions factors, including rainfall intensity level and river level. When the rainfall intensity level is higher than the preset threshold or the river water level is higher than the warning level, the time window length is shortened; when the rainfall intensity level is lower than the preset threshold and the river water level is lower than the warning level, the time window length is extended. Within the time window, the real-time status feature sequence of each monitoring node is extracted and used as the real-time status feature of the corresponding node in the dynamic heterogeneous graph.

6. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, In S4, the graph attention mechanism with physical causal mask constraints includes: The physical causal mask constraint, based on the water flow direction determined by the pipeline topology connection relationship and the physical reachability determined by the comparison result of the propagation delay and the preset delay threshold, masks non-physically related node connections when calculating attention weights.

7. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, S5 includes: A dynamic adjustment coefficient is generated based on the current hydrological conditions factors. Using the dynamic adjustment coefficient, the preset risk classification threshold is corrected in real time to obtain the currently effective first classification threshold corresponding to the failure probability index and the second classification threshold corresponding to the failure consequence index. The failure probability index is compared with the first grading threshold and converted into a probability level; the failure consequence index is compared with the second grading threshold and converted into a consequence level. The real-time risk level of each asset is determined based on the probability level and consequence level.

8. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, In step S5, based on the real-time risk level, control instructions are matched from a preset multi-objective optimization strategy library, and closed-loop control operations are executed, including: Based on the real-time risk level, the data sampling frequency of the sensors associated with each asset is dynamically adjusted: when the real-time risk level is higher than a preset threshold, the sampling frequency is increased; when the real-time risk level is lower than the preset threshold, the sampling frequency is decreased or the sensor is switched to sleep mode. And / or, based on the real-time risk level, dynamically activate or put into sleep mode the backup sensors: when the real-time risk level is higher than a preset threshold, activate the surrounding associated backup sensors to form a local encrypted monitoring network; when the real-time risk level is lower than the preset threshold, switch the backup sensors to sleep mode.

9. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, Before S2, it also includes: The preset risk red line conditions include at least one of the following: pipe diameter does not meet the design standard, pipe section structure fails, pipe section blockage causes water flow cross-sectional area loss to exceed 50%, and pipe section overflows within the design standard range. When an asset in the target city's water system triggers the aforementioned risk red line conditions, the asset is directly deemed invalid, and an emergency control order is executed.

10. The method for dynamic risk assessment of urban water assets based on physically augmented graph networks according to claim 1, characterized in that, Following S5 are: Collect feedback data after the execution of the differentiated closed-loop control instructions, including actual fault occurrence and actual fault loss data; Using the Bayesian update algorithm, the weight parameters in the pre-trained physical augmentation graph neural network model are corrected in reverse based on the feedback data; When the accumulated amount of feedback data reaches a preset threshold, an incremental learning mechanism is triggered, which updates only the network layer parameters near the output end in the pre-trained physical augmentation graph neural network model, while keeping the network layer parameters near the input end unchanged, in order to prevent catastrophic forgetting.

Images