A method and system for cascading failure deduction of water supply network across modal causal coupling

Abstract

This invention discloses a method and system for inferring cascading failures in water supply networks based on cross-modal causal coupling, relating to the field of safety control in urban water supply networks. The method first constructs a heterogeneous graph of the water supply network, abstracting physical pipe segments as core nodes and identifying heterogeneous edges between segments, obtaining their static and dynamic attributes. Then, a dedicated scale-adaptive coding layer completes the spatiotemporal alignment of multi-source heterogeneous data, generating a standardized feature matrix. Using a cross-modal causal directed acyclic graph as a hard constraint, a dynamic coupling adjacency matrix is ​​generated through a cross-modal graph attention layer. A spatiotemporal dynamic gating unit is used to achieve dynamic weighted coupling of multimodal features. Finally, based on the coupling features and the adjacency matrix, the cascading failure evolution is simulated, critical inflection points are identified, and graded early warning signals are output. This invention solves the problems of modeling distortion, false associations, and lack of inflection point prediction in traditional techniques, significantly improving the inference accuracy and engineering reliability, and is applicable to safety control and emergency early warning in urban water supply networks.

Description

Technical Field

[0001] This invention belongs to the fields of urban water supply network safety control and smart water technology, specifically involving a method for inferring cascade failures, identifying critical inflection points, and providing hierarchical early warning of water supply networks based on heterogeneous graph modeling, cross-modal causal coupling, and spatiotemporal graph convolution. Background Technology

[0002] The ongoing urbanization process has led to the continuous expansion of urban water supply networks and their increasing service life, resulting in frequent pipe bursts and leaks. A single point-like pipe burst can easily trigger a chain reaction through hydraulic redistribution in the network, leading to widespread cascading failures, causing large-scale water outages, water pollution, and other accidents, seriously threatening urban water supply security and the stability of people's livelihoods.

[0003] Existing water supply network safety analysis technologies have many core defects and can no longer meet the actual engineering needs of urban water supply network safety control. Specifically, these defects include: 1) Traditional graph modeling, which uses water sources, pumping stations, and users as nodes and pipe segments as edges, reduces the main pipe segment where failure occurs to connecting edges, failing to accurately depict the failure propagation correlation between pipe segments, resulting in low failure prediction accuracy and large deviation from actual operating conditions; 2) Due to significant differences in the spatiotemporal granularity of data from hydraulics, structural health, spatial geography, and operation and maintenance, existing simple splicing and uniform interpolation methods easily lose key information such as long-term degradation of pipe segments and the impact of discrete events, resulting in severe fusion characteristics. 3) Graph neural network-based analysis models are mostly pure data-driven black-box models without physical constraints, which are prone to generating false correlations. The prediction results often violate the basic principles of hydraulics and materials science. The few schemes that incorporate physical constraints only use physical formulas as regularization terms of loss functions, which cannot eliminate the problem at its root and have extremely low engineering credibility. 4) Existing technologies mostly remain at the level of ex-post or real-time assessment of the risk level and steady-state stability of the current node of the pipeline network. They cannot predict in advance the critical inflection point of the system from local controllability to global collapse in cascading failure, making it difficult to grasp the golden window period for emergency response and failing to meet the actual needs of water supply network safety control.

[0004] In summary, existing technologies cannot solve the core problems of accurate fusion of multi-source data, hard constraints of physical laws, accurate prediction of dynamic evolution, and early warning of critical inflection points in cascading failure simulation of water supply networks. There is an urgent need to build a complete, highly reliable, and implementable cascading failure simulation technology solution. Summary of the Invention

[0005] To address the aforementioned deficiencies in existing technologies, the present invention aims to provide a cross-modal causal coupling method and system for cascading failure prediction of water supply networks, thereby solving the problems of traditional modeling distortion, multi-source data fusion distortion, false model correlation, and lack of critical inflection point prediction capability.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] A method for extrapolating cascading failures in a cross-modal causal coupling water supply network includes the following steps:

[0008] Step S1: Construct the heterogeneous graph structure of the water supply network: This involves connecting the physical pipe segments in the water supply network... P ij Abstracted as the core nodes of a graph structure V ij The physical connection relationship, hydraulic transmission relationship, and spatial adjacency relationship between pipe segments are abstracted into heterogeneous edges connecting adjacent core nodes; the static inherent properties and dynamic operation properties of the core nodes, as well as the static association properties and dynamic coupling properties of the heterogeneous edges are obtained.

[0009] Step S2: Spatiotemporal Scale Adaptation of Multi-Source Heterogeneous Data: Acquire multimodal source data of the pipeline network, including high-frequency hydraulic time-series data, low-frequency structural health monitoring data, static spatial geographic data, discrete event-based operation and maintenance management data, and medium-frequency water quality safety time-series data, as well as external event-related data affecting pipeline network operation; for the spatiotemporal scale differences between each modal data and core nodes, design a dedicated scale adaptation coding layer for each modality to achieve accurate spatiotemporal alignment between multimodal data and heterogeneous graph structures, and generate standardized feature matrices corresponding to each modality; calculate the initial failure probability of the pipeline segment based on external event-related data, and integrate the initial failure probability into the standardized feature matrix of the corresponding modality;

[0010] Step S3: Cross-modal feature fusion with causal constraints: A cross-modal causal directed acyclic graph (DAG) specific to the water supply network is pre-constructed to clarify the physical causal transmission paths between data from different modalities that conform to the principles of hydraulics, materials science, and geological engineering. Using the DAG as a hard constraint, a causal path-aware cross-modal graph attention layer is designed, allowing only features to be transmitted and aggregated along compliant causal paths, generating a dynamic coupling adjacency matrix A that characterizes the real-time coupling and correlation strength between core nodes. coup ;

[0011] Step S4: Node-level spatiotemporal dynamic feature coupling: Design spatiotemporal dynamic gating units at the core node level of the pipe segment. Based on the inherent properties of the pipe segment, real-time hydraulic conditions, and external event characteristics, adaptively calculate the contribution weight of each modal data at each core node and each time step; add a modal redundancy suppression term to attenuate the weight of modes with feature overlap exceeding the threshold, and complete the dynamic weighted coupling of multimodal features through the gating mechanism to output the cross-modal coupled spatiotemporal feature vector of each core node of the water supply network;

[0012] Step S5: Cascade Failure Deduction and Critical Early Warning: Based on Coupled Cross-Modal Spatiotemporal Eigenvectors and Dynamically Coupled Adjacency Matrix A coupA convolutional layer for the spatiotemporal graph of cascading failure propagation is constructed to simulate the spatiotemporal evolution of failure events from the initial node to the entire pipeline network. The rate of change of the global feature gradient of the pipeline network is calculated through information manifold learning to identify the critical inflection point T between local controllability and global collapse in cascading failure. cri Based on the simulation results, a four-level early warning mechanism is set up to output the spatiotemporal evolution path of cascading failures, the range of future time-series failures, the probability of system collapse, and graded early warning signals.

[0013] Further optimization, in step S1, the core node V ij The static inherent properties are engineering design parameters that do not change with the operating state of the pipeline network, including pipe sections. P ij geographic coordinates (x ij ,y ij ), Design inner diameter D ij Physical length L ij Initial tensile strength σ of the pipe ij0 and the design maximum overcurrent capacity Q ij_des。

[0014] The dynamic operating attributes of the core node are time-series operating parameters obtained through real-time sensor acquisition and simulation calculation, including pipe sections. P ij The real-time flow Q at time t ij (t), Real-time flow rate v ij (t), head loss along the route h f-ij (t), Remaining wall thickness δ ij (t), uniform corrosion rate r ij_corr (t) and fatigue damage degree D ij_fat (t).

[0015] The pipe section P ij Real-time flow rate v ij The formula for calculating (t) is as follows:

[0016] ;

[0017] In the formula, D ij For pipe section P ij Design inner diameter, Q ij (t) represents pipe segment P at time t. ij Real-time traffic.

[0018] The pipe section P ij Head loss along the route h f-ij(t) is calculated using the Darcy-Wiesbach formula:

[0019] ;

[0020] In the formula, λ ij For pipe section P ij Friction coefficient (dimensionless), g is the acceleration due to gravity, L ij For pipe section P ij length.

[0021] The static association attributes of the heterogeneous edge include pipe segments. P ij and P mn The original space Euclidean distance d ij-mn Design hydraulic correlation degree and affiliation with the same operation and maintenance unit;

[0022] The dynamic coupling properties of heterogeneous edges include pipe segments P ij and P mn Real-time traffic allocation ratio, pressure transmission coefficient, and failure correlation probability at time t.

[0023] Further optimization is achieved by designing the modality-specific scale adaptation coding layer in step S2 as follows:

[0024] 1) For high-frequency hydraulic time-series data with sampling intervals on the order of minutes, a time-aligned encoding layer is designed to map the data to the time step dimension of the corresponding core nodes in the heterogeneous graph, generating a hydraulic feature matrix F that matches the number of core nodes and the time step size. hyd。

[0025] 2) For low-frequency structural health data with a quarterly monitoring period, a cubic spline interpolation coding layer is designed to smoothly map the low-frequency monitoring data to a minute-level time step consistent with the hydraulic data, ultimately generating the structural health feature matrix F. str The interpolation formula is:

[0026] ;

[0027] In the formula, S ij (t) represents pipe segment P ij The interpolation result at time t, a k b k c k d k Let t be the spline coefficient for the k-th interpolation interval. k For pipe section P ij The kth detection time point.

[0028] 3) For discrete event-based operation and maintenance data, design an event-triggered coding layer to encode segment P. ij The discrete events are transformed into a feature weight matrix F for continuous time steps. ops。

[0029] 4) For static spatial geographic data, an inverse distance weighted spatial coding layer is designed. This layer sequentially processes the spatial distance between pipe segments through min-max normalization, calculates the initial spatial coding weights, and performs a second min-max normalization on the initial weights. This maps the static attributes of the pipe segments to the corresponding core nodes and uniformly maps them to the [0,1] interval, generating a spatial geographic feature matrix F. geo The calculation formula is:

[0030] ;

[0031] ;

[0032] ;

[0033] In the formula, w ijmn_nor For pipe section P ij With pipe section P mn The final normalized spatial encoding weights; w ijmn For pipe section P ij With pipe section P mn Initial spatial encoding weights; w min w represents the minimum initial spatial coding weight for the entire pipeline network segment. max d represents the maximum value of the initial spatial coding weight for the entire pipeline segment; μ is the distance attenuation coefficient (dimensionless) of the inverse distance weighted spatial coding, with a baseline value of 2 and a range of 1.5~3.0; ij-mn ′ represents a pipe section P ij With pipe section P mn The normalized space Euclidean distance, d ij-mn For pipe section P ij With pipe section P mn The original space Euclidean distance, d min d represents the minimum spatial distance between pipe sections in the entire pipeline network. max This represents the maximum spatial distance between sections of the entire pipeline network.

[0034] 5) For medium-frequency water quality safety data, with sampling intervals on the order of hours, a water flow direction alignment coding layer is designed to map water quality characteristics to corresponding pipe sections along the water supply path. P ijcore node V ij Generate the water quality safety feature matrix F wat .

[0035] Further optimization involves step S2, where the initial failure probability of the pipe segment is calculated and incorporated into the corresponding feature matrix for four types of external events: earthquake, rainstorm, freeze-thaw, and third-party construction. The specific calculation formula is as follows:

[0036] 1) Earthquake event: Initial failure probability P of the pipe segment fail_seis The formula for calculating (ij,t) is:

[0037] ;

[0038] In the formula, k seis Let a be the seismic attenuation coefficient. max For peak ground acceleration, σ ij (t) represents the pipe segment P ij Current remaining tensile strength, σ ij0 For pipe section P ij Initial tensile strength.

[0039] 2) Rainstorm event: Initial failure probability P of the pipeline segment fail_rain The formula for calculating (ij,t) is:

[0040] ;

[0041] In the formula, R real (t) represents the real-time rainfall at time t, R des (ij) represents the pipe segment P ij The designed flood control rainfall for the area is k rain h represents the impact coefficient of heavy rainfall. ij For pipe section P ij Burial depth, h min This refers to the minimum safe burial depth for water supply pipeline sections.

[0042] 3) Freeze-thaw events: First, calculate the pipe segment damage using the Miner linear cumulative damage criterion. P ij Fatigue damage degree D ij_fat (t), then calculate the failure probability P fail_freeze (ij,t), the calculation formula is:

[0043] ;

[0044] In the formula, e r (ij) represents the pipe segment Pij The actual number of cycles experienced under the r-th stress, E r (ij) represents the pipe segment P ij The fatigue life corresponding to the r-th stress, k freeze is the freeze-thaw effect coefficient; E is the total number of stress cycles.

[0045] 4) Third-party construction events: Initial failure probability P of the pipe segment fail_cons The formula for calculating (ij,t) is:

[0046] ;

[0047] In the formula, k cons d is the construction impact factor. cons (ij,t) represents the construction point and pipe section at time t. P ij The distance is d0, where d0 is the safe distance threshold.

[0048] Further optimization is achieved by defining the constraint mechanism and dynamic coupling adjacency matrix generation rules for the causal directed acyclic graph (DAG) in step S3 as follows:

[0049] 1) The compliant one-way causal transmission path is:

[0050] Spatial geographic mode → Structural health mode → Hydraulic operating condition mode → Pipeline segment failure; Structural health mode → Hydraulic operating condition mode → Water quality safety mode; Operation and maintenance event mode → Hydraulic operating condition mode → Secondary failure risk of pipeline segment.

[0051] 2) Set up an independent attention head for each compliant causal path, and directly reset the weights of features passed to non-compliant paths to zero; add a causal consistency penalty term L to the model's total loss function. causal The formula is:

[0052] ;

[0053] In the formula, N represents the total number of core nodes, and M represents the total number of non-compliant causal paths. For core node V ij In the p Attention weights on non-compliant paths.

[0054] 3) Dynamically Coupled Adjacency Matrix A coup The calculation formula is:

[0055] ;

[0056] In the formula, Θ is the learnable parameter tensor, and F mergerepresents the multimodal feature matrix after causal constraint fusion, and Softmax and ReLU are nonlinear activation functions.

[0057] Further optimization is achieved by defining the contribution weight calculation and feature coupling rules for the spatiotemporal dynamic gating unit in step S4 as follows:

[0058] 1) For each core node and each time step, the initial contribution weight W of each modality. q The formula for calculating (ij,t) is:

[0059] ;

[0060] In the formula, W q (ij,t) represents the core node of the q-th mode at time t. V ij The initial contribution weights at point X, σ is the sigmoid activation function, MLP is a multilayer perceptron, and X ij (t) represents the core node at time t. V ij The inherent properties of the pipe section, A ij (t) represents the core node at time t. V ij Real-time hydraulic operating condition characteristics, E ij (t) represents the core node at time t. V ij External event characteristics.

[0061] 2) Calculate the feature cosine similarity between modes |cos(F) q1 ,F q2 )∣,when∣cos(F q1 ,F q2 When |>0.8, the redundancy suppression term L is used. red Weight decay is applied to highly redundant mode pairs; the formula for calculating the modal redundancy suppression term is:

[0062] ;

[0063] In the formula, L red This is a modal redundancy suppression term, used to reduce the weights of modes whose feature overlap exceeds a threshold; cos(F q1 ,F q2 W represents the cosine similarity between the modal features of classes q1 and q2; q1 W q2 F represents the initial contribution weights of the modes corresponding to q1 and q2, respectively; q1 and F q2 These represent the standardized feature matrices for different modes q1 and q2, respectively.

[0064] 3) Core Nodes V ij Cross-modal coupled spatiotemporal eigenvector H ij The formula for calculating (t) is:

[0065] ;

[0066] In the formula, F q (ij,t) represents the core node of the q-th mode at time t. V ij Standardization features of the location; W q (ij,t) represents the core node of the q-th mode at time t. V ij The contribution weight of each location.

[0067] Further optimization is achieved by using the following formula for iteratively calculating the failure propagation probability of the cascaded failure propagation spatiotemporal graph convolutional layer in step S5:

[0068] ;

[0069] In the formula, P casc (ij, t+1) is the core node V ij The probability of cascading failure propagation at time t+1, N( V ij ) as the core node V ij The set of adjacent nodes, V kl As an adjacency to the core node, A coup (ij,kl) is the core node V ij and V kl The coupling correlation strength, H kl (t) represents the core node at time t. V kl The coupling characteristics of W casc With b casc Here, σ represents the learnable weight matrix and the bias term, respectively, and σ is the sigmoid activation function.

[0070] Set the cascading failure trigger threshold P th When P casc (ij,t+1)≥P th At that time, determine the core node. V ij corresponding pipe section P ij When a secondary failure occurs, the heterogeneous graph structure and the dynamic coupling adjacency matrix are updated synchronously, and the full-chain evolution of cascade failure is simulated iteratively.

[0071] Further optimization is achieved by implementing the following rules for identifying cascading failure critical inflection points and providing graded early warnings in step S5:

[0072] 1) The formula for calculating the global characteristic gradient change rate C(t) of the pipeline network is:

[0073] ;

[0074] In the formula, C(t) is the gradient change rate of the global characteristics of the pipeline network at time t. Core node at time t V ij The gradient of the cross-modal coupling feature, where N is the total number of core nodes;

[0075] When C(t) exceeds the preset mutation threshold within three consecutive time steps, the time corresponding to the first step of the three consecutive time steps is determined to be the critical inflection point T of cascade failure. cri。

[0076] 2) The triggering conditions for the Level 4 early warning mechanism are:

[0077] Blue alert: No initial failure event, all core nodes V ij Cascade failure propagation probability P casc (ij,t)<0.2, C(t) is below the preset mutation threshold, and there is no risk of global collapse;

[0078] Yellow alert: An initial failure event has occurred, the failure scope is limited to the independent metering area DMA of a single water supply network, C(t) is lower than the preset mutation threshold, and there is no critical inflection point trigger signal;

[0079] Orange alert: The failure range is about to span two or more DMA partitions, the probability of failure propagation in adjacent control segments is ≥0.4 and less than the failure trigger threshold P. th Predict the remaining critical time T res (t) equals 1 time step, which poses a potential risk of global collapse;

[0080] Red alert: The failure area spans two or more DMA partitions and has reached the critical inflection point T. cri Remaining critical time T res (t)=0, the system is about to enter a global crash state.

[0081] A cross-modal causal coupling cascade failure prediction system for water supply networks, used to execute the above method, includes:

[0082] The heterogeneous graph construction module is used to abstract water supply network segments into core nodes of a heterogeneous graph and the relationships between segments into heterogeneous edges, construct the heterogeneous graph structure, and obtain and store the full-dimensional attribute data of core nodes and heterogeneous edges.

[0083] The multi-source data adaptation module is used to acquire multimodal source data of the pipeline network, design a dedicated scale adaptation coding layer for each mode, complete the spatiotemporal alignment of multimodal data and generate standardized feature matrices; calculate the initial failure probability of the pipeline segment for four types of external events and incorporate them into the corresponding feature matrices;

[0084] The causal constraint fusion module is used to pre-construct a cross-modal causal directed acyclic graph (DAG) of the water supply network, design a causal path-aware cross-modal graph attention layer, and generate a dynamically coupled adjacency matrix.

[0085] The dynamic feature coupling module is used to design the spatiotemporal dynamic gating unit at the node level of the pipe segment. It adaptively calculates the contribution weight of each mode and adds a mode redundancy suppression term to complete the dynamic weighted coupling of multimodal features and outputs the spatiotemporal feature vector of node cross-modal coupling.

[0086] The cascaded failure simulation and early warning module is used to construct a convolutional layer of the spatiotemporal graph of cascaded failure propagation, simulate the spatiotemporal evolution law of failure, identify the critical inflection point of system collapse through information manifold learning, and output the cascaded failure simulation results and graded early warning signals.

[0087] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0088] Compared with the prior art, the present invention has the following beneficial effects:

[0089] 1. This invention abstracts physical pipe segments as core nodes of a heterogeneous graph and the relationships between pipe segments as heterogeneous edges, constructing a heterogeneous graph structure specifically for water supply networks. This breaks through the limitations of traditional modeling that reduces pipe segments to edges, accurately depicts the failure propagation and correlation characteristics between pipe segments, and improves the basic accuracy of cascading failure inference from the root of modeling.

[0090] 2. To address the spatiotemporal scale differences of multi-source heterogeneous data, a dedicated scale-adaptive coding layer is designed for each modality. Combined with techniques such as cubic spline interpolation and inverse distance weighted spatial coding, the data achieves accurate spatiotemporal alignment. At the same time, the initial failure probability of pipe segments in external events is quantified to avoid feature distortion caused by simple splicing and uniform interpolation, and key information such as long-term degradation of pipe segments and the impact of discrete events is fully preserved.

[0091] 3. Based on the fundamental principles of hydraulics and materials science, a cross-modal causal directed acyclic graph is constructed and used as a hard constraint. A causal path-aware attention layer for the cross-modal graph is designed. The weights of non-compliant paths are cleared to zero and a causal consistency penalty term is added. This eliminates the generation of false associations from the model structure, makes the model output results conform to the laws of engineering physics, and greatly improves the engineering credibility of the model.

[0092] 4. By designing spatiotemporal dynamic gating units at the node and time step levels, adaptive adjustment of modal weights is achieved, and modal redundancy suppression terms are added to avoid feature redundancy, thus solving the weight distortion problem caused by existing global fusion and realizing accurate and dynamic coupling of multimodal features.

[0093] 5. By constructing a convolutional layer of the spatiotemporal graph of cascading failure propagation, iterative deduction of the entire chain of cascading failure is realized. By calculating the gradient change rate of global features of the pipeline network through information manifold learning, the critical inflection point of the system from local controllability to global collapse is accurately identified, breaking through the limitations of existing technologies that can only be evaluated in real time after the fact, and reserving a golden window period for emergency response.

[0094] 6. Based on the failure range and critical inflection point prediction results, a four-level early warning mechanism is set up to output the spatiotemporal evolution path of cascading failures, the future time-series failure range, the system collapse probability and graded early warning signals, providing quantitative basis for the safety control, emergency dispatch and fault handling of water supply networks, and realizing the progress from "passive post-event repair" to "proactive and forward prevention". Attached Figure Description

[0095] Figure 1 This is a simulation diagram of the pipeline topology and spatial distribution of rainstorm events in an embodiment of the present invention;

[0096] Figure 2 A flowchart of a cross-modal causal coupling cascade failure prediction method for water supply networks;

[0097] Figure 3 This is the final normalized spatial coding weight heatmap between pipe segments in this invention;

[0098] Figure 4 This is a simulation diagram of the feature transfer attention weight distribution under causal constraints according to the present invention;

[0099] Figure 5 This is a simulation diagram of the time-series evolution of the failure probability of pipe segments in the cascade failure process of this invention;

[0100] Figure 6 This is a simulation result diagram of the global characteristic gradient change rate of the pipeline network according to the present invention. Detailed Implementation

[0101] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0102] Example 1: This example uses a rainstorm event scenario in a town's water supply network as an application case to provide a detailed explanation of the cross-modal causal coupling method for cascading failures in water supply networks according to the present invention. This example is only used to explain the present invention and is not intended to limit the present invention.

[0103] In this embodiment, the water supply network includes 4 physical pipe segments, which are abstracted as 4 core nodes. The specific parameters are shown in Table 1 below. Figure 1 As shown.

[0104] Table 1. Corresponding parameters of pipe sections and core nodes

[0105] ;

[0106] In this embodiment, a rainstorm is used as an external event for simulation. The specific parameters of the rainstorm event are as follows:

[0107] Rainfall period: 6th-14th hour of the simulation cycle (05:00-13:00), with the peak time being the 9th hour;

[0108] Peak real-time rainfall: Coordinates of the rainstorm center at 9 hours (100, 30), peak intensity R real (9h) = 82mm / h;

[0109] Design flood control rainfall for each pipe section area: R des (12) = 50 mm / h, R des (23) = 40 mm / h (P 23 Corresponding area), R des (24) = 35 mm / h, R des (34) = 35 mm / h;

[0110] Minimum safe burial depth: h min =1.2m.

[0111] like Figure 2 As shown, a method for extrapolating cascading failures in a cross-modal causal coupling water supply network includes the following steps:

[0112] Step S1: Construct the heterogeneous graph structure of the water supply network, and include the physical pipe segment P. ij Abstracted as the core node V of a heterogeneous graph ij The relationships between pipe segments are abstracted as heterogeneous edges, and the attribute data of core nodes and heterogeneous edges are obtained. Specifically, this includes:

[0113] Step S1.1: Core Node: V 12 V 23 V 24 V 34 ;

[0114] Heterogeneous edge construction: 1) Physically connected edges: V 12 -V 23 V 23 -V 24 V 23 -V 34 V 24 -V 34 1) The physical connectivity of the corresponding pipe sections through hydraulic nodes; 2) Hydraulic transmission edge: V 12 -V 23 V 23 -V 24 V 23 -V 34 V 24 -V 34 3) Spatial adjacent edge: V 23 -V 24 V 24 -V 34 The spatial geographical proximity of the corresponding pipe sections.

[0115] Step S1.2: Taking the 9th hour after the peak of the rainstorm as an example, the measured real-time flow rates of each pipe section are as follows: Q 12 (9h) = 0.13m 3 / s, Q 23 (9h) = 0.10 m³ / s, Q 24 (9h) = 0.07m 3 / s, Q 34 (9h) = 0.09m 3 / s, the core hydraulic parameters of each pipe section are calculated based on the following formula:

[0116] Among them, pipe section P 12 (Core Node V) 12 The corresponding parameter is: real-time flow rate. v 12 (9h) = 4 × 0.13 / (3.1416 × 0.4) 2 =0.52 / 0.5027≈1.034m / s, head loss along the friction path

[0117] h f-12 (9h) = 0.02 × 120 / 0.4 × 1.034 2 / (2×9.8)=6×1.069 / 19.6≈0.327m.

[0118] Pipe section P 23 (Core Node V) 23 The corresponding parameter is: real-time flow rate. v 23 (9h) = 4 × 0.10 / (3.1416 × 0.3)2 =0.4 / 0.2827≈1.415m / s, head loss along the friction path

[0119] h f-23 (9h) = 0.022 × 100 / 0.3 × 1.415 2 / (2×9.8)≈0.749m.

[0120] Pipe section P 24 (Core Node V) 24 The corresponding parameter is: real-time stream. v 24 (9h) = 1.426 m / s, head loss along the friction h f-24 (9h) = 0.672m.

[0121] Pipe section P 34 (Core Node V) 34 The corresponding parameter is: real-time stream. v 34 (9h) = 0.935 m / s, head loss along the friction h f-34 (9h) = 0.331m.

[0122] Step S2: Perform spatiotemporal scale adaptation on the multi-source heterogeneous data, generate standardized feature matrices for each modality, and calculate the pipe section under four types of external events. P ij The initial failure probability.

[0123] Step S2.1: Multimodal data encoding processing

[0124] 1) For high-frequency hydraulic time-series data: Through a time-aligned coding layer, minute-level hydraulic data of each pipe section is mapped to the time step of the core node, generating a hydraulic feature matrix F. hyd Dimension 4×288 (4 core nodes × 24h × 12 5min time steps).

[0125] 2) For low-frequency structural health data: The detection period is quarterly. A cubic spline interpolation encoding layer is designed to smoothly map the low-frequency detection data to a minute-level time step consistent with the hydraulic data, ultimately generating the structural health feature matrix F. str .like Figure 3 As shown.

[0126] The interpolation formula is: ;

[0127] In the formula, S ij (t) represents pipe segment P ij The interpolation result at time t, a k b k c k d kLet t be the spline coefficient for the k-th interpolation interval. k For pipe section P ij The kth detection time point.

[0128] pipe segment P 12 (Core Node V) 12 Taking pipe segment P as an example, let's illustrate this. 12 The quarterly corrosion rate test data are as follows: corrosion rate of 0.12 mm / year at 0h (January 1), corrosion rate of 0.13 mm / year at 2160h (March 31), and corrosion rate of 0.14 mm / year at 4320h (June 30).

[0129] Interval step size calculation: h0 = t1 - t0 = 2160h, h1 = t2 - t1 = 2160h

[0130] Natural boundary conditions: ,

[0131] In the formula, The interpolation function at the z-th detection time point t z The second derivative at point M z Let z be the bending moment corresponding to the z-th detection time point. Let 0 be the normal value of the natural boundary condition. z = 0, 1, 2, where z is the sequence number of the detection time point.

[0132] Establishment and solution of the three bending moment equations: For the scenario with three time points and two corresponding time intervals, the three bending moment equations are as follows:

[0133] ;

[0134] in: ;

[0135] In the formula, λ1 and μ1 are the interval coefficients of the three bending moment equations, h0 and h1 are the interval step sizes of adjacent detection time points, t0, t1, and t2 are the three time points for the structural health detection of the pipe section, y0, y1, and y2 are the detected values ​​of the corrosion rate of the pipe section corresponding to the three time points, and d1 is the constant term of the three bending moment equations.

[0136] Substituting the boundary condition M0=M2=0, we get: M1=0;

[0137] Solving for the spline coefficients of the first interpolation interval (k=1, t∈[0,2160h]):

[0138] Constant term: a1 = y0 = 0.12;

[0139] coefficient of the first term: ;

[0140] Quadratic coefficient: ;

[0141] Cubic coefficient: ;

[0142] Target time interpolation calculation: Target interpolation time t = 1080h, substituting into the formula, we get: S 12 (1080) = 0.12 + 4.63 × 10 -6 ×1080+0×1080 2 +0×1080 3 ≈0.125 mm / year.

[0143] Achieve smooth alignment of quarterly low-frequency data to hourly time steps, fully preserving the long-term degradation characteristics of pipeline sections.

[0144] 3) For static spatial geographic data: Calculate the original spatial Euclidean distance between each pipe segment: d 12-23 =50m, d 12-24 =98.49m, d 12-34 =108.17m, d 23-24 =36.06m, d 23-34 =44.72m, d 24-34 =56.57m. Extreme distance between pipe segments in the entire pipeline network: minimum distance d min =36.06m, maximum distance d max =108.17m.

[0145] pipe segment P 23 With P 24 For example, the original Euclidean distance d 23-24 =36.06m;

[0146] Step 1: Calculate the normalized distance d 23-24 = (36.06 - 30.06) / (108.17 - 36.06) = 0;

[0147] Step 2: Initial spatial weight calculation takes the maximum limit value of the entire pipeline network as 258.32;

[0148] Step 3: Secondary normalization calculation.

[0149] Initial weight extreme value of the entire pipeline network: w min =0.0025, w max =258.32, w 2324-norm =1.0.

[0150] Finally, pipe segment P was obtained. 23 With P 24 The spatial coding weight is 1.0, indicating that the two pipe sections have the highest spatial correlation and the highest probability of failure due to the same rainstorm event. For example... Figure 3 As shown.

[0151] 4) For discrete event-based maintenance data: Two maintenance events within the simulation cycle (the 8th maintenance event) will be included. 12 Repair, 16th hP 34 (Repair) is transformed into a feature weight matrix F through an event-triggered encoding layer. ops .

[0152] 5) For medium-frequency water quality safety data: Through a water flow direction alignment encoding layer, hourly water quality data is mapped to each core node along the water supply path, generating a water quality safety feature matrix F. wat .

[0153] Step S2.2: Calculation of the initial failure probability of the pipeline segment during a rainstorm event

[0154] Based on the formula for calculating the initial failure probability of a rainstorm event, the initial failure probability of each pipe segment in the 9th hour is calculated as follows:

[0155] ;

[0156] In the formula, R real (t) represents the real-time rainfall at time t, R des (ij) represents the pipe segment P ij The designed flood control rainfall for the area is k rain h represents the impact coefficient of heavy rainfall. ij For pipe section P ij Burial depth, h min This refers to the minimum safe burial depth for water supply pipeline sections;

[0157] In this embodiment, R real (t) = 82 mm / h, k rain =0.8, h min =1.2.

[0158] Pipe section P 12 (V) 12 ):P fail_rain (12,9h)=1-exp(-0.2785)≈1-0.7569=0.2431≈24.3%;

[0159] Pipe section P 23 (V) 23 ):P fail_rain (23,9h)=45.2%;

[0160] Pipe section P 24 (V) 24 ):P fail_rain (24,9h)=65.8%;

[0161] Pipe section P 34 (V) 34 ):P fail_rain (34,9h)=39.3%;

[0162] The aforementioned initial failure probabilities are mapped to the corresponding core nodes and incorporated into the structural health feature matrix F. str .

[0163] Step S3: Based on the cross-modal causal directed acyclic graph (DAG) constraint, realize multimodal feature fusion and generate a dynamically coupled adjacency matrix A. coup Specifically, this includes:

[0164] Step S3.1: Implementation of hard constraints in causal DAG

[0165] Construct compliant unidirectional causal transmission paths: Spatial Geographic Mode → Structural Health Mode → Hydraulic Condition Mode → Pipeline Failure; Structural Health Mode → Hydraulic Condition Mode → Water Quality Safety Mode; Operation and Maintenance Event Mode → Hydraulic Condition Mode → Secondary Pipeline Failure Risk. For non-compliant paths (such as Hydraulic Mode → Structural Mode, Water Quality Mode → Structural Mode, etc.), the feature transmission weights are directly reset to zero. For example... Figure 4 As shown, where Figure 4 (a) in the figure represents the attention weight distribution of the prior application without causal hard constraints. The prior application is a Chinese patent, application number: 2025118997906, invention title: A method and system for dynamic risk assessment of water networks based on spatiotemporal graph convolution. Figure 4 Figure (b) shows the attention weight distribution under the hard constraint of the causal DAG in this invention. As can be seen from the figure, this invention reduces the proportion of false associations in non-compliant paths from 38.2% in the prior scheme to 0.4% by forcibly clearing the feature transmission weights of non-compliant causal paths to zero. This eliminates false associations from the root of the model structure and ensures that feature transmission fully conforms to the physical causal laws of the water supply network.

[0166] Step S3.2: Calculation of Causal Consistency Penalty Term

[0167] For each compliant causal path, three independent attention heads are set up, corresponding to the upstream and downstream hydraulic path, the structural failure cascade, and the operation and maintenance boundary, respectively. The feature output dimension of a single head is 128 dimensions, and the outputs of the three heads are concatenated before output. The dropout coefficient within the layer is 0.2. The feature propagation of non-compliant paths is directly cleared to zero. A causal consistency penalty term L is added to the total loss function of the model. causal The formula is:

[0168] ;

[0169] In the formula, N represents the total number of core nodes, and M represents the total number of non-compliant causal paths. For core node Vij In the p Attention weights on non-compliant paths.

[0170] In this embodiment, the total number of core nodes N=4, the total number of non-compliant paths M=8, and after the model training converges, the average attention weight of the non-compliant paths drops to 0.01, and the final penalty term L... causal =4×8×0.01=0.32.

[0171] The calculation process is as follows: The total loss function of the model is defined as: L total =L MSE +αL causal L MSE The mean squared error loss is used for predicting failure probability, and α=0.3 is the weight coefficient of the penalty term, which is optimal after training and verification.

[0172] Initial penalty term: During the initial training of the model, the average attention weight W for non-compliant paths. ijinvalid =0.1, Initial penalty term: L causal =4×8×0.1=3.2.

[0173] Convergence penalty: After 1000 iterations of training, the model converges when the loss function drops below 1e-4, and the average attention weight of non-compliant paths decreases to 0.01. Final penalty term: L causal =4×8×0.01=0.32, the proportion of non-compliant path feature transmission decreased from 38.2% of prior patents to 0.4%, completely blocking false associations.

[0174] Step S3.3: Dynamically Coupled Adjacency Matrix Generation

[0175] Dynamically Coupled Adjacency Matrix A coup The calculation formula is:

[0176] ;

[0177] In the formula, Θ is the learnable parameter tensor, and F merg represents the multimodal feature matrix after causal constraint fusion, and Softmax and ReLU are nonlinear activation functions.

[0178] In this embodiment, the multimodal feature matrix Fmerge (4×128 dimensions) after causal constraint fusion is substituted into the generation formula, and after normalization by ReLU and Softmax activation functions, a 4×4 dynamic coupling adjacency matrix A is generated. coup :

[0179] ;

[0180] The diagonal elements of the matrix represent the nodal autocorrelation coefficients, with a value of 1.0; the off-diagonal elements represent the failure coupling strength between pipe segments, with a value range of [0,1].

[0181] Step S4: Through node-level spatiotemporal dynamic gating units, dynamic weighted coupling of multimodal features is achieved, outputting the cross-modal coupled spatiotemporal feature vector of the core node. Specifically, this includes:

[0182] Step S4.1: For each core node and each time step, determine the initial contribution weight W for each modality. q (ij,t), the calculation formula is:

[0183] ;

[0184] In the formula, W q (ij,t) represents the core node of the q-th mode at time t. V ij The initial contribution weights at point X, σ is the sigmoid activation function, MLP is a multilayer perceptron, and X ij (t) represents the core node at time t. V ij The inherent properties of the pipe section, A ij (t) represents the core node at time t. V ij Real-time hydraulic operating condition characteristics, E ij (t) represents the core node at time t. V ij External event characteristics.

[0185] In this embodiment, there are 5 modal categories, including: structural health, hydraulic conditions, spatial geography, water quality safety, and operation and maintenance events.

[0186] The core node V at the 9th hour of the peak rainfall 24 Taking the example calculation: a 2-layer MLP is used (input layer 8-dimensional → hidden layer 1 (16-dimensional, ReLU activation) → hidden layer 2 (8-dimensional, ReLU activation) → output layer 5-dimensional), with the standardized pipe segment intrinsic attribute feature vector X 24 (9h), Real-time hydraulic condition feature vector A 24 (9h), Rainstorm Event Feature Vector E 24 (9h) is the input. After forward propagation and sigmoid activation, the initial contribution weights of the five modes are obtained: structural health 0.72, hydraulic condition 0.68, spatial geography 0.78, water quality safety 0.22, and operation and maintenance event 0.15.

[0187] Step S4.2: Modal redundancy suppression calculation

[0188] Calculate the cosine similarity of features between modes |cos(F) m1 ,F m2 )∣,when∣cos(F m1 ,F m2 When |>0.8, the redundancy suppression term L is used. red Weight decay is applied to highly redundant mode pairs; the formula for calculating the modal redundancy suppression term is:

[0189] ;

[0190] In the formula, L red This is a modal redundancy suppression term, used to reduce the weights of modes whose feature overlap exceeds a threshold; cos(F q1 ,F q2 W represents the cosine similarity between the modal features of classes q1 and q2; q1 W q2 F represents the initial contribution weights of the modes corresponding to q1 and q2, respectively; q1 and F q2 These represent the standardized feature matrices for different modes q1 and q2, respectively.

[0191] In this embodiment, the characteristic cosine similarity between the structural health mode and the spatial geographic mode is calculated to be 0.82, which exceeds the redundancy threshold of 0.8, and is therefore determined to be a highly redundant mode pair; the cosine similarity between the other modes is all <0.7, indicating no redundancy.

[0192] Redundancy suppression term L red =0.82×(0.72+0.78)=1.23, using the attenuation formula Weight decay is applied to the two modes.

[0193] Final weight correction result: Final weight W of structural health mode 1-final =0.72×(1-1.23 / 10)=0.72×0.877≈0.63; Spatial geographic modality final weight W 3-final =0.68, and the remaining non-redundant modal weights remain unchanged from their initial values.

[0194] Step S4.3: Weighted Coupling of Multimodal Features:

[0195] core node V ij Cross-modal coupled spatiotemporal eigenvector H ij The formula for calculating (t) is:

[0196] ;

[0197] In the formula, F q (ij,t) represents the core node of the q-th mode at time t.V ij Standardization features of the location; W q (ij,t) represents the core node of the q-th mode at time t. V ij The contribution weight of each location.

[0198] In this embodiment, the core node V is taken as the 9th hour of the peak rainfall. 24 Let's take an example to calculate:

[0199] Output a 128-dimensional cross-modal coupled spatiotemporal feature vector. Similarly, calculate the coupling feature vector H of the remaining core nodes. 12 (9h), H 23 (9h), H 34 (9h).

[0200] Step S5: Based on coupling characteristics and the dynamic adjacency matrix, deduce the spatiotemporal evolution of cascade failure, identify critical inflection points, and trigger a four-level hierarchical early warning system. Specifically, this includes:

[0201] Step S5.1: Initial Failure Event Setting

[0202] 9h, Pipeline segment P 24 (V) 24 The initial failure probability of the heavy rain is 65.8% > 0.5, so the pipe section is judged to have burst and failed. This serves as the initial triggering event for the cascading failure, and the scope of the initial failure is limited to the DMA2 partition.

[0203] Step S5.2: Iterative calculation of failure propagation probability

[0204] The formula for iteratively calculating the failure propagation probability of a convolutional layer in a cascaded failure propagation spatiotemporal graph is as follows:

[0205] ;

[0206] In the formula, P casc (ij, t+1) is the core node V ij The probability of cascading failure propagation at time t+1, N( V ij ) as the core node V ij The set of adjacent nodes, V kl As an adjacency to the core node, A coup (ij,kl) is the core node V ij and V kl The coupling correlation strength, H kl (t) represents the core node at time t.V kl The coupling characteristics of W casc With b casc Here, σ represents the learnable weight matrix and the bias term, respectively, and σ is the sigmoid activation function.

[0207] In this embodiment, there are two convolutional layers of the cascaded failure propagation spatiotemporal graph, corresponding to the second-order neighborhood of failure propagation. The convolutional kernel is a 3×3 spatiotemporal dual-dimensional convolutional kernel, and the iteration inference time step is 10 minutes.

[0208] The cascading failure trigger threshold is set to 0.5. When the cascading failure propagation probability is greater than or equal to the failure trigger threshold, the core section is determined to be... V ij corresponding pipe section P ij When a secondary failure occurs, the heterogeneous graph structure and the dynamically coupled adjacency matrix are updated synchronously, and the full-chain evolution of cascade failure is iteratively simulated. The threshold for determining the global characteristic mutation degree of the pipeline network is 0.8. When the global characteristic mutation degree C(t) exceeds this threshold for three consecutive time steps, the moment when the threshold is first exceeded is determined to be the critical inflection point T of cascade failure. cri The total number of core nodes in the pipeline network is N=4, corresponding to V 12 V 23 V 24 V 34 DMA partition is set to: V 12 V 23 Belongs to DMA1 partition, V 24 V 34 Belongs to DMA2 partition.

[0209] In this embodiment, a 10-minute time step is used for the extrapolation, based on the dynamic coupling adjacency matrix Acoup and the coupling eigenvector H. ij (t), calculate the cascading failure propagation probability of the core node at each time point, and the results are shown in Table 2 below. Figure 5 As shown. Among them. Figure 5 The horizontal axis represents the extrapolation time (in minutes) after the initial failure event, and the vertical axis represents the cascading failure propagation probability of each pipe segment's corresponding core node. Figure 5 The initial failure pipe segment P is fully presented in the middle. 24 The temporal evolution of the failure probability of other pipe sections after cascading failure is triggered is shown in the figure. The failure trigger threshold of 0.5 and the trigger interval corresponding to the fourth level warning are marked, which intuitively shows the spatiotemporal diffusion law of cascading failure in water supply network.

[0210] Table 2. Iterative Calculation Results of Cascade Failure Propagation Probability

[0211] ;

[0212] Step S5.3: Critical Inflection Point Identification

[0213] The formula for calculating the rate of change of the global characteristic gradient of the pipeline network, C(t), is as follows:

[0214] ;

[0215] In the formula, C(t) is the gradient change rate of the global characteristics of the pipeline network at time t. Core node at time t V ij The gradient of the cross-modal coupling feature, where N is the total number of core nodes.

[0216] In this embodiment, N=4; the mutation threshold was verified to be 0.8 based on the measured dataset. Δt=10min is the fixed time step for cascading failure simulation, and the initial steady-state benchmark is the node coupling characteristics of the pipeline network under normal operation 10 minutes before the initial failure (t=-10min).

[0217] Based on the time-step iterative calculation results, the time-series results of the global gradient change rate C(t) of the pipeline network in this embodiment are as follows: 0 min (9:00) after initial failure, C(t) = 0.28; 10 min (9:10), C(t) = 0.62; 20 min (9:20), C(t) = 0.85; 30 min (9:30), C(t) = 0.92; 40 min (9:40), C(t) = 0.94.

[0218] According to the judgment rule set in this invention: when C(t) first exceeds the preset mutation threshold of 0.8, and subsequently exceeds the mutation threshold of 0.8 for two consecutive steps, the total duration corresponding to the moment when the threshold is first exceeded is determined to be the critical inflection point T of cascade failure. cri In this embodiment, C(t) first exceeds the 0.8 threshold at 20 minutes after the initial failure, and the subsequent steps at 30 minutes and 40 minutes are consistently higher than the threshold, which meets the characteristics of a critical mutation.

[0219] Final determination: Critical inflection point T of cascade failure cri =20min (corresponding to the 9th hour and 20th minute of absolute time), that is, 20 minutes after the initial failure, the pipeline system will change from a locally controllable state to a globally irreversible collapse state.

[0220] Determining the Golden Window for Emergency Response: Based on the evolution sequence of cascade failures in the pipeline network, the changing patterns of cascade failure propagation probability, and the results of critical inflection point determination, in this embodiment, the time interval from a significant increase in the cascade failure propagation probability and the failure range about to cross the DMA zone to the triggering of the critical inflection point of the cascade failure is 10-20 minutes after the initial failure, corresponding to the absolute time of 9 hours 10 minutes to 9 hours 20 minutes. This interval is the golden window for emergency response; among them, 10 minutes after the initial failure (9 hours 10 minutes) is the optimal intervention time point, at which time the failure has not yet crossed the zone, the cascade failure propagation probability has not reached the failure threshold, and the blocking effect of emergency intervention is optimal. Figure 6 As shown, the horizontal axis represents the estimated time (in minutes) after the initial failure event, the left vertical axis represents the gradient change rate of the global characteristics of the pipeline network, the right vertical axis represents the proportion of failed pipeline segments, and the dashed line indicates the location corresponding to the mutation threshold of 0.8. The critical inflection point T of cascading failure is clearly marked in the figure. cri =20min, and the golden window period for emergency response of 10~20min, intuitively present the critical change characteristics of the pipeline network from local controllability to global collapse, and verify the effectiveness of the inflection point identification method of the present invention.

[0221] Step S5.4: Level 4 warning triggered

[0222] Based on the four-level early warning mechanism triggering conditions set by this invention, and combined with the failure evolution results and critical inflection point determination results of this embodiment, the triggering time and system state of each early warning level are as follows:

[0223] Blue Alert: From hour 0 to hour 59, no initial failure events occurred in the pipeline network. The cascading failure propagation probability P of all core nodes is [not specified]. casc (ij,t)<0.2, no risk of global crash, continuously triggering blue alerts;

[0224] Yellow Alert: 9 hours 0 minutes (0 minutes after initial failure), pipe segment P 24 An initial tube burst failure occurred, the failure range was limited to a single DMA2 partition, C(t) did not exceed the mutation threshold, there was no risk of cross-regional spread, and a yellow warning was triggered.

[0225] Orange alert: 9 hours 10 minutes (10 minutes after initial failure), adjacent control section P 23 P 34 The cascading failure propagation probabilities all exceed the failure threshold of 0.4 and are close to 0.5. The failure range is about to span DMA1 and DMA2 partitions. The remaining critical time T to the critical inflection point is predicted. res =10min, which is 1 time step, there is a potential risk of global crash, triggering an orange alert;

[0226] Red Alert: 9 hours 20 minutes (20 minutes after initial failure), pipe segment P 23 This triggers a secondary failure, extending the failure range across two DMA partitions, and triggers the critical inflection point T of a cascading failure. cri =20min, remaining critical time Tres=0min, the system is about to enter a global crash state, triggering a red alert.

[0227] Application No.: 2025118997906, Invention Title: A Method and System for Dynamic Risk Assessment of Water Networks Based on Spatiotemporal Graph Convolution, which is the inventor's prior patent. This application is the result of further research based on the prior application. The prior application focuses on the fusion of water quality and flow characteristics, but does not address the problem of multimodal spatiotemporal differences; this application designs a dedicated adaptive coding layer for five types of multimodal data to achieve precise spatiotemporal alignment. The prior application lacks hard constraints based on physical laws and relies on data-driven approaches, which can easily lead to false correlations; this application constructs a cross-modal causal DAG to structurally limit the compliance transmission path. The prior application assesses risk using a stability index but fails to identify critical inflection points; this application accurately captures the critical inflection point of global collapse through the gradient change rate of global features in the pipeline network.

[0228] This application employs heterogeneous graph modeling combined with hard causal constraints to eliminate spurious associations and address the distortion issues inherent in traditional modeling, resulting in deductions that better reflect engineering realities. Based on a dedicated scale-adaptive coding layer, it retains key information such as long-term pipe segment degradation and discrete events, avoiding distortion in fused features. It can identify critical inflection points in advance, reserving a golden window for emergency response and overcoming the limitations of real-time evaluation in comparative schemes. Furthermore, it quantifies the impact of four types of external events, including earthquakes and rainstorms, and its four-level early warning mechanism adapts to different failure ranges, enhancing its engineering practicality.

[0229] Example 2: A cross-modal causal coupling cascaded failure prediction system for water supply networks, used to execute the method described in Example 1, including:

[0230] The heterogeneous graph construction module is used to abstract water supply network segments into core nodes of a heterogeneous graph and the relationships between segments into heterogeneous edges, construct the heterogeneous graph structure, and obtain and store the full-dimensional attribute data of core nodes and heterogeneous edges.

[0231] The multi-source data adaptation module is used to acquire multimodal source data of the pipeline network, design a dedicated scale adaptation coding layer for each mode, complete the spatiotemporal alignment of multimodal data and generate standardized feature matrices; calculate the initial failure probability of the pipeline segment for four types of external events and incorporate them into the corresponding feature matrices;

[0232] The dynamic feature coupling module is used to design the spatiotemporal dynamic gating unit at the node level of the pipe segment. It adaptively calculates the contribution weight of each mode and adds a mode redundancy suppression term to complete the dynamic weighted coupling of multimodal features and outputs the spatiotemporal feature vector of node cross-modal coupling.

[0233] The causal constraint fusion module is used to pre-construct a cross-modal causal directed acyclic graph (DAG) of the water supply network, design a causal path-aware cross-modal graph attention layer, and generate a dynamically coupled adjacency matrix.

[0234] The cascaded failure simulation and early warning module is used to construct a convolutional layer of the spatiotemporal graph of cascaded failure propagation, simulate the spatiotemporal evolution law of failure, identify the critical inflection point of system collapse through information manifold learning, and output the cascaded failure simulation results and graded early warning signals.

[0235] Example 3: A computer-readable storage medium storing computer instructions thereon, characterized in that, when the computer instructions are executed by a processor, they implement the steps of the method described in Example 1.

[0236] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for cascading failure inference of water supply network across modal causal coupling, characterized in that, Includes the following steps: Step S1: Construct the heterogeneous graph structure of the water supply network: This involves connecting the physical pipe segments in the water supply network... P ij Abstracted as the core nodes of a graph structure V ij The physical connection relationship, hydraulic transmission relationship, and spatial adjacency relationship between pipe segments are abstracted into heterogeneous edges connecting adjacent core nodes; the static inherent properties and dynamic operation properties of the core nodes, as well as the static association properties and dynamic coupling properties of the heterogeneous edges are obtained. Step S2: Spatiotemporal scale adaptation of multi-source heterogeneous data: acquire multimodal source data of the pipeline network, including high-frequency hydraulic time series data, low-frequency structural health monitoring data, static spatial geographic data, discrete event-based operation and maintenance management data, medium-frequency water quality safety time series data, and external event-related data affecting pipeline network operation; To address the spatiotemporal scale differences between each modal data and the core node, a dedicated scale-adaptive coding layer is designed for each modality to achieve accurate spatiotemporal alignment between multimodal data and heterogeneous graph structures, and to generate standardized feature matrices corresponding to each modality; the initial failure probability of the pipe segment is calculated based on external event-related data, and the initial failure probability is integrated into the standardized feature matrix of the corresponding modality; Step S3: Cross-modal feature fusion with causal constraints: Pre-construct a cross-modal causal directed acyclic graph (DAG) specific to the water supply network, and clarify the physical causal transmission paths between data of each modality that conform to the principles of hydraulics, materials science, and geological engineering; use the causal directed acyclic graph (DAG) as a hard constraint, design a causal path-aware cross-modal graph attention layer, and only allow features to be transmitted and aggregated along compliant causal paths to generate a dynamic coupling adjacency matrix that characterizes the real-time coupling correlation strength between core nodes; Step S4: Node-level spatiotemporal dynamic feature coupling: Design spatiotemporal dynamic gating units at the core node level of the pipe segment. Based on the inherent properties of the pipe segment, real-time hydraulic conditions and external event characteristics, adaptively calculate the contribution weight of each modal data at each core node and each time step. A modal redundancy suppression term is added to reduce the weight of modes whose feature overlap exceeds the threshold. The dynamic weighted coupling of multimodal features is completed through a gating mechanism, and the cross-modal coupling spatiotemporal feature vectors of each core node of the water supply network are output. Step S5: Cascade Failure Inference and Critical Early Warning: Based on the coupled cross-modal spatiotemporal feature vector and dynamic coupling adjacency matrix, a convolutional layer of the spatiotemporal graph of cascade failure propagation is constructed to simulate the spatiotemporal evolution of failure events from the initial node to the entire pipeline network. The global feature gradient change rate of the pipeline network is calculated through information manifold learning to identify the critical inflection point of cascade failure from local controllability to global collapse. Based on the inference results, a four-level early warning mechanism is set up to output the spatiotemporal evolution path of cascade failure, the range of future time-series failures, the probability of system collapse, and the graded early warning signals.

2. The cross-modality causally coupled water distribution network cascade failure inference method of claim 1, wherein, In step S1, the core node V ij The static inherent properties are engineering design parameters that do not change with the operating state of the pipeline network, including pipe sections. P ij geographic coordinates (x ij ,y ij ), Design inner diameter D ij Physical length L ij Initial tensile strength σ of the pipe ij0 and the design maximum overcurrent capacity Q ij_des ; The dynamic operating attributes of the core node are time-series operating parameters obtained through real-time sensor acquisition and simulation calculation, including pipe sections. P ij The real-time flow Q at time t ij (t), Real-time flow rate v ij (t), head loss along the route h f-ij (t), Remaining wall thickness δ ij (t), uniform corrosion rate r ij_corr (t) and fatigue damage degree D ij_fat (t); The pipe section P ij Real-time flow rate v ij The calculation formula of (t) is as follows: ； wherein D ij is the pipe segment P ij is the design inner diameter, Q ij is the real-time flow rate of the pipe segment P ij at time t; The pipe section P ij head loss h f-ij (t) using the Darcy-Weisbach formula: ； where λ ij is the pipe segment P ij the frictional resistance coefficient, g is the gravitational acceleration, L ij is the pipe segment P ij length; The static association attributes of the heterogeneous edge include pipe segments. P ij and P mn The original space Euclidean distance d ij-mn Design hydraulic correlation degree and affiliation with the same operation and maintenance unit; The dynamic coupling properties of heterogeneous edges include pipe segments P ij and P mn Real-time traffic allocation ratio, pressure transmission coefficient, and failure correlation probability at time t.

3. The cross-modality causally coupled water distribution network cascade failure inference method of claim 1, wherein, In step S2, the design of the scale-adaptive coding layer for each modality is as follows: 1) For high-frequency hydraulic time series data, the sampling interval is minute level, and the time series alignment coding layer is designed to map the data to the time step dimension of the corresponding core nodes of the heterogeneous graph, generating a hydraulic feature matrix F matching the number of core nodes and time steps hyd ; 2), For low-frequency structure health data, the detection period is quarterly, and the cubic spline interpolation coding layer is designed to smooth the low-frequency detection data to the minute-level time step consistent with the hydraulic data, and finally generate the structure health feature matrix F str The interpolation formula is: ； In the formula, S ij (t) represents pipe segment P ij The interpolation result at time t, a k b k c k d k Let t be the spline coefficient for the k-th interpolation interval. k For pipe section P ij The kth detection time point; 3) For discrete event-based operation and maintenance data, design an event-triggered coding layer to encode segment P. ij The discrete events are transformed into a feature weight matrix F for continuous time steps. ops ; 4) For static spatial geographic data, an inverse distance weighted spatial coding layer is designed. This layer sequentially processes the spatial distance between pipe segments through min-max normalization, calculates the initial spatial coding weights, and performs a second min-max normalization on the initial weights. This maps the static attributes of the pipe segments to the corresponding core nodes and uniformly maps them to the [0,1] interval, generating a spatial geographic feature matrix F. geo The calculation formula is: ； ； ； where w ijmn_nor is the pipe segment P ij and the final normalized spatial encoding weight of the pipe segment P mn ; w ijmn For pipe section P ij With pipe section P mn Initial spatial encoding weights; w min w represents the minimum initial spatial coding weight for the entire pipeline network segment. max d represents the maximum value of the initial spatial coding weight for the entire pipeline segment; μ is the distance attenuation coefficient of the inverse distance weighted spatial coding, with a baseline value of 2 and a range of 1.5~3.0; ij-mn ′ represents a pipe section P ij With pipe section P mn The normalized space Euclidean distance, d ij-mn For pipe section P ij With pipe section P mn The original space Euclidean distance, d min d represents the minimum spatial distance between pipe sections in the entire pipeline network. max This represents the maximum spatial distance between sections of the entire pipeline network. 5) For medium-frequency water quality safety data, with sampling intervals on the order of hours, a water flow direction alignment coding layer is designed to map water quality characteristics to corresponding pipe sections along the water supply path. P ij core node V ij Generate the water quality safety feature matrix F wat .

4. The cross-modality causal coupling water distribution network cascade failure inference method of claim 3, wherein, In step S2, the initial failure probability of the pipe segment is calculated and incorporated into the corresponding feature matrix for four types of external events: earthquake, rainstorm, freeze-thaw, and third-party construction. The specific calculation formula is as follows: 1) Seismic event: initial failure probability of pipe segment P fail_seis The formula for (ij, t) is: ； In the formula, k seis Let a be the seismic attenuation coefficient. max For peak ground acceleration, σ ij (t) represents the pipe segment P ij Current remaining tensile strength, σ ij0 For pipe section P ij Initial tensile strength; 2) Rainstorm event: initial failure probability of pipe segment P fail_rain The calculation formula of (ij, t) is: ； In the formula, R real (t) represents the real-time rainfall at time t, R des (ij) represents the pipe segment P ij The designed flood control rainfall for the area is k rain h represents the impact coefficient of heavy rainfall. ij For pipe section P ij Burial depth, h min This refers to the minimum safe burial depth for water supply pipeline sections; 3) Freeze-thaw events: First, calculate the pipe segment damage using the Miner linear cumulative damage criterion. P ij Fatigue damage degree D ij_fat (t), then calculate the failure probability P fail_freeze (ij,t), the calculation formula is: ； In the formula, e r (ij) represents the pipe segment P ij The actual number of cycles experienced under the r-th stress, E r (ij) represents the pipe segment P ij The fatigue life corresponding to the r-th stress, k freeze The freeze-thaw effect coefficient is represented by E, which is the total number of stress cycles. 4) Third party construction event: initial pipe segment failure probability P fail_cons The formula for (ij,t) is: ； In the formula, k cons d is the construction impact factor. cons (ij,t) represents the construction point and pipe section at time t. P ij The distance is d0, where d0 is the safe distance threshold.

5. The cross-modality causally coupled water distribution network cascade failure inference method of claim 1, wherein, In step S3, the constraint mechanism and the generation rules of the dynamically coupled adjacency matrix of the causal directed acyclic graph (DAG) are as follows: 1) The compliant one-way causal transmission path is: Spatial geographic mode → Structural health mode → Hydraulic operating condition mode → Pipeline segment failure; Structural health mode → Hydraulic operating condition mode → Water quality safety mode; Operation and maintenance event mode → Hydraulic operating condition mode → Risk of secondary failure of pipeline segment; 2) Set an independent attention head for each compliance causal path, and directly clear the weight of the feature transmission of the non-compliance path; add a causal consistency penalty term L in the total loss function of the model causal , the formula is: ； In the formula, N represents the total number of core nodes, and M represents the total number of non-compliant causal paths. For core node V ij In the p Attention weight on non-compliant paths; 3) Dynamic coupling adjacency matrix A coup The calculation formula is: ； In the formula, Θ is the learnable parameter tensor, and F merge represents the multimodal feature matrix after causal constraint fusion, and Softmax and ReLU are nonlinear activation functions.

6. The cross-modality causal coupling water distribution network cascade failure inference method of claim 1, wherein, In step S4, the contribution weight calculation and feature coupling rules of the spatiotemporal dynamic gating unit are as follows: 1) For each core node, each time step, the initial contribution weight W of each modality q The calculation formula of (ij, t) is: ； In the formula, W q (ij,t) represents the core node of the q-th mode at time t. V ij The initial contribution weights at point X, σ is the sigmoid activation function, MLP is a multilayer perceptron, and X ij (t) represents the core node at time t. V ij The inherent properties of the pipe section, A ij (t) represents the core node at time t. V ij Real-time hydraulic operating condition characteristics, E ij (t) represents the core node at time t. V ij External event characteristics; 2) Calculate the inter-modal feature cosine similarity |cos(F q1 ,F q2 ) |, when |cos(F q1 ,F q2 ) | > 0.8, attenuate the weight of the high-redundancy modal pair through the redundancy suppression term L red ; the formula for calculating the modal redundancy suppression term is: ； In the formula, L red This is a modal redundancy suppression term, used to reduce the weights of modes whose feature overlap exceeds a threshold; cos(F q1 ,F q2 W represents the cosine similarity between the modal features of classes q1 and q2; q1 W q2 F represents the initial contribution weights of the modes corresponding to q1 and q2, respectively; q1 and F q2 These represent the standardized feature matrices for different modes q1 and q2, respectively. 3) core node V ij the cross-modal coupled spatio-temporal feature vector H ij The calculation formula of the cross-modal coupled spatio-temporal feature vector H ； where F q (ij,t) is the normalized feature of the qth modality at time t at the core node V ij of the core node W q (ij,t) is the contribution weight of the qth modality at the core node V ij at time t.

7. The cross-modality causally coupled water distribution network cascade failure inference method of claim 1, wherein, In step S5, the formula for iteratively calculating the failure propagation probability of the cascaded failure propagation spatiotemporal graph convolutional layer is as follows: ； In the formula, P casc (ij, t+1) is the core node V ij The probability of cascading failure propagation at time t+1, N( V ij ) as the core node V ij The set of adjacent nodes, V kl As an adjacency to the core node, A coup (ij,kl) is the core node V ij and V kl The coupling correlation strength, H kl (t) represents the core node at time t. V kl The coupling characteristics of W casc With b casc These are the learnable weight matrix and the bias term, respectively, and σ is the sigmoid activation function; Set the cascading failure trigger threshold P th When P casc (ij,t+1)≥P th At that time, determine the core node. V ij corresponding pipe section P ij When a secondary failure occurs, the heterogeneous graph structure and the dynamic coupling adjacency matrix are updated synchronously, and the full-chain evolution of cascade failure is simulated iteratively.

8. The cross-modality causally coupled water distribution network cascade failure inference method of claim 7, wherein, In step S5, the rules for identifying cascade failure critical inflection points and providing graded early warning are as follows: 1) The formula for calculating the global characteristic gradient change rate C(t) of the pipeline network is: ； In the formula, C(t) is the gradient change rate of the global characteristics of the pipeline network at time t. Core node at time t V ij The gradient of the cross-modal coupling feature, where N is the total number of core nodes; When C(t) exceeds the preset mutation threshold for 3 consecutive time steps, the time corresponding to the first step in the 3 consecutive steps is determined as the critical inflection point T of cascading failure cri ; 2) The triggering conditions for the Level 4 early warning mechanism are: Blue alert: No initial failure event, all core nodes V ij Cascade failure propagation probability P casc (ij,t)<0.2, C(t) is below the preset mutation threshold, and there is no risk of global collapse; Yellow alert: An initial failure event has occurred, the failure scope is limited to the independent metering area DMA of a single water supply network, C(t) is lower than the preset mutation threshold, and there is no critical inflection point trigger signal; Orange alert: The failure range is about to span two or more DMA partitions, the probability of failure propagation in adjacent control segments is ≥0.4 and less than the failure trigger threshold P. th Predict the remaining critical time T res (t) equals 1 time step, which poses a potential risk of global collapse; Red alert: failure range spans 2 or more DMA partitions, reaching critical corner T cri , remaining critical time T res (t) = 0, system is about to enter global collapse.

9. A cross-modality causally coupled water supply network cascading failure inference system, characterized in that, For performing the method according to any one of claims 1-8, comprising: The heterogeneous graph construction module is used to abstract water supply network segments into core nodes of a heterogeneous graph and the relationships between segments into heterogeneous edges, construct the heterogeneous graph structure, and obtain and store the full-dimensional attribute data of core nodes and heterogeneous edges. The multi-source data adaptation module is used to acquire multimodal source data of the pipeline network, design a dedicated scale adaptation coding layer for each mode, complete the spatiotemporal alignment of multimodal data and generate standardized feature matrices; calculate the initial failure probability of the pipeline segment for four types of external events and incorporate them into the corresponding feature matrices; The causal constraint fusion module is used to pre-construct a cross-modal causal directed acyclic graph (DAG) of the water supply network, design a causal path-aware cross-modal graph attention layer, and generate a dynamically coupled adjacency matrix. The dynamic feature coupling module is used to design the spatiotemporal dynamic gating unit at the node level of the pipe segment. It adaptively calculates the contribution weight of each mode and adds a mode redundancy suppression term to complete the dynamic weighted coupling of multimodal features and outputs the spatiotemporal feature vector of node cross-modal coupling. The cascaded failure simulation and early warning module is used to construct a convolutional layer of the spatiotemporal graph of cascaded failure propagation, simulate the spatiotemporal evolution law of failure, identify the critical inflection point of system collapse through information manifold learning, and output the cascaded failure simulation results and graded early warning signals.

10. A computer readable storage medium having stored thereon computer instructions, wherein, When executed by a processor, the computer instructions implement the steps of the method as described in any one of claims 1-8.

Images