An intelligent monitoring method and system for water conservancy facilities based on the Internet of Things

Abstract

The application discloses a kind of water conservancy facilities intelligent monitoring method and system based on Internet of Things, belong to water conservancy engineering technical field, including through sensor data acquisition, generate physical dependency graph, and it is fused with short-time dependency graph as spatial perception graph, establish spatial perception model, obtain abnormal propagation link and abnormal score set, extract high confidence abnormal link and in spatial perception model aggregation obtain global model, generate scheduling decision through global model.The water conservancy facilities intelligent monitoring method and system based on Internet of Things can be in multi-facility distributed monitoring scene, the spatial-structure dependency relationship between the multidimensional monitoring data inside water conservancy facilities is fully modeled, the spatial accurate positioning and propagation path tracing of abnormal event are realized, and the abnormal detection capability in single facility is significantly improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of water conservancy projects, and particularly relates to a water conservancy facility intelligent monitoring method and system based on the Internet of Things. BACKGROUND

[0002] With the development of the Internet of Things, sensor networks and cloud computing technologies, intelligent monitoring of water conservancy facilities has gradually become an important means to ensure the safety of water conservancy projects and improve management efficiency. Currently, the industry generally deploys multiple types of sensors on key water conservancy facilities such as reservoirs, dams and channels, collects real-time data such as water level, water quality, flow, seepage, leakage, deformation and stress-strain in multiple dimensions, and transmits the data to a cloud server through wireless communication technologies such as NB-IoT, 5G and LoRa. Through these data, the water conservancy management system can monitor the operation state of the facility, provide fault early warning and maintenance scheduling. However, the existing technology still has many deficiencies in practical application. On the one hand, most monitoring systems still use the "single-point monitoring + time series analysis" method, focusing on the time trend of independent sensor data, and it is difficult to effectively capture the spatial structure dependence between sensors, resulting in difficulty in accurately detecting and tracing when facing spatial anomalies such as internal leakage paths, crack propagation and stress concentration of complex water conservancy facilities.

[0003] On the other hand, water conservancy facilities generally have characteristics such as large-scale distribution, large differences in geographical environment and uneven operation and maintenance resources. Due to insufficient historical data, some remote or small water conservancy facilities are difficult to establish effective monitoring models and rely on traditional "rule threshold" or "simplified model", which has insufficient early warning capability. At the same time, the existing centralized model training method lacks a collaborative mechanism between distributed facilities, making it difficult to share abnormal knowledge of multiple facilities under the premise of privacy protection, resulting in a serious "information island" problem. More importantly, the current intelligent monitoring system often ignores the fusion of physical simulation results and actual observation data, lacks modeling of the hydrological-structure interaction within water conservancy facilities, and the abnormal detection results lack physical interpretability, which is prone to false positives and false negatives.

[0004] Therefore, we propose a water conservancy facility intelligent monitoring method and system based on the Internet of Things to solve the above problems. SUMMARY

[0005] The purpose of the present application is to solve the problem of inaccurate detection and false positives and false negatives in the prior art, and to propose a water conservancy facility intelligent monitoring method and system based on the Internet of Things.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] A water conservancy facility intelligent monitoring method based on the Internet of Things, comprising the following steps:

[0008] S1: Deploy the structural design parameters, environmental data, and water conservancy data of the sensor collection facility, calculate the seepage field and stress field inside the facility based on the structural design parameters, environmental data, and sensor deployment location of the facility, generate a physical dependency graph, the nodes of the physical dependency graph are the sensor set, and the edge set corresponds to the edge weight obtained from the consistency of the seepage field and the stress field;

[0009] S2: Based on the edge weight and spatial distance of two points in the physical dependency graph, a dynamic correlation coefficient is obtained, and a short-time dependency graph is constructed based on the water conservancy data collected by the sensor, the nodes of which are the sensor set, and the edge weight of the edge set is obtained through the dynamic correlation coefficient. Nonlinear fusion is performed on the edge weight of the physical dependency graph and the dynamic correlation coefficient of the short-time dependency graph to obtain a fused edge weight. All sensor node sets are taken as nodes to generate a spatial perception graph;

[0010] S3: A feature matrix of all sensor nodes is established based on the real-time water conservancy data collected by the sensor, and a spatial embedding calculation is performed using the graph convolution mechanism on the spatial perception graph to obtain the embedded node features. A spatial perception model is established based on the SAC and GNN models, and a chain structure regularization term is introduced into the model to force the spatial perception model to strengthen the similarity of the start and end features of the abnormal link during the training process. The chain abnormal propagation structure is extracted from the spatial perception model based on the water conservancy data, and the abnormal propagation link is obtained. Based on the embedded node features, each node is scored to output the abnormal score of each node;

[0011] S4: Extract high-confidence abnormal links from the abnormal propagation links, and perform weighted aggregation on the high-confidence abnormal links and the trained spatial perception model to obtain a global model. The high-confidence abnormal links of different facilities are aggregated to obtain a global abnormal link knowledge base;

[0012] S5: Input the spatial perception graph and feature matrix of the target facility into the global model to output the spatial anomaly score. Based on the spatial anomaly score and the abnormal link features in the global abnormal link knowledge base, a spatial risk level graph is constructed, and a dispatching decision is generated based on the spatial distribution in the spatial risk level graph.

[0013] Preferably, the water conservancy data collected in step S1 includes water level, flow rate, seepage rate, structural stress, and deformation.

[0014] Preferably, the physical dependency graph in step S1 introduces a regularization term to reduce the edge weight of abnormal paths that are physically inconsistent. By actively reducing the weight of hydrological and structural inconsistent paths, abnormal propagation paths with strong physical field consistency are preserved.

[0015] Preferably, in step S2, the dynamic collaborative features are obtained based on the edge weight and spatial distance of two points in the physical dependency graph, combined with the time delay between nodes.

[0016] Preferably, the instantaneous physical dependence strength and the difference between dynamic synergy features are introduced when generating the space perception graph in step S2. When the difference between the physical dependence and the dynamic synergy feature is large, the invalid edges in the graph structure are actively reduced, and the accuracy of capturing abnormal propagation links by the space perception graph is improved.

[0017] Preferably, the high-confidence abnormal link is extracted based on the following conditions in step S4: the fusion edge weight is higher than the dependence threshold; the abnormal score is greater than the average of the abnormal scores within the facility; and the abnormal link has spatial continuity.

[0018] Preferably, the specific steps of generating the scheduling decision in step S5 include:

[0019] S501: According to the spatial distribution of the space risk level graph, a spatial clustering method is used to divide abnormal high-risk areas;

[0020] S502: For each abnormal high-risk area, the operation and maintenance priority is calculated based on the balance between abnormal severity and the concentration of operation and maintenance resources;

[0021] S503: Output the space warning graph to reflect the risk level of each monitoring point, and output the dynamic operation and maintenance scheduling priority list.

[0022] A water conservancy facility intelligent monitoring system based on the Internet of Things, comprising:

[0023] A data acquisition module is configured to deploy sensors to collect structural design parameters, environmental data, and water conservancy data of the facility, calculate the internal seepage field and stress field of the facility based on the structural design parameters, environmental data, and sensor deployment location, generate a physical dependence graph, and the nodes of the physical dependence graph are a set of sensors, and the edge weights of the edge set are obtained from the consistency of the seepage field and the stress field;

[0024] A graph model construction module is configured to obtain a dynamic correlation coefficient based on the edge weight and spatial distance of two points in the physical dependence graph, construct a short-time dependence graph based on the water conservancy data collected by the sensors, and the nodes of the short-time dependence graph are a set of sensors, and the edge weights of the edge set are obtained from the dynamic correlation coefficient. Nonlinear fusion is performed on the edge weights of the physical dependence graph and the dynamic correlation coefficients of the short-time dependence graph to obtain fusion edge weights, and all sensor node sets are taken as nodes to generate a space perception graph.

[0025] The anomaly detection module is configured to establish a feature matrix for all sensor nodes using water conservancy data collected in real time by sensors. A graph convolution mechanism is used on the spatial perception graph to perform spatial embedding calculations to obtain embedded node features. A spatial perception model is established based on the SAC and GNN models, and a chain structure regularization term is introduced into it to force the spatial perception model to strengthen the similarity of the start and end features of the anomaly link during training. The water conservancy data is input into the spatial perception model to extract the chain anomaly propagation structure and obtain the anomaly propagation link. Based on the embedded node features, each node is scored and an anomaly score is output for each node.

[0026] The model optimization module is configured to extract high-confidence anomaly links from anomaly propagation links, perform weighted aggregation of the high-confidence anomaly links with the trained spatial perception model to obtain a global model, and aggregate the high-confidence anomaly links of different facilities to obtain a global anomaly link knowledge base;

[0027] The decision module is configured to input the spatial perception map and feature matrix of the target facility into the global model, output the spatial anomaly score, construct a spatial risk level map based on the spatial anomaly score and the abnormal link features in the global abnormal link knowledge base, and generate scheduling decisions based on the spatial distribution in the risk level map within the space.

[0028] In summary, the technical effects and advantages of the present invention are as follows: the method and system for intelligent monitoring of water conservancy facilities based on the Internet of Things can fully model the spatial-structural dependency relationship between the multi-dimensional monitoring data within water conservancy facilities in a multi-facility distributed monitoring scenario, achieve precise spatial positioning of abnormal events and traceability of the propagation path, and significantly improve the anomaly detection capability within a single facility. At the same time, the present invention also constructs a distributed cross-facility intelligent collaboration mechanism, which, under the premise of protecting data privacy, improves the model sharing and knowledge transfer capabilities between multiple facilities, helping remote water conservancy facilities with low resources and scarce data to quickly obtain intelligent monitoring and early warning capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flow chart of the method of the present invention;

[0030] Figure 2 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0032] like Figure 1 As shown, an intelligent monitoring method for water conservancy facilities based on the Internet of Things includes:

[0033] S1: Deploy the structural design parameters of the sensor collection facility, environmental data and water conservancy data, calculate the seepage field and stress field inside the facility based on the structural design parameters of the facility, environmental data and sensor deployment location, generate a physical dependency graph, the nodes of the physical dependency graph are the sensor set, and the edge set corresponds to the edge weight obtained by the consistency of the seepage field and the stress field;

[0034] S2: Based on the edge weight and spatial distance of two points in the physical dependency graph, a dynamic correlation coefficient is obtained, and a short-time dependency graph is constructed based on the water conservancy data collected by the sensor, the nodes of which are the sensor set, and the edge weight of the edge set is obtained through the dynamic correlation coefficient. Nonlinear fusion is performed on the edge weight of the physical dependency graph and the dynamic correlation coefficient of the short-time dependency graph to obtain a fusion edge weight, and all sensor node sets are taken as nodes to generate a spatial perception graph;

[0035] S3: A feature matrix of all sensor nodes is established based on the real-time water conservancy data collected by the sensor, and a spatial embedding calculation is performed using the graph convolution mechanism on the spatial perception graph to obtain the node features after embedding. A spatial perception model is established based on the SAC and GNN models, and a chain structure regularization term is introduced into the spatial perception model to force the spatial perception model to strengthen the feature similarity of the start and end points of the abnormal link during the training process. The water conservancy data is input into the spatial perception model to extract the chain abnormal propagation structure therein, and an abnormal propagation link is obtained. Based on the embedded node features, each node is scored to output an abnormal score of each node;

[0036] S4: Extract high-confidence abnormal links from the abnormal propagation links, and aggregate the high-confidence abnormal links with the trained spatial perception model to obtain a global model, and aggregate the high-confidence abnormal links of different facilities to obtain a global abnormal link knowledge base;

[0037] S5: Input the spatial perception graph and feature matrix of the target facility into the global model to output a spatial anomaly score, and based on the spatial anomaly score and the abnormal link features in the global abnormal link knowledge base, a spatial risk level graph is constructed, and a scheduling decision is generated according to the spatial distribution in the spatial risk level graph.

[0038] In this embodiment, the following is specific:

[0039] Step S1

[0040] The abnormal diffusion path in the Internet of Water Infrastructure is usually affected by the strong coupling effect of "seepage-stress" multi-physical field. The traditional method based on sensor distance or data correlation is difficult to describe the real spatial dependence. To solve this problem, the scheme first uses the structural design parameters of water conservancy facilities (such as dam CAD, boundary conditions, and rock and soil physical parameters), environmental data (rainfall, water level change, etc.), and sensor deployment location set V to calculate the seepage field F s and stress field F σ inside the facility based on the simulation module, which reflects the leakage trend and structural stress concentration area, respectively. s F 3 is the seepage velocity vector field, with the unit of m σ / s, and F phy is the stress distribution field, with the unit of Pa. The two physical fields provide basic physical information for subsequent graph modeling.

[0041] Based on the deployed sensor nodes V, the scheme proposes a multi-physical field driven spatial dependence graph generation method to construct the physical dependence graph G phy = (V, E phy ), where E s reflects the physical dependence relationship between sensor pairs. The edge weight calculation formula designed by the scheme is:

[0042]

[0043] where f s (i,j) is the seepage intensity of nodes v i to v j in the seepage field F 3 (the integral of flow between two points along the main direction of seepage, with the unit of m σ / s), f σ (i,j) is the stress gradient between two nodes in the stress field F ij (with the unit of Pa / m), α controls the contribution proportion of seepage and stress two physical fields, and β is the physical regularization factor.

[0044] To further ensure the physical rationality of the graph structure, the scheme introduces the regularization term R ij based on the consistency of the physical field to reduce the edge weight of the abnormal path with physical inconsistency:

[0045]

[0046] where and represent the local gradient vectors of seepage field and stress field along edge e 3 (with the gradient unit of m 2 / s / m and Pa / m), and γ > 0 is a regularization term strength adjustment factor. This design actively reduces the weight of the hydrology and structure inconsistent (such as the conflict between seepage direction and stress gradient direction) path through the penalize method, and retains the abnormal propagation path with strong physical field consistency, so that the generated G phy is more in line with the actual evolution of water conservancy facilities.

[0047] For example, in the left bank seepage monitoring scene of the dam, the simulation shows that the seepage main direction between v1 and v2 is consistent with the stress gradient direction, R 12 ≈1, while between v2 and v3, the seepage and stress directions differ significantly, R 23 <0.3. This mechanism makes w 12 >w 23 , in subsequent anomaly detection, v1 and v2 are preferentially considered as potential abnormal propagation paths, effectively avoiding the risk of "graph structure mismatching the actual abnormal path". The physical dependency graph G phy output by this step has real physical consistency constraints, the nodes are sensor sets V, and the edge set E phy corresponds to the edge weight w ij which is jointly determined by seepage, stress, and physical field consistency. The edge weight matrix W phy describes the physical correlation strength between each pair of sensors in space. This graph is directly input into the subsequent "physical-observation fusion perception graph", ensuring that the subsequent spatial perception and anomaly detection processes are based on the physical field distribution characteristics of water conservancy facilities, significantly improving the accuracy and engineering adaptability of the scheme in abnormal path tracing and spatial dependence modeling.

[0048] Step S2

[0049] This step is based on the physical dependency graph G phy output by the previous step = (V, E phy ), which truly reflects the spatial structure characteristics of "seepage-stress collaborative dependence" in water conservancy facilities through multi-physical field joint weighting. At the same time, the system also receives real-time observation data X obs from Internet of Things sensors = {x i |i∈V}, where each x i is the multi-dimensional observation feature (such as water level, flow, seepage rate, structural stress, deformation, etc.) of sensor v i within a fixed time window. The goal of this step is to organically fuse the physical dependence information with the statistical characteristics of the observation data, generating a spatial perception graph G fuse with "physical consistency + data dynamics" = (V, E fuse ), laying the foundation for subsequent spatial anomaly detection.

[0050] The scheme first constructs a short-time dependence graph G obs= (V, E obs ), whose edge set E obs reflects the dynamic data correlation between sensors in the current time window. Unlike traditional correlation calculation, this scheme designs a "time delay-offset sensitive correlation coefficient" for the "abnormal diffusion with spatial-temporal offset" feature in water conservancy monitoring:

[0051]

[0052] where τ ij is the time delay factor between node pairs v i and v j , which is automatically derived based on the edge weight w ij and spatial distance d ij of the two points in the physical dependency graph, reflecting the abnormal propagation time lag under the hydrological-stress link. For example, between the upstream sensor v i and the downstream sensor v j of the dam, τ ij reflects the dynamic physical offset feature of "how long the upstream anomaly affects the downstream".

[0053] The w ij in the physical dependency graph G phy represents the physical dependency strength, while c ij represents the dynamic coordination feature on the observation level. Directly superimposing the two may lead to information redundancy or weight imbalance. Therefore, the scheme proposes a "physical consistency dynamic adjustment fusion operator (PID-F)", which designs an innovative nonlinear fusion mechanism to generate the fusion edge weight e ij :

[0054] e ij = σ(λ·w ij +(1-λ)·|c ij |-γ·|w ij -|c ij ||)

[0055] where λ is the fusion coefficient, γ>0 is the physical-observation consistency penalty factor, and σ(·) is the normalized activation function. The innovation lies in the introduction of the |w ij -|c ij || difference term. When the physical dependency and the observation statistical dependency differ greatly, the system actively reduces e ij to reduce the "observation-physical separation" invalid edges in the graph structure and improve the accuracy of the fusion graph in capturing abnormal propagation links.

[0056] For example, in dam foundation abnormal seepage monitoring, if the physical dependency between v1 and v2 is strong (w 12 = 0.8), but the observation data correlation is weak (c 12= 0.2), the traditional method of simple weighting may still retain the edge, but the present scheme introduces a penalty by |w 12 -|c 12 | | = 0.6 to reduce e 12 , highlighting the strong link between physical and observation consistency, and improving the identification accuracy of abnormal links. In water conservancy facilities, this mechanism is particularly suitable for multi-source physical anomaly scenarios, such as upstream dam collapse causing downstream pressure anomalies, channel seepage causing adjacent structure deformation, etc. The system can effectively distinguish between "strong physical correlation but weak observation anomaly" and "strong observation anomaly but weak physical field dependence", and focus on preserving high-value abnormal propagation paths with physical-observation consistency. Finally, the scheme outputs the fusion graph G fuse = (V, E fuse ), whose edge set E fuse is described by the edge weight matrix E = {e ij}, which fully embeds the dual characteristics of "physical dependence + observation dynamic dependence". This graph will be directly used as input for the subsequent step "spatial anomaly detection", ensuring that the subsequent graph model has real physical structure support and dynamic observation consistency features when processing spatial anomaly perception, significantly improving the accuracy of anomaly localization and tracing.

[0057] Step S3

[0058] This step takes the spatial perception graph G fuse = (V, E fuse ) and its edge weight matrix E = {e ij} generated in step 2, where V is the set of all deployed sensor nodes in the water conservancy facility, E fuse is the edge set of the fusion graph, and the edge weight e ij fuses physical dependence and observation statistical features. In addition, the node feature matrix H fuse = {h i |i∈V} contains real-time observation data of sensors such as water level, seepage flow, stress and strain, etc. The goal is to design a graph neural network (GNN) based on G fuse that not only realizes spatial anomaly detection but also extracts "spatial anomaly propagation links" as subsequent federated collaboration anomaly knowledge. The scheme targets the typical anomaly characteristics of water conservancy facilities - anomalies usually present a "spatial diffusion chain" structure, i.e. abnormal events spread along seepage paths, stress concentration areas, or high dependence paths in the abnormal perception graph in the physical structure, rather than randomly distributed globally. To this end, a "spatial anomaly chain perception graph neural network (SAC-GNN)" is proposed, which designs a spatial anomaly link perception mechanism based on graph convolution to accurately extract the "chain-like anomaly propagation pattern" in water conservancy facilities. The specific steps are as follows

[0059] The spatial embedding calculation uses the graph convolution mechanism on the physical fusion perception graph, and each node vi The embedding representation h i ' is defined as:

[0060]

[0061] Among them, σ(·) is the activation function, φ(·) and ψ(·) are nonlinear mappings, and δ is the link perception factor. The innovation lies in the introduction of the second-order neighbor link information e ij e jk , capturing the “physics-observation fusion”, node v i V j Propagate to v k This mechanism is particularly suitable for the abnormal diffusion path of "crack-seepage-stress" in water conservancy facilities, overcoming the problem that traditional first-order graph convolution is unable to perceive long-chain anomalies.

[0062] To further enhance the model's ability to detect chain-like abnormal structures, the solution introduces a "chain structure regularization term" into the training objective:

[0063]

[0064] Among them, P tri Represents the fusion graph G fuse All the items that satisfy e ij >τ and e jk >τ's high-dependence ternary link (i.e., v i →v j →v k ) and η is the chain regularization strength. This regularization term strengthens the similarity of the embedding features of the starting and ending nodes in the chain structure, encouraging the model to focus on potential abnormal propagation chains in space rather than just focusing on local adjacency relationships.

[0065] In the dam leakage + stress anomaly scenario, if v1→v2→v3 forms an abnormal link along the seepage channel and the stress concentration area, and the edge weight e 12 、e 23 Significantly higher than adjacent paths (such as e 24 、e 35 ), then the link is included in P tri , the similarity between h1' and h3' will be automatically optimized during model training to capture the potential abnormal diffusion chain from v1 to v3. Combined with the above embedding features, the solution is based on the embedded node feature h i ', output the anomaly score s of each node through a dedicated anomaly scoring function i , generate anomaly score set S = {s i{ s | i e V}. Meanwhile, the scheme automatically extracts abnormal propagation links P trace , defined as the set of all chain edges satisfying e ij >τ and s i ,s j are both higher than the mean value. This link not only reflects the distribution of spatial anomalies within the facility, but also restores the "propagation path" of anomalies, providing structured anomaly information for subsequent cross-facility anomaly knowledge sharing in the federated learning phase. The final output includes: a set of spatial anomaly scores S, abnormal propagation links P trace , and the trained spatial perception model f θ (i.e., the parameter set of the SAC-GNN model).

[0066] Step S4

[0067] This step receives the spatial anomaly detection model f θ (i.e., the parameter set of the SAC-GNN model) from multiple water facilities, as well as the set of spatial anomaly scores S = {s i |i e V} and abnormal propagation links P trace The goal is to design a cross-facility "model parameter + spatial anomaly knowledge" collaborative optimization mechanism in a distributed facility environment. Unlike traditional federated optimization of model parameters, the abnormal propagation link P trace in water facilities has structured spatial information, and to achieve effective knowledge sharing, the scheme proposes a "structure-model joint migration mechanism (SMC-Fed)" to achieve collaborative optimization of model capabilities and anomaly knowledge.

[0068] In each facility, first extract the local model f θ and high-confidence anomaly link knowledge K i , K i derived from P trace , specifically retaining links (i,j) that satisfy the following conditions:

[0069] e ij >τ1, the fusion edge weight is higher than the dependence threshold;

[0070] s i ,s j scores are both higher than the mean value of the facility's anomaly scores;

[0071] Anomaly links have spatial continuity (there is a continuous i→j→k spatial chain).

[0072] Each facility uploads f θ and K i to the cloud coordination server as local federated input. The cloud first performs basic weighted aggregation on the model parameters, outputting the global model shared across facilities:

[0073]

[0074] wherein, n i is the number of samples of the i-th facility.

[0075] The innovation of the scheme lies in the introduction of "abnormal link-aware collaborative regularization". While aggregating the model, the uploaded K i is aggregated to generate a global abnormal knowledge base K global . K global contains spatial chain abnormal links from different facilities. The cloud introduces a cross-facility link-aware regularization term in model training:

[0076]

[0077] wherein, h i ' and h j ' are node spatial embeddings under the global model , and μ is the regularization strength. This mechanism guides the abnormal links of each facility through global abnormal link knowledge: so that it not only inherits the "abnormal detection capability" at the parameter level, but also describes the cross-facility link abnormal structure characteristics in spatial embedding.

[0078] For example, if facilities A and B both have the "upstream crack → leakage link → downstream settlement" pattern, the cloud aggregates their links into K global , which has a cross-facility shared "spatial chain abnormal distribution feature". Even if deployed to other facilities C with scarce abnormal links, it can still quickly perceive potential abnormal chains. Finally, the output is a federated global model and a global abnormal link knowledge base K global , which are directly input for subsequent deployment and intelligent early warning, realizing the global collaboration of "distributed abnormal knowledge + model parameters".

[0079] Step S5

[0080] This step inputs the global model and abnormal link knowledge base K global output by step 4. is a spatial anomaly detection model that integrates the model capabilities of multiple facilities, and K global is a high-confidence abnormal link set aggregated from multiple facilities. The goal of this step is to use these two to complete spatial intelligent early warning within the facility, output "partition risk level", and provide "abnormal distribution-based dynamic maintenance scheduling suggestions" for subsequent operation and maintenance.

[0081] The specific scheme first performs spatial perception on the target facility's spatial perception graph Gfuse and the characteristic matrix H fuse Perform reasoning and output spatial anomaly score S = {s i |i∈V}, that is, the spatial anomaly score of each node. Then, the scheme is based on the anomaly score S and K global Abnormal link features in the space risk level map R = {r i |i∈V}, reflects the intelligent warning level of each node. The solution proposes the "Link Enhanced Anomaly Aggregation Mechanism (LEA-Rank)", which is defined as:

[0082]

[0083] Among them, s i is the node anomaly score, For node i in K global The directly associated link neighbors in [1] are represented by [1], and λ is the link anomaly enhancement factor. This mechanism improves the risk perception of nodes on abnormal links, which is consistent with the abnormal characteristics of water conservancy facilities: "spatial anomalies spread along links."

[0084] After generating R, the solution designs a "Link Abnormal Distribution Driven Scheduling Strategy (LAD-Scheduler)" based on the spatial distribution in R. The specific steps are as follows:

[0085] First, based on the distribution of R, a spatial clustering method (such as density clustering based on anomaly scores) is used to divide the abnormal high-risk area Z = {Z1, Z2, ..., Z m};

[0086] Then, for each risk zone Z k , calculate its operation and maintenance priority index U k :

[0087]

[0088] in, is the sum of the warning scores of all nodes in the area, A(Z k ) is Z k The physical space coverage area is β, and β is the coverage area penalty factor. This function can balance the "severity of anomalies" and "centralization of operation and maintenance resources" to avoid wasting resources in scattered areas.

[0089] Taking the monitoring of a certain dam as an example, the node anomaly score and link density in the upstream dam shoulder area Z1 are high, the area is concentrated, and the U1 score is high. The system prioritizes allocating operation and maintenance resources to this area. Although there are a small number of high-scoring nodes in the dam bottom area Z2, the spatial distribution is discrete, the U2 score is lower than U1, and the priority is relatively low.

[0090] Finally, the solution outputs:

[0091] R-based spatial intelligent early warning map reflecting the risk level of each monitoring point;

[0092] U k based on the dynamic operation and maintenance priority list, the partitioned and quantitative response strategy is provided for the operation and maintenance team, and the closed-loop scheduling decision of early warning, response and maintenance is supported.

[0093] The technical solutions in the embodiments of the present application have at least the following technical effects or advantages: the method can fully model the spatial-structure dependency relationship between multi-dimensional monitoring data inside the water conservancy facility in a multi-facility distributed monitoring scene, realize spatial accurate positioning and propagation path tracing of abnormal events, and significantly improve the abnormal detection capability in a single facility. Meanwhile, the present application also constructs a distributed cross-facility intelligent cooperation mechanism, which improves the model sharing and knowledge migration capability between multiple facilities under the premise of data privacy protection, and helps remote water conservancy facilities with low resources and data scarcity to quickly obtain intelligent monitoring and early warning capability.

[0094] The embodiments of the present application also provide a water conservancy facility intelligent monitoring system based on Internet of Things, as shown in Figure 2 , comprising:

[0095] The data acquisition module is configured to deploy sensors to collect structural design parameters, environmental data and water conservancy data of the facility, calculate the seepage field and stress field inside the facility based on the structural design parameters, environmental data and sensor deployment position of the facility, generate a physical dependency graph, and the nodes of the physical dependency graph are a set of sensors, and the edge weights of the edge set are obtained from the consistency of the seepage field and the stress field;

[0096] The graph model construction module is configured to obtain a dynamic correlation coefficient based on the edge weight and spatial distance of two points in the physical dependency graph, construct a short-time dependency graph through the water conservancy data collected by the sensors, the nodes of the short-time dependency graph are a set of sensors, the edge weight of the edge set of the short-time dependency graph is obtained through the dynamic correlation coefficient, and the edge weight of the physical dependency graph and the dynamic correlation coefficient of the short-time dependency graph are nonlinearly fused to obtain a fused edge weight, all sensor node sets are taken as nodes to generate a spatial perception graph;

[0097] The anomaly detection module is configured to establish a feature matrix of all sensor nodes through the water conservancy data collected by the sensors in real time, perform spatial embedding calculation through a graph convolution mechanism on the spatial perception graph to obtain node features after embedding, establish a spatial perception model based on a SAC and a GNN model, and introduce a chain structure regularization term in the spatial perception model to force the spatial perception model to strengthen the feature similarity of the start point and the end point of the abnormal link in the training process, input the water conservancy data into the spatial perception model to extract a chain abnormal propagation structure therein, obtain an abnormal propagation link, and score each node based on the node features after embedding to output an abnormal score of each node;

[0098] The model optimization module is configured to extract high-confidence abnormal link from the abnormal propagation link, aggregate the high-confidence abnormal link with the trained spatial perception model to obtain a global model, and aggregate the high-confidence abnormal links of different facilities to obtain a global abnormal link knowledge base;

[0099] The decision module is configured to input the spatial perception graph and the feature matrix of the target facility into the global model, output a spatial anomaly score, construct a spatial risk level graph based on the spatial anomaly score and the abnormal link features in the global abnormal link knowledge base, and generate a scheduling decision according to the spatial distribution in the spatial risk level graph.

[0100] The technical solutions in the embodiments of the present application have at least the following technical effects or advantages: The present application introduces hydrology-structure physical prior information through physical simulation technology, fuses simulation results with actual monitoring data for modeling, constructs an abnormal detection mechanism with stronger physical consistency, reduces false positives, improves the interpretation of abnormal detection, and forms a systematic solution of "Internet of Things perception-spatial anomaly detection-collaborative knowledge transfer-physical consistency guarantee", which can effectively overcome various problems in the prior art and comprehensively improve the intelligent monitoring, fault early warning and operation and maintenance capabilities of water conservancy facilities.

[0101] The working principle is as follows: In a multi-facility distributed monitoring scene, the spatial-structure dependency relationship between multi-dimensional monitoring data inside water conservancy facilities is fully modeled, a distributed cross-facility intelligent collaboration mechanism is constructed, the model sharing and knowledge transfer capabilities between multiple facilities are improved under the premise of data privacy protection, physical simulation technology is used to introduce hydrology-structure physical prior information, simulation results are fused with actual monitoring data for modeling, an abnormal detection mechanism with stronger physical consistency is constructed, and a systematic solution of "Internet of Things perception-spatial anomaly detection-collaborative knowledge transfer-physical consistency guarantee" is formed, which can effectively overcome various problems in the prior art.

[0102] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An intelligent monitoring method for water conservancy facilities based on the Internet of Things, characterized in that: include: S1: Deploy sensors to collect the facility's structural design parameters, environmental data, and water conservancy data. Based on the facility's structural design parameters, environmental data, and sensor deployment locations, calculate the facility's internal seepage and stress fields, and generate a physical dependency graph. The nodes of the physical dependency graph are sensor sets, and the edge weights corresponding to the edge sets are derived from the consistency of the seepage and stress fields. S2: Based on the edge weights and spatial distances between two points in the physical dependency graph, a dynamic correlation coefficient is obtained. A short-term dependency graph is constructed using water conservancy data collected by sensors. Its nodes are sensor sets, and the edge weights of the edge set are obtained using the dynamic correlation coefficient. The edge weights of the physical dependency graph and the dynamic correlation coefficient of the short-term dependency graph are nonlinearly fused to obtain fused edge weights. The spatial perception graph is generated using all sensor node sets as nodes. S3: Build a feature matrix for all sensor nodes using water conservancy data collected in real time by sensors. Use a graph convolution mechanism on the spatial perception graph to perform spatial embedding calculations to obtain embedded node features. Build a spatial perception model based on the SAC and GNN models, and introduce a chain structure regularization term to force the spatial perception model to strengthen the similarity of the start and end features of abnormal links during training. Input the water conservancy data into the spatial perception model to extract the chain abnormal propagation structure and obtain the abnormal propagation link. Based on the embedded node features, score each node and output an abnormality score for each node. S4: Extract high-confidence abnormal links from the abnormal propagation links, perform weighted aggregation on the high-confidence abnormal links and the trained spatial perception model to obtain a global model, and aggregate the high-confidence abnormal links of different facilities to obtain a global abnormal link knowledge base; S5: Input the spatial perception map and feature matrix of the target facility into the global model, output the spatial anomaly score, construct a spatial risk level map based on the spatial anomaly score and the abnormal link features in the global abnormal link knowledge base, and generate scheduling decisions based on the spatial distribution in the spatial risk level map.

2. The method for intelligent monitoring of water conservancy facilities based on the Internet of Things according to claim 1, characterized in that: The hydraulic data collected in step S1 include water level, flow rate, seepage rate, structural stress and deformation, and the environmental data include rainfall and water level changes.

3. The method for intelligent monitoring of water conservancy facilities based on the Internet of Things according to claim 1, characterized in that: In step S1, the physical dependency graph introduces a regularization term to reduce the edge weights of physically inconsistent abnormal paths. By actively reducing the weights of hydrologically and structurally inconsistent paths, abnormal propagation paths with strong physical field consistency are retained.

4. The method for intelligent monitoring of water conservancy facilities based on the Internet of Things according to claim 1, characterized in that: In step S2, the time delay factor is obtained based on the edge weight and spatial distance of two points in the physical dependency graph and combined with the time delay between the nodes, and the covariance of the time delay factor is standardized to obtain the dynamic correlation coefficient.

5. The method for intelligent monitoring of water conservancy facilities based on the Internet of Things according to claim 4 is characterized in that: When generating the spatial perception map in step S2, a difference term between the instantaneous physical dependency strength and the dynamic collaborative feature is introduced. When the difference between the physical dependency and the dynamic collaborative feature is large, the invalid edges in the graph structure are actively reduced to improve the spatial perception map's ability to accurately capture abnormal propagation links.

6. The method for intelligent monitoring of water conservancy facilities based on the Internet of Things according to claim 1, characterized in that: In step S4, high-confidence abnormal links are extracted based on the following conditions: the fusion edge weight is higher than the dependency threshold; the abnormal scores are all greater than the average abnormal scores within the facility; and the abnormal links have spatial continuity.

7. The method for intelligent monitoring of water conservancy facilities based on the Internet of Things according to claim 1, characterized in that: The specific steps of generating the scheduling decision in step S5 include: S501: Based on the spatial distribution of the spatial risk level map, a spatial clustering method is used to divide abnormally high-risk areas; S502: For each abnormally high-risk area, calculate its operation and maintenance priority based on balancing the severity of the abnormality and the concentration of operation and maintenance resources; S503: Output a spatial early warning map to reflect the risk level of each monitoring point and output a dynamic operation and maintenance scheduling priority list.

8. An intelligent monitoring system for water conservancy facilities based on the Internet of Things, characterized in that: include: A data acquisition module is configured to deploy sensors to collect structural design parameters, environmental data, and water conservancy data of the facility. Based on the structural design parameters, environmental data, and sensor deployment locations of the facility, the seepage field and stress field inside the facility are calculated to generate a physical dependency graph. The nodes of the physical dependency graph are sensor sets, and the edge weights corresponding to the edge sets are obtained by the consistency of the seepage field and stress field. The graph model construction module is configured to obtain a dynamic correlation coefficient based on the edge weights and spatial distances between two points in the physical dependency graph. A short-term dependency graph is constructed using water conservancy data collected by sensors. Its nodes are sensor sets, and the edge weights of the edge sets are obtained using the dynamic correlation coefficients. The edge weights of the physical dependency graph and the dynamic correlation coefficients of the short-term dependency graph are nonlinearly fused to obtain fused edge weights. The spatial perception graph is generated using all sensor node sets as nodes. The anomaly detection module is configured to establish a feature matrix for all sensor nodes using water conservancy data collected in real time by sensors. A graph convolution mechanism is used on the spatial perception graph to perform spatial embedding calculations to obtain embedded node features. A spatial perception model is established based on the SAC and GNN models, and a chain structure regularization term is introduced into it to force the spatial perception model to strengthen the similarity of the start and end features of the anomaly link during training. The water conservancy data is input into the spatial perception model to extract the chain anomaly propagation structure and obtain the anomaly propagation link. Based on the embedded node features, each node is scored and an anomaly score is output for each node. The model optimization module is configured to extract high-confidence anomaly links from anomaly propagation links, perform weighted aggregation of the high-confidence anomaly links with the trained spatial perception model to obtain a global model, and aggregate the high-confidence anomaly links of different facilities to obtain a global anomaly link knowledge base; The decision module is configured to input the spatial perception map and feature matrix of the target facility into the global model, output the spatial anomaly score, construct a spatial risk level map based on the spatial anomaly score and the abnormal link features in the global abnormal link knowledge base, and generate scheduling decisions based on the spatial distribution in the spatial risk level map.

Images