A method and system for safety data governance in water conservancy and hydropower projects based on cloud-edge collaborative computing
By leveraging cloud-edge collaborative computing, edge nodes are used to process multi-source safety monitoring data from water conservancy and hydropower projects. Combined with lightweight convolutional neural networks and structural monitoring sensors, edge event data sets are generated, and data priority optimization is performed. This solves the real-time and bandwidth pressure problems under centralized data processing architecture, and achieves efficient safety monitoring and early warning.
Patent Information
- Application Number
- CN202610612805.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-07
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-05-07
AI Technical Summary
Traditional safety monitoring models for water conservancy and hydropower projects rely on centralized data processing architectures, which are difficult to cope with the real-time and bandwidth pressures of complex on-site environments. They also lack the ability to deeply understand and integrate video visual features and structural monitoring data, resulting in redundant data occupying network bandwidth and easily generating false alarms or missed alarms.
A cloud-edge collaborative computing approach is adopted, which utilizes edge nodes to preprocess multi-source security monitoring data and lightweight convolutional neural networks to identify abnormal events. Combined with the anomaly determination of the rate of change of structural monitoring sensor data, an edge event dataset is generated. Data priority is optimized through the phase space critical manifold, and finally, data fusion and risk assessment are performed in the cloud.
It enables local early warning at the edge, reduces network bandwidth consumption, avoids missed reports of security risks caused by data transmission delays, and provides real-time security assurance for water conservancy and hydropower projects.
Smart Images

Figure CN122554475B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of data governance and edge computing technology, specifically to a method and system for the governance of safety data in water conservancy and hydropower projects based on cloud-edge collaborative computing. Background Technology
[0002] With the continuous expansion of the scale of water conservancy and hydropower projects, their safety monitoring faces the severe challenge of processing massive amounts of heterogeneous data from multiple sources (such as video streams, displacement, and seepage pressure). Traditional monitoring models often rely on centralized data processing architectures, which are difficult to cope with the real-time and bandwidth pressures of complex on-site environments. In practical applications, existing edge computing nodes typically only perform simple threshold alarms or raw data forwarding, lacking the ability to deeply understand and fuse video visual features and structural monitoring data. This results in a large amount of invalid or low-value redundant data occupying network bandwidth, and due to the lack of accurate identification of historical water level change paths (such as unloading hysteresis effects) and critical abnormal states, false alarms or missed alarms are easily generated. Summary of the Invention
[0003] The purpose of this invention is to address the problems existing in the background technology by proposing a method and system for the governance of safety data in water conservancy and hydropower projects based on cloud-edge collaborative computing.
[0004] The technical solution of this invention: A method for safety data governance of water conservancy and hydropower projects based on cloud-edge collaborative computing, comprising: S1. Collect multi-source safety monitoring data of water conservancy and hydropower projects, and use edge nodes to preprocess the multi-source safety monitoring data to generate edge data; S2. Based on edge data, a lightweight convolutional neural network is used to identify abnormal events through video target detection and generate a set of video events. At the same time, the change rate of structural monitoring sensor data is combined with an asymmetric change rate anomaly judgment method based on water level unloading path memory to identify structural state anomalies and generate a set of state events. An edge event data set is generated based on the set of video events and the set of state events. S3. Construct a data priority function based on the edge data and edge event data set, and use the phase space critical manifold to perform topology locking optimization on the priority to generate a cloud data set; S4. Based on cloud-based data sets, data fusion is performed using a cloud platform to form a governance dataset; S5. Construct a safety status vector based on the governance dataset, and build a multi-indicator fusion risk assessment model based on the safety status vector to comprehensively assess and classify the structural safety status of water conservancy and hydropower projects.
[0005] As a further improvement to this technical solution, in step S1, multi-source safety monitoring data of water conservancy and hydropower projects are collected, and edge nodes are used to preprocess the multi-source safety monitoring data to generate edge data, including the following steps: Collect structural monitoring data of water conservancy and hydropower projects, including at least displacement, seepage pressure value, crack width, and tilt angle data; simultaneously collect on-site visual data to form a multi-source safety monitoring dataset; preprocess the multi-source safety monitoring dataset, and uniformly encapsulate the preprocessed multi-source safety monitoring data to generate edge data.
[0006] As a further improvement to this technical solution, in step S2, a lightweight convolutional neural network is used based on edge data to identify abnormal events through video object detection and generate a video event set, including the following steps: S2.1 Extract on-site visual data from edge data, decode the on-site visual data at edge nodes, perform frame extraction at fixed time intervals to generate a video frame sequence, and preprocess the video frame sequence. S2.2 Perform target detection inference on the preprocessed video frames based on a lightweight convolutional neural network, and output a candidate target set. ; S2.3. For target detection inference in video frames, a lightweight target association mechanism is introduced to construct a set of target cross-frame trajectories. And based on the target cross-frame trajectory set Extract the temporal behavioral features of the target to form an enhanced target state representation. ; S2.4, Based on enhanced target state representation Perform anomaly identification and generate anomaly events; S2.5. Summarize all identified abnormal events to form a video event set. .
[0007] As a further improvement to this technical solution, in step S2, structural state anomaly identification is performed by combining the structural monitoring sensor data change rate with an asymmetric change rate anomaly determination method based on water level unloading path memory, generating a state event set, and generating an edge event data set based on the video event set and the state event set, including the following steps: S2.6. Based on the preprocessed structural monitoring data, construct a continuous sequence under a unified time axis, and calculate the rate of change characteristics of each structural monitoring data based on discrete time difference. And based on the rate of change characteristics Generate multidimensional variable feature vectors; S2.7 Collect historical health operation data, construct normal fluctuation baselines for various structural monitoring data, and form a reference range for the rate of change; S2.8 Identify structural anomaly events using an asymmetric rate of change anomaly determination method based on water level unloading path memory, based on the rate of change reference interval; S2.9 Summarize all identified structural anomaly events to form a state event set. ; S2.10. Based on unified timestamps and spatial region identifiers, analyze video event sets. and state event set After spatiotemporal alignment and fusion, an edge event data set is generated. .
[0008] As a further improvement to this technical solution, in step S2.8, the asymmetric rate of change anomaly determination method based on water level unloading path memory is used to identify structural anomaly events, including the following steps: S2.81, Water level monitoring sequence collection Constructing water level change rate based on time difference Based on water level change rate Construct path state functions to generate path state sequences. ; S2.82, Based on path state sequence Constructing exponentially decaying path memory variables ; S2.83, Based on exponentially decaying path memory variables Characteristics of the rate of change Perform asymmetric rate of change correction to generate path-dependent asymmetric rate of change. ; S2.84, Based on the rate of change characteristic and water level change rate Construct a directional consistency index between structural response and water level change And based on the directional consistency index Constructing a direction anomaly indicator function ; S2.85, Based on exponentially decaying path memory variables asymmetric rate of change and direction anomaly indicator function Construct a comprehensive decision function It is used to identify structural anomalies.
[0009] As a further improvement to this technical solution, in step S3, a data priority function is constructed based on the edge data and edge event data set, and the priority is optimized by topology locking using the phase space critical manifold to generate a cloud data set, including the following steps: S3.1. Perform timestamp alignment and spatial correlation matching on edge data and edge event data sets, and construct a fused data set based on a unified indexing mechanism; S3.2. Based on the edge event dataset, edge events are classified into different risk levels; S3.3, For each piece of fused data in the fused data set To construct its comprehensive value factors; S3.4. Generate a data priority function based on a comprehensive value factor and by introducing a resource constraint function; S3.5. By constructing a phase space critical manifold with change rate characteristics, and implementing priority topology locking on the structural monitoring data within the phase space critical manifold, the data can maintain a stable priority under critical conditions, and an optimized priority can be generated. S3.6 Sort all fused data in descending order according to the optimized priority, generate a scheduling sequence, and build a real-time scheduling queue. Based on scheduling queue Filter out the cloud data set.
[0010] As a further improvement to this technical solution, in S3.5, by constructing a phase space critical manifold with change rate characteristics and implementing priority topology locking on the structural monitoring data within this phase space critical manifold, the following steps are included: S3.51, Based on the characteristics of asymmetric rate of change Construct a normalized rate of change vector ; S3.52, Based on normalized rate of change vector The infinite norm is used to reflect the overall anomaly intensity, and a critical state region is constructed. ; S3.53, Based on the critical state region For structural monitoring data, construct its locked state variables. ; S3.54. During the evolution of the phase space region attribution status of structural monitoring data over time, monitor and lock the state variables. Changes; S3.55, Based on locked state variables The changes in the result are used to construct a piecewise priority function to generate the optimized priority. .
[0011] As a further improvement to this technical solution, in step S4, data fusion is performed using a cloud platform based on the cloud data set to form a governance dataset, including the following steps: S4.1 Upload the cloud data collection to the cloud platform to uniformly access cloud data from different edge nodes; S4.2 To address the discrepancy between the timestamp and spatial location of cloud data, a spatiotemporal correction function is constructed. Through time alignment interpolation and spatial coordinate normalization, spatiotemporal consistency correction is performed on the cloud data, and a spatiotemporal consistent dataset is output. S4.3. Based on the spatiotemporal consistency dataset, a unified governance representation is generated through a weighted fusion mechanism, and the unified governance representation is encapsulated for governance to generate a structured governance dataset.
[0012] As a further improvement to this technical solution, in step S5, a safety state vector is constructed based on the governance dataset, and a multi-index fusion risk assessment model is constructed based on the safety state vector to comprehensively assess and classify the safety status of water conservancy and hydropower engineering structures, including the following steps: S5.1. Based on the governance dataset, feature decoupling and index extraction are performed using a feature separation network to construct a multi-dimensional safety index system, including structural stress indexes. Displacement and deformation index seepage pressure index Crack development indicators Furthermore, physical quantity normalization processing is performed on different indicators to obtain standardized indicator vectors. ; S5.2, Based on Standardized Indicator Vectors Construct structural safety state vector ; S5.3, Based on structural safety state vector Construct a multi-indicator integrated risk assessment model to generate a comprehensive risk value; S5.4. The comprehensive risk value is compared and divided with the preset classification threshold set to comprehensively assess and classify the structural safety status of water conservancy and hydropower projects, and the structural safety classification results are output for real-time early warning and decision support of water conservancy and hydropower projects.
[0013] On the other hand, the present invention provides a water conservancy and hydropower project safety data governance system based on cloud-edge collaborative computing, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the above-mentioned water conservancy and hydropower project safety data governance method based on cloud-edge collaborative computing.
[0014] Compared with existing technologies, the above-mentioned technical solution of the present invention has the following beneficial technical effects: by utilizing edge computing nodes deployed at the engineering site, localized preprocessing, target detection, and abnormal event generation are performed on the collected multi-source heterogeneous data (such as video streams, piezometer data, and displacement gauge data); only the key "event data" (such as crack propagation, personnel intrusion, and sudden pressure changes) extracted by the edge computing nodes are sent to the cloud; not only is network bandwidth resources significantly saved, but more importantly, by using lightweight convolutional neural networks and asymmetric rate of change determination methods for preliminary risk identification at the edge, effective local early warning can be achieved, avoiding the underreporting of safety hazards caused by data transmission delays, and providing dual protection for the real-time safe operation of water conservancy and hydropower projects. Attached Figure Description
[0015] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation
[0016] The technical solutions of the embodiments 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0017] Example 1: Please refer to Figure 1 As shown in the figure, this embodiment provides a method for the governance of safety data in water conservancy and hydropower projects based on cloud-edge collaborative computing, including the following steps: S1. Collect multi-source safety monitoring data of water conservancy and hydropower projects, and use edge nodes to preprocess the multi-source safety monitoring data to generate edge data; In this embodiment, multi-source safety monitoring data of water conservancy and hydropower projects are collected, and edge nodes are used to preprocess the multi-source safety monitoring data to generate edge data, including the following steps: By integrating data from structural monitoring sensors such as displacement gauges, piezometers, crack gauges, inclinometers, and water level gauges through an IoT sensing system, as well as data from video surveillance equipment and mobile inspection terminals, structural monitoring data for water conservancy and hydropower projects can be collected. This data includes at least displacement and seepage pressure values. Crack width Tilt state data Simultaneously, on-site visual data is collected through video surveillance equipment, and manual inspection records are collected through mobile inspection terminals, thus forming a multi-source safety monitoring dataset. At edge nodes (referring to computing processing units deployed at or near the site of water conservancy and hydropower projects, used for real-time preprocessing, lightweight analysis, and preliminary event identification of the collected multi-source monitoring data, and for data interaction with the cloud platform), time alignment processing is performed on data with different sampling frequencies based on a unified timestamp, and spatial calibration is performed in conjunction with engineering measurement point numbers, spatial coordinates, or structural partition information to achieve spatiotemporal consistency mapping of monitoring data across devices and multiple regions. The multi-source safety monitoring dataset is preprocessed (including at least data cleaning and abnormal data correction: noise filtering based on sliding windows, data normalization and scale unification processing: for different types of safety monitoring data, Min-Max normalization or Z-score standardization methods are used to perform scale unification processing on the data, so that data of different dimensions are mapped to a unified feature space), and the preprocessed multi-source safety monitoring data is uniformly packaged to generate edge data.
[0018] S2. Based on edge data, a lightweight convolutional neural network is used to identify abnormal events through video target detection and generate a set of video events. At the same time, the change rate of structural monitoring sensor data is combined with an asymmetric change rate anomaly judgment method based on water level unloading path memory to identify structural state anomalies and generate a set of state events. An edge event data set is generated based on the set of video events and the set of state events. In this embodiment, a lightweight convolutional neural network is used based on edge data to identify abnormal events through video object detection and generate a video event set, including the following steps: S2.1 Extract on-site visual data from edge data, decode the on-site visual data at edge nodes, and perform frame extraction at fixed time intervals to generate a video frame sequence. (First, identify and parse the data format of the incoming on-site visual data stream. For different encoding formats (such as H.264 / H.265), call the corresponding decoder to complete the stream decapsulation and frame-by-frame decoding to obtain the original continuous video frames. Then, perform time alignment on the decoded frame sequence based on a unified timestamp mechanism, and perform frame extraction according to a preset sampling strategy, that is, downsample the original frame sequence at fixed time intervals or target frame rates.) The process involves filtering (uniformly sampling when the original frame rate is higher than the target frame rate, and interpolating or skipping frames based on timestamps when the frame rate is unstable) to generate a temporally continuous video frame sequence with controllable sampling density. During this process, the timestamp, device identifier, and spatial location information of each frame are recorded synchronously to form spatiotemporally labeled video frame data, providing standardized input for subsequent target detection and temporal analysis. The video frame sequence is then preprocessed (including image size unification, pixel normalization, and noise reduction; and lightweight adaptation processing of the input data is performed based on edge computing resource constraints to reduce the computational load of model inference). S2.2 Deploy a lightweight convolutional neural network (lightweight CNN) at the edge nodes. Based on the lightweight convolutional neural network, perform object detection inference on the preprocessed video frames and output a set of candidate targets. ( For the first The bounding boxes of each target (objects detected in a video frame, such as people, equipment, cracks, water stains, foreign objects, etc.)). Label the target category (such as people, equipment, cracks, water stains, foreign objects, etc.). (for the corresponding confidence level), and using a preset confidence threshold. (In this embodiment, The candidate targets were screened using a value of 0.7 (calibrated through field experiments) to obtain a set of effective targets. The lightweight convolutional neural network takes preprocessed video frames as input (the input layer is a fixed-size image tensor). First, it passes through a lightweight feature extraction backbone network, which consists of multiple layers of depthwise separable convolutional modules, batch normalization, and non-linear activation functions (such as ReLU6) stacked together. Stride convolutions or lightweight downsampling modules are introduced between several layers to achieve feature map dimensionality reduction and receptive field expansion. Next, it enters a feature fusion layer, employing a simplified multi-scale feature fusion structure (such as a lightweight FPN) to upsample and laterally connect features at different scales, enhancing its ability to represent small and multi-scale targets. Finally, the detection head outputs the prediction results. The detection head typically consists of several parallel lightweight convolutional layers, used to output target bounding box parameters (position regression), class probability (classification branch), and confidence score (target existence score), forming an output tensor, i.e., the candidate target set. ; S2.3. For target detection inference in video frames, a lightweight target association mechanism (such as IoU-based matching or simplified Kalman filter prediction) is introduced to construct a set of target cross-frame trajectories. (Specifically: for the current moment) Effective target set The trajectory is prepared for matching with the set of trajectories established in the previous time step; for each existing trajectory, a simplified Kalman filter (retaining only position and velocity state variables) is used to predict its position in the current frame, resulting in a predicted bounding box. (First, construct a low-dimensional state vector for each established target trajectory, containing only the target's center position and velocity information, such as the low-dimensional state vector) ,in The coordinates of the bounding box center are For the corresponding velocity components; a state transition model is constructed based on the assumption of uniform motion, and the predicted state at the current moment is obtained by linearly extrapolating the state at the previous moment through the state transition matrix. Subsequently, the prediction center location is extracted from the prediction state. The predicted bounding box size is then recovered by combining the width and height parameters of the bounding box from the previous time step (or by slowly updating the width and height through modeling), and the predicted bounding box is generated. , For the goal at the previous moment The width of the bounding box, For the goal at the previous moment The bounding box height is determined; then, a matching cost matrix is constructed based on the intersection-over-union (IoU) ratio between the predicted bounding box and the current detection box, and a greedy matching algorithm is used to associate them one by one to obtain matching pairs (the set of predicted bounding boxes obtained from the state prediction at the previous time step). With the current frame detection result set Perform pairwise combinations, calculate the matching metric between each pair of candidates, and calculate the intersection-union ratio using the IoU method. Based on this, a matching cost matrix is constructed. Subsequently, a threshold screening is performed on the matching cost matrix to remove invalid matching pairs that do not meet the minimum IoU threshold 'a' (0.3 in this embodiment). Based on this, a greedy matching algorithm is used to sort all candidate matching pairs by cost from smallest to largest. The matching pair with the lowest current cost and whose corresponding predicted and detected bounding boxes are not occupied is selected as the valid association and removed from the candidate set. This process is repeated until no matching pairs are available or the termination condition is met. Finally, a set of one-to-one matching pairs is obtained, used to update the target trajectory and state, achieving cross-frame target association. For successfully matched targets, the current detection result is updated to the corresponding trajectory, forming a continuous trajectory sequence. ( This is the time of the final frame of the target trajectory. The system sets the starting frame time of the target trajectory and updates its state parameters; for unmatched detected targets, it initializes a new trajectory; for unmatched historical trajectories, it temporarily retains or deletes them according to a preset loss tolerance frame number; during the trajectory update process, it synchronously calculates the target's temporal attributes (such as velocity, displacement, dwell time, etc.), and finally constructs a stable and continuous set of target cross-frame trajectories. ), and based on the target cross-frame trajectory set Extract the temporal behavioral features of the target to form an enhanced target state representation. (Specifically: First, extract the sequence of bounding box center coordinates from consecutive frames in chronological order for each target cross-frame trajectory.) Instantaneous velocity is calculated based on the difference between adjacent frames. ( The time step is represented, and the average velocity and velocity variance are further calculated to characterize motion stability; secondly, the dwell time is calculated based on the number of consecutive frames of the trajectory within the same region. It is used to identify loitering or abnormal stationary behavior; at the same time, the trajectory offset is obtained by calculating the displacement vector between the trajectory start and end points. Based on this, the aforementioned temporal features are combined with the detection results of the current frame. Perform fusion and splicing to construct enhanced feature vectors And through the time-series feature mapping function The target state is then uniformly encoded and normalized to form an enhanced target state representation. Among them, the time-series feature mapping function It adopts a lightweight temporal coding and fusion structure, and its overall architecture consists of an input feature concatenation layer, a temporal modeling layer, an attention-weighted fusion layer, and an output mapping layer: First, the enhanced feature vector composed of object detection results and trajectory statistical features is processed. Numerical normalization and category embedding are performed, and vector concatenation is used to form a unified input representation. Subsequently, in the temporal modeling layer, a lightweight GRU is used to dynamically encode the feature sequences within continuous time windows to capture the temporal dependencies of the target motion state. The fused temporal features are then dimensionality-reduced and projected through a fully connected mapping layer to generate an enhanced target state representation. ); S2.4, Based on enhanced target state representation Anomaly identification is performed, and anomaly events are generated. Anomaly judgment rules include at least unauthorized personnel entering critical areas (based on area-person category matching), structural surface anomalies (such as crack propagation, water seepage, and other visual features), abnormal equipment stagnation or boundary crossing, and lack of safety protection (such as not wearing safety equipment). When any rule is triggered, a corresponding anomaly event is generated, along with event attribute information (timestamp, location, target category, confidence level, etc.). Specifically: for... Perform structured analysis to extract target category information, spatial location information, and motion behavior features (including speed). Duration of stay trajectory offset (etc.) and confidence information, combined with the spatial region to which the target belongs and the current time window. A spatiotemporal constraint relationship is established; then, based on a pre-set anomaly knowledge base, a multi-dimensional comparative analysis is performed on the above features, including regional safety constraint judgment (whether the target has entered a restricted or dangerous area), behavioral threshold comparison analysis (such as sudden speed changes, abnormal lingering, or trajectory deviation from the normal range), and category risk mapping analysis (such as prioritizing the identification of abnormal behaviors of high-risk target categories); on this basis, various anomaly features are fused and judged, and when any dimension feature exceeds its corresponding normal behavior boundary, the target is considered to have abnormal behavior at the current moment; finally, the target information that triggers the anomaly is encapsulated into an event to generate an anomaly event. Each event is appended with attributes such as timestamp, spatial location, trigger type, anomaly feature description, and confidence level, thereby achieving the transformation and output of enhanced state representation into structured anomaly events. The pre-defined anomaly knowledge base is constructed using a hierarchical structured rule base architecture, consisting of three levels: a basic rule layer, a behavior pattern layer, and a composite event layer. The basic rule layer stores atomic judgment rules directly related to the target behavior, including regional constraint rules (such as mapping relationships between restricted areas and restricted work areas), motion threshold rules (speed, acceleration, and mutation thresholds), dwell time rules, and trajectory deviation rules, all expressed uniformly in a parameterized form. The behavior pattern layer introduces multi-feature combination logic on top of the basic rules. The knowledge base is used to characterize typical abnormal behavior patterns, such as "rapid entry and prolonged stay," "slow-speed abnormal wandering," and "trajectory reversal." It forms a behavior-level judgment template by logically combining and weighting multiple basic rules. The composite event layer constructs cross-type association rules for complex scenarios, associating and mapping video abnormal events with multi-source information such as structural monitoring anomalies and equipment status anomalies, forming a joint anomaly pattern library of "vision, structure, and equipment." At the same time, the knowledge base supports a dynamic update mechanism based on historical anomaly samples. By summarizing rules and adaptively correcting parameters for newly added anomaly events, it achieves continuous expansion and optimization of anomaly rules, thereby improving the coverage and robustness of anomaly identification in complex construction or monitoring environments. S2.5. Summarize all identified abnormal events to form a video event set. .
[0019] Furthermore, structural anomalies are identified by combining the rate of change of structural monitoring sensor data with an asymmetric rate of change anomaly determination method based on water level unloading path memory, generating a set of state events, and generating an edge event data set based on the video event set and the state event set, including the following steps: S2.6. Based on the preprocessed structural monitoring data, construct a continuous sequence under a unified time axis (i.e., a sequence of structural monitoring data arranged continuously according to uniformly spaced time points), and calculate the rate of change characteristics of each structural monitoring data point based on discrete time difference. ( Indicates the first Class structure monitoring data at time The observed values, (corresponding to the rate of change), and based on the characteristics of the rate of change. Generate multidimensional variation feature vectors ( (The number of structural monitoring data). S2.7. Collect historical health operation data and construct normal fluctuation baselines for various structural monitoring data to form a reference interval for the rate of change. Specifically, firstly, extract historical structural monitoring data samples that are in a state of "no obvious abnormalities, structural stability, and safe operation" from the long-term operation monitoring database of water conservancy and hydropower projects, covering different water level conditions, seasonal changes, and typical operating cycles; then, perform time alignment and noise reduction processing on various monitoring data (such as displacement, seepage pressure, crack width, tilt angle, etc.), remove abrupt anomalies and sensor drift data, and form a highly reliable historical health dataset; based on this, construct the time series statistical distribution characteristics for each type of structural monitoring indicator, including mean, standard deviation, quantile intervals, and rate of change distribution, and calculate its historical rate of change sequence through a sliding time window; further, perform probability density estimation or interval statistical analysis on the rate of change sequence to determine its fluctuation range under normal operating conditions, forming the corresponding rate of change reference interval. Finally, the reference intervals of various monitoring indicators are uniformly encapsulated to construct a multidimensional rate of change baseline model, which is used as a comparison benchmark in subsequent anomaly judgment to realize the quantitative expression of the "normal fluctuation boundary" of structural health status. S2.8. Based on the rate of change reference interval, the asymmetric rate of change anomaly determination method based on water level unloading path memory is used to identify structural anomaly events (water level unloading path memory refers to the state variable that is time-decayed weighted recording of the historical path state of the evolution stage from high water level to low water level during the water level change process, which is used to characterize the cumulative response and hysteresis effect of the structure during the water level decline process). In this embodiment, the asymmetric rate of change anomaly determination method based on water level unloading path memory, using a rate of change reference interval, primarily addresses the problem in traditional anomaly determination methods for water conservancy and hydropower engineering safety monitoring that neglect the influence of historical water level change paths (especially unloading / dewatering stages) on the hysteresis effect and nonlinear recovery behavior of structural responses. This method constructs an exponentially decaying path memory variable to achieve selective forgetting and retention of water level loading and unloading history, and introduces an asymmetric rate of change correction mechanism: under the unloading path, negative change components are amplified, while a directional consistency index is used to determine whether structural responses and water level changes are inversely coupled. This effectively avoids normal hysteresis responses caused by water level fluctuations being misjudged as anomalies, or real anomalies being submerged during the unloading stage, significantly improving the robustness and sensitivity of structural safety assessment under complex hydraulic conditions, and ensuring the accuracy of risk identification for dams and other hydraulic structures during water storage and dewatering cycles. The method for identifying structural anomaly events using an asymmetric rate of change based on water level unloading path memory includes the following steps: S2.81, Water level monitoring sequence collection Constructing water level change rate based on time difference Based on water level change rate Construct path state functions to generate path state sequences. : ; The path state function Used to discretize and characterize the water level change process, where: when the water level change rate hour( The preset water level change rate threshold (0.05 m / h) indicates that the water level is continuously rising, corresponding to the structure entering the loading stage (water storage stage), where the water pressure on the structure gradually increases; when When the water level change is within the threshold tolerance range, the water level tends to be stable, the corresponding structure is in a stable loading stage, and external hydraulic disturbances are weak; when When the water level is continuously decreasing, the structure enters the unloading phase (water discharge phase), and the water pressure on the structure is gradually released. This three-valued state division maps the continuous water level change process into a discrete path state sequence. It is used to characterize the periodic evolution characteristics of loading, stabilization and unloading of structures under different water level evolution stages, and to provide basic state input for subsequent path memory variable construction and asymmetric rate of change correction; S2.82, Based on path state sequence Constructing exponentially decaying path memory variables ( (The memory decay coefficient was determined through expert experience). An exponentially decaying path memory variable was constructed. This involves extending the current water level change phase to a loading or unloading history with time-decaying memory, thus allowing anomaly detection to rely not only on instantaneous operating conditions but also on the path evolution process experienced by the structure. Specifically, By applying an exponential weighted decay to historical path states, the recent loading or unloading states have a greater impact on the current structural response, while the impact of long-term states gradually weakens, thus achieving a "memory retention and forgetting balance" for the water level fluctuation process. In structural state anomaly identification, this variable is used to characterize the hysteresis effect and cumulative stress state of the dam body or foundation under different water level evolution paths, providing path-dependent input for subsequent asymmetric change rate correction and anomaly enhancement judgment, avoiding misjudgments caused by relying solely on instantaneous change rates. S2.83, Based on exponentially decaying path memory variables Characteristics of the rate of change Perform asymmetric rate of change correction to generate path-dependent asymmetric rate of change. ;in, The value represents the overall loading and unloading history of the current structure and the intensity of its cumulative effect; based on this, the original structure's rate of change characteristics are monitored. Introducing a path dependency correction mechanism, when When the time frame indicates that the structure is in a state dominated by a loading or neutral stable path, the structural response is considered to be within the normal evolution range, and the rate of change remains at the original observed value. Unchanged; when When the structure is in a state dominated by the unloading path, it indicates that the structure exhibits hysteresis recovery or nonlinear unloading response characteristics. The negative change component is then enhanced and corrected by multiplying the original rate of change by an enhancement factor. ,in The unloading enhancement coefficient is used to adjust the amplification of the negative response by the unloading path, thereby highlighting the abnormal hysteresis or asymmetric response behavior of the structure during water level drop; ultimately, it generates the path-dependent asymmetric rate of change. This enables the modeling and enhanced representation of differentiated responses to structural monitoring data under different water level evolution paths; S2.84, Based on the rate of change characteristic and water level change rate Construct a directional consistency index between structural response and water level change It is used to enable the same rate of change to present differentiated interpretation capabilities under different historical backgrounds of loading or unloading at different water levels, thereby enhancing the sensitivity to the hysteresis effect, reverse response and nonlinear recovery process during the unloading stage; at the same time, by directionally enhancing or suppressing the rate of change under specific paths (such as unloading state caused by rapid precipitation), it can separate the environmental driving force from the actual abnormal response of the structure, and improve the discriminative power of anomaly identification and the ability to resist working condition interference. Judgment rule: When When the direction of the structural response rate of change is consistent with the direction of the water level rate of change, it means that the structural response and the change of external water level load have a positive coupling relationship, which is judged as a normal hydraulic and structural coupling state; when When the water level change rate is within the threshold range or the structural response change is not significant, the system is in a stable or weakly disturbed state and does not trigger anomaly detection; when When the direction of the structural response is opposite to the trend of water level change, that is, the reverse coupling phenomenon of response decrease during the process of water level rise or response increase during the process of water level fall, this state reflects the structural response hysteresis, path dependence effect or potential nonlinear abnormal evolution characteristics. Therefore, it is marked as an anomaly enhancement factor to increase the anomaly score weight at this moment and trigger key monitoring and risk warning. And based on the directional consistency index Constructing a direction anomaly indicator function ; S2.85, Based on exponentially decaying path memory variables asymmetric rate of change and direction anomaly indicator function Construct a comprehensive decision function Used to identify structural anomalies: (In the formula, For the amount of structural monitoring data, As the weight of structural monitoring data, For the first Each structural monitoring data at time The dynamic upper limit threshold for normalization is used for normalization. , The water level influence coefficient is obtained by fitting measured data from historical flood events, and its unit is h / m. The path memory impact weight is used to adjust the degree of impact of historical load or unload paths on the current risk. (Weighting to enhance directional consistency). Among them, when ( (The threshold for structural anomalies is determined by fitting experimental data) to identify a structural anomaly state. S2.9 Summarize all identified structural anomaly events to form a state event set. ; S2.10. Based on unified timestamps and spatial region identifiers, analyze video event sets. and state event set After spatiotemporal alignment and fusion, an edge event data set is generated. Specifically: First, extract the video event sets respectively. The timestamp and spatial region identifier of each event (such as the structural partition number covered by the camera or the engineering coordinate range after pixel coordinate system mapping), and the set of state events. Each event is assigned a timestamp and spatial region identifier (e.g., the structural partition, elevation, or dam section number where the measuring point is located). Then, using a globally unified timeline as a benchmark, the event timestamps are discretized to a preset time granularity (e.g., seconds or milliseconds), and a spatiotemporal matching window is constructed. For event pairs within the same time window that share the same spatial region identifier or exhibit spatial inclusion or adjacency, their spatiotemporal similarity is calculated. When the similarity exceeds a preset threshold b, the event is considered a fusionable synchronous event. For successfully matched events, the video event and status event are merged into a composite event unit, retaining their original attributes (e.g., event type, confidence level, monitoring values, etc.) and adding a fusion identifier. Unmatched events are directly retained as independent event units. Finally, all event units (composite or independent) are sorted chronologically and uniformly encapsulated into structured event data records. Each record includes a timestamp, spatial region identifier, event source type (video, status, or fusion), detailed attribute fields, and preliminary risk assessment labels, thereby generating an edge event data set. .
[0020] S3. Construct a data priority function based on the edge data and edge event data set, and use the phase space critical manifold to perform topology locking optimization on the priority to generate a cloud data set; In this embodiment, a data priority function is constructed based on the edge data and edge event data set, and the priority is optimized by topology locking using the phase space critical manifold to generate a cloud data set, including the following steps: S3.1 The edge data and edge event data sets are time-stamp aligned and spatially correlated. A fused data set is constructed based on a unified indexing mechanism. Each fused data sample contains both state observation information (i.e., a multi-dimensional continuous structure monitoring data set) and event triggering information (attributes such as the type of abnormal event and risk level recorded in the edge event data set), providing a unified input basis for subsequent risk and value assessment. Specifically, the timestamps in the edge data and edge event data are first standardized, and data from different sampling frequencies and different device sources are uniformly mapped to a global time reference. Then, discretization and resampling are performed according to a preset time granularity. A unified timeline index is formed. Subsequently, based on engineering measurement point numbers, structural partition identifiers, or spatial coordinate information, edge data and event data are spatially standardized and mapped to a unified spatial index set to achieve data comparability across devices and regions. For multi-source data under the same index, field-level alignment and structured reorganization methods are used to uniformly encapsulate structural monitoring values and event triggering information, forming a fused data unit containing state observations and event attributes. Finally, by traversing and aggregating all time-space index units, a standardized fused data set is constructed to provide a unified input basis for subsequent priority assessment and risk modeling. S3.2. Based on the edge event dataset, edge events are classified into different risk levels. Specifically, each edge event is classified into three levels—low risk, medium risk, and high risk—according to the event type, the degree of exceeding the limit of structural monitoring data change rate, duration, and spatial importance: Low risk: only includes a single visual anomaly (such as personnel briefly approaching the warning area, not wearing a safety helmet, etc.) and no related structural monitoring data anomalies, or the structural monitoring data change rate is less than 1 times the normal baseline; Medium risk: the structural monitoring data change rate exceeds the normal baseline by 1 to 3 times, or visual anomalies are accompanied by slight fluctuations in monitoring data, or multiple low-risk events occur in the same area but repeat within a short period of time; High risk: the structural monitoring data change rate exceeds the normal baseline by more than 3 times, or there are serious anomalies such as crack expansion, sudden changes in seepage pressure, and sudden changes in displacement, or continuous abnormal events occur in key areas (such as dam foundations and floodgates), or visual and structural monitoring are jointly determined to be major safety hazards (such as unauthorized personnel entering a restricted area while monitoring data is abnormal). S3.3, For each piece of fused data in the fused data set To construct its comprehensive value factors; specifically: firstly, the integrated data... The state observation information and event triggering information are structurally analyzed to extract multi-dimensional feature indicators, including basic attributes such as the intensity of structural monitoring change rate, spatial importance, time urgency, and data integrity. Subsequently, various features are normalized to eliminate dimensional differences, and corresponding weight coefficients are assigned based on their impact on system safety. On this basis, value contribution sub-items are constructed, including a risk contribution factor (reflecting the intensity of the impact of an event or state on structural safety), an information gain factor (reflecting the degree of data contribution to uncertainty reduction), a time sensitivity factor (reflecting the urgency of the data within the current time window), and a spatial criticality factor (reflecting the importance of the region where the data is located). Furthermore, a weighted fusion method is used to comprehensively calculate the above sub-items, resulting in a comprehensive value factor for the fused data. S3.4. Based on the comprehensive value factor, and by introducing a resource constraint function, a data priority function is generated: (In the formula, For the first fused data at any time The data priority function value describes the relative importance and processing priority of the data in cloud scheduling. To indicate the first The comprehensive value factor of the fused data is used to characterize the overall contribution of the data to risk identification, information gain, and structural security analysis. The global risk factor (whose value is normalized to between 0 and 1) is calculated using the following steps: 1. Calculate the weighted density of edge events within a fixed time window: Collect all events in the edge event dataset within the past 10 minutes, assign weights of 1, 2, and 4 respectively based on their risk level (low, medium, high), sum the weights of each event, and divide by the time window length to obtain the event weighted density; 2. Calculate the maximum normalized amplitude of the asymmetric rate of change in all structural monitoring data: Find the absolute value of the asymmetric rate of change in various monitoring data such as displacement, seepage pressure, and crack width. The factor with the largest ratio to its dynamic normal upper limit; calculate the proportion of high-risk events at the current moment: count the number of events classified as "high-risk" among all edge events at the current moment, divide by the total number of events to obtain a proportion; multiply the above three quantities by preset weighting coefficients and sum them, with each weighting coefficient ranging from 0.2 to 0.6, and the sum of the three being 1. The default values are: event weighting density weight 0.4, maximum rate of change amplitude weight 0.4, and high-risk event proportion weight 0.2; input the weighted sum from the previous step into the hyperbolic tangent function, and the final output value is the global risk factor. (ensuring it remains between 0 and 1) is used to characterize the dynamic changes in the overall security risk level over time. This is the risk modulation coefficient, used to adjust the global risk factor. The amplification or suppression effect on data priority. For resource constraint functions (resource constraint functions) By constructing a unified model of the multidimensional resource consumption during the cloud processing of fused data, the construction process includes: firstly, obtaining the... fused data at any time The corresponding computing resource consumption is used to characterize the CPU or GPU computing load required for model inference, feature calculation, and data fusion. Secondly, the communication resource consumption is obtained to characterize the network bandwidth and transmission latency overhead during the data upload from the edge node to the cloud. Simultaneously, the storage resource consumption is obtained to characterize the storage space resources occupied by the data during cloud caching or historical storage. Based on this, corresponding weight coefficients are introduced for different types of resource consumption to reflect the relative importance of computing, communication, and storage resources in the overall scheduling. Finally, a unified resource constraint function is generated through weighted fusion and normalization of the above three types of resource consumption. The computational resource consumption, communication resource consumption, and storage resource consumption in the resource constraint function are obtained through a collaborative monitoring mechanism between edge nodes and the cloud. This is a stability correction constant (a very small positive number) used to avoid cases where the denominator is zero or the value is unstable, and to ensure the continuity and computability of the priority function. S3.5. By constructing a phase space critical manifold with change rate characteristics, and implementing priority topology locking on the structural monitoring data within the phase space critical manifold, the data can maintain a stable priority under critical conditions, and an optimized priority can be generated. In this embodiment, step S3.5 addresses the problem in cloud-edge collaborative water conservancy and hydropower project safety data governance that traditional priority scheduling strategies rely solely on instantaneous risks or fixed thresholds. This leads to data in a "critical anomaly state" (i.e., monitoring indicators approaching but not yet exceeding alarm limits) being delayed or ignored due to priority fluctuations, thus missing early warning signals. This method constructs an infinite norm phase space of the normalized rate of change vector and defines a critical manifold region, introducing a hysteresis-controlled "topology locking" mechanism: when data is continuously in the critical region, its priority is frozen to the value at the entry time, and it is only unlocked after two consecutive moments outside the critical region. This effectively avoids repeated priority jumps caused by noise or instantaneous fluctuations, ensuring that data in a critical evolution state can be stably and preferentially uploaded to the cloud. This significantly improves the forward-looking perception capability and scheduling robustness for potential sudden anomalies, preventing early weak anomaly signals from being overwhelmed by resource competition at the edge, thereby enhancing the timeliness and reliability of water conservancy project safety early warnings. The process involves constructing a phase space critical manifold based on the rate of change characteristics and implementing priority topology locking on structural monitoring data within this critical manifold, including the following steps: S3.51, Based on the characteristics of asymmetric rate of change Construct a normalized rate of change vector ( (The number of structural monitoring data) is used to map the rate of change of multiple indicators to a unified phase space coordinate system, thereby characterizing the relative position of the current structural monitoring data sample in the multidimensional anomaly evolution space; S3.52, Based on normalized rate of change vector The infinite norm is used to reflect the overall anomaly intensity, and a critical state region is constructed. , This is the critical bandwidth parameter; obtained by applying the normalized rate of change vector. The infinity norm (i.e., the largest magnitude of the rate of change in each dimension) is calculated to characterize the current "maximum anomalous component" of the structure, and this is used to define a critical band close to the anomalous threshold in the multidimensional feature space, thereby dividing the structural state into a normal region, a critical transition region, and an anomalous region; among which the critical state region It is used to identify critical evolutionary states that have not yet fully crossed the boundary but are close to the risk threshold, enabling the system to perceive potential risks in advance before an anomaly actually occurs, and to provide criteria for subsequent priority locking, key monitoring and early warning triggering, thereby improving the foresight and stability of anomaly identification; S3.53, Based on the critical state region For structural monitoring data, construct its locked state variables. To achieve hysteresis control: ; Among them, being continuously in the critical state means being in the critical state for two consecutive adjacent sampling times ( and ), normalized rate of change vector of structural monitoring data All fall within the critical state region The inner (i.e., the infinite norm satisfies) This time continuity constraint is used to distinguish between a truly stable state remaining within the critical region and a single instantaneous crossing caused by noise or random fluctuations, thereby reliably identifying that the structure has entered an important critical stage of abnormal evolution and triggering a priority freeze operation; it indicates that at two consecutive adjacent sampling times ( and ), normalized rate of change vector of structural monitoring data None of them are in the critical state region Inner (i.e., infinite norm) or This continuity condition is used to confirm that the data has stably exited the critical region, rather than briefly jumping out due to transient noise; once this condition is met, the state variable is locked. This will be set to 0, thereby unfreezing the priority and restoring dynamic updates of the priority. The construction of this locked state variable ensures that once the locked flag is set to 1 due to continuous entry into the critical section, it must leave the critical section for two consecutive moments before it can be set to 0, thus avoiding frequent state switching caused by instantaneous noise. S3.54. During the evolution of the phase space region attribution status of structural monitoring data over time, monitor and lock the state variables. Changes: When When the data changes from 0 to 1, it means that the data is in the critical state region for the first time in two consecutive adjacent sampling times. Triggering priority freeze operation: Record the moment of entering lock. (i.e., the second consecutive moment in the critical zone), and update the freeze priority value. The frozen value will be used in subsequent... The period remains unchanged until The freeze is lifted when the value changes from 1 to 0. S3.55, Based on locked state variables The changes in the result are used to construct a piecewise priority function to generate the optimized priority. ; when At this time, the data is in a "perpetually critical" state, and the priority remains frozen. ;when Priority is restored and updated dynamically due to locked state variables. By adopting a hysteresis rule (requiring two consecutive moments of detachment before being cleared), the priority can be effectively avoided from repeated unfreezing due to instantaneous fluctuations at a single moment, thus achieving a dynamic balance between "stability and sensitivity". S3.6 Sort all fused data in descending order according to the optimized priority, generate a scheduling sequence, and build a real-time scheduling queue. (Real-time scheduling queue) This indicates that all fused data will be prioritized according to their optimized order. The ordered data set formed after sorting in descending order, where For the first One data item being scheduled, , The total number of data items to be scheduled, and which meet the corresponding priority. Based on scheduling queues The cloud data set is selected by first determining the resource constraints for data access based on the current available resources in the cloud (including computing resources, communication bandwidth, and storage capacity), such as the maximum schedulable data volume, the maximum data transmission volume, or the processing latency threshold; then, starting from the head of the scheduling queue, data is selected in descending order of priority. The system continuously accumulates the corresponding resource consumption (such as data size and computing overhead) in real time to determine whether the preset resource constraints are met. When the accumulated resource consumption does not exceed the constraint threshold, the data is added to the cloud data set; otherwise, the selection process stops. Finally, a cloud data set that meets the resource constraints and is selected according to priority is formed for subsequent cloud fusion analysis and risk assessment.
[0021] S4. Based on cloud-based data sets, data fusion is performed using a cloud platform to form a governance dataset; In this embodiment, data fusion is performed using a cloud platform based on a cloud-based dataset to form a governance dataset, including the following steps: S4.1 Upload the cloud data set to the cloud platform and perform unified access to cloud data from different edge nodes (during the access process, perform field mapping and data structure standardization for heterogeneous data formats (such as structured data, time-series data and event data)). S4.2. To address the discrepancy between cloud data acquisition timestamps and spatial positioning, a spatiotemporal correction function is constructed. Through time-aligned interpolation and spatial coordinate normalization, spatiotemporal consistency correction is performed on cloud data, outputting a spatiotemporally consistent dataset to ensure comparability and synchronization of data from different edge nodes. Specifically: First, the timestamp and spatial location information of each fused data point received from the cloud are extracted. The timestamp may originate from asynchronous sampling records from different edge nodes, while the spatial information is obtained based on the measurement point number or geographic coordinates. Then, a unified time axis is constructed in the time dimension, and time alignment processing is performed on data from different sampling frequencies with a fixed time granularity. For data with missing time points, linear interpolation, spline interpolation, or sliding window completion methods are used for reconstruction, mapping all data to a unified time reference. In the spatial dimension, spatial coordinates from different sources are normalized, uniformly mapping engineering measurement points, structural zoning, or GPS coordinates to a standard spatial coordinate system to eliminate spatial representation differences between different systems. Based on this, the time alignment results and spatial normalization results are processed through the spatiotemporal correction function. Perform joint mapping and correction (spatial-temporal correction function) Using data timestamps and spatial location information as input, a joint transformation function is constructed through time alignment mapping and spatial coordinate normalization mapping to map the original data into spatiotemporally consistent data with a unified time axis and a standard spatial coordinate system. The spatiotemporal correction function performs corrections on each data point in both time and space dimensions. The specific logic is as follows: Time correction: First, a globally unified time reference is set, such as the standard time of a cloud server. Then, considering the clock drift between edge nodes and the cloud, an average clock drift rate (between 0.999 and 1.001) is obtained through periodic network time protocol synchronization measurements. For each data point, its original recording time is subtracted from the global reference time, divided by the average drift rate, and then added back to the global reference time to obtain the corrected timestamp. Subsequently, due to the potential inconsistency in sampling frequencies among different sensors, a piecewise cubic Hermite interpolation method was employed to align the corrected timestamps to a preset, equally spaced time axis, with a default interval of 1 second. Spatial correction: First, the original local coordinates of the sensors were transformed to the origin of the engineering global reference system through translation transformation. The components of the translation vector were taken from the reference point deviation given in the sensor installation calibration file. Then, a scaling transformation was performed: a diagonal scaling matrix was constructed, with its diagonal elements determined based on the conversion relationship between the original and target units (uniformly meters) of each coordinate axis: if the original input was in meters, it was set to 1; if the original input was in millimeters, it was set to 0.001. Next, a rotation transformation was performed: a three-dimensional rotation matrix was constructed based on the horizontal azimuth and pitch angles recorded in the sensor installation calibration file. The translated coordinates were then multiplied sequentially by the scaling matrix and the rotation matrix to obtain the corrected global spatial coordinates. If the sensor has no rotational relationship with the global reference frame, the rotation matrix is taken as the identity matrix; finally, the time-corrected unified timestamp and the spatially corrected standard spatial coordinates are appended to the metadata field of the original data, and the output is spatiotemporally consistent data, so that each data has a unified time index and spatial identifier; finally, the spatiotemporally consistent dataset is output, so that all cloud data are comparable and consistent in time and space dimensions. S4.3. Based on the spatiotemporal consistency dataset, a unified governance representation is generated through a weighted fusion mechanism, and the unified governance representation is then encapsulated to generate a structured governance dataset. Specifically, each data point in the spatiotemporal consistency dataset is first structurally parsed to extract its state observation features (such as displacement, seepage pressure, cracks, etc.), event features (such as anomaly type, risk level), and spatiotemporal attribute features (standard timestamps and spatial coordinates). All features are then uniformly normalized to eliminate dimensional differences. Subsequently, a unified governance representation is generated through weighted fusion to comprehensively characterize the overall contribution of a single data point to structural safety assessment. Based on this, the unified governance representation undergoes governance encapsulation processing, which involves organizing and standardizing the fusion results according to a predefined data structure (the predefined data structure is set according to the safety analysis requirements of water conservancy and hydropower projects, dividing the data into dimensions such as time, space, structural state, events, and fusion results, and uniformly standardizing the types, value ranges, and expression methods of the corresponding fields for each dimension), forming structured data units containing basic attributes, risk attributes, and spatiotemporal attributes; finally, all encapsulated data units are aggregated and indexed and reorganized to generate a structured governance dataset to support subsequent safety status modeling and risk assessment analysis.
[0022] S5. Construct a safety state vector based on the governance dataset, and build a multi-index fusion risk assessment model based on the safety state vector to comprehensively assess and classify the structural safety status of water conservancy and hydropower projects, so as to guide the safe operation scheduling and risk prevention and control decisions of water conservancy and hydropower projects. In this embodiment, a safety state vector is constructed based on the governance dataset, and a multi-indicator fusion risk assessment model is built based on the safety state vector to comprehensively assess and classify the structural safety status of water conservancy and hydropower projects, including the following steps: S5.1. Based on the governance dataset, feature decoupling and index extraction are performed using a feature separation network to construct a multi-dimensional safety index system, including structural stress indexes. Displacement and deformation index seepage pressure index Crack development indicators Furthermore, physical quantity normalization processing is performed on different indicators to obtain standardized indicator vectors. To eliminate the impact of dimensional differences on model input, the following steps are taken: First, the structured data organized by time series in the governance dataset is decomposed into multi-scale temporal components. Methods such as sliding window or wavelet decomposition are used to break down the original data into short-term fluctuation components, long-term trend components, and periodic components to separate information at different time scales. Then, the decomposed multi-scale features are input into a feature separation network. An encoding layer unifies the representation of state observation features and event features from different sources, and an attention weighting mechanism is introduced to dynamically assign weights to key features, enabling the model to automatically focus on key variables that significantly impact structural safety. Based on this, different feature subspaces are decoupled to separate independent indicators such as structural stress, displacement deformation, seepage pressure, and crack evolution. Finally, the decoupled features are standardized, mapped, and recombined to form a multi-dimensional safety indicator system. The feature separation network employs a modular architecture combining fully connected and multi-head attention to decouple and extract four independent safety indicators—structural stress, displacement deformation, seepage pressure, and crack development—from the multi-dimensional temporal fusion data of the governance dataset. Specifically, the network input consists of multidimensional temporal data within a fixed time window. The input data first passes through a one-dimensional convolutional network and a lightweight Transformer encoder to extract local and global temporal dependencies, outputting a temporal feature matrix. This matrix is then fed into an attention decomposition layer with a multi-head self-attention mechanism. This layer generates weight vectors based on the dynamic importance of the input features and multiplies them element-wise with the features, enhancing key features and suppressing redundant features to obtain a weighted feature vector. This weighted feature vector is then fed in parallel into four structurally identical and independent fully connected sub-networks. Each sub-network consists of two linear layers and a LeakyReLU activation function, outputting a feature subspace vector. These four subspaces correspond to structural stress, displacement deformation, seepage pressure, and crack development indices. During training, a decoupling loss function based on the Hilbert-Schmidt independence criterion is introduced to minimize the mutual information between subspaces, while a reconstruction loss is used to ensure that the decoupled features do not lose their original information. The output layer concatenates the four subspace vectors and then performs linear mapping and Softmax normalization to generate the final multidimensional safety index vector. The entire network is deployed in the cloud and pre-trained to convergence using historical security monitoring data. During inference, inputting the multi-dimensional fusion data of the current window will output standardized security indicators in real time, thus providing clear indicator inputs for subsequent structural security status assessment. S5.2, Based on Standardized Indicator Vectors Construct structural safety state vector ; S5.3, Based on structural safety state vector Construct a multi-indicator integrated risk assessment model to generate a comprehensive risk value. ( An index for a multi-dimensional safety indicator system used in risk assessment. For the first (Weight of security indicators) S5.4, Comprehensive Risk Value With preset graded threshold set ( This serves as the dividing line between low and medium risk. This serves as the dividing line between medium and high risk. In this embodiment, the threshold separating high risk from extremely high risk is defined as follows: , , This is used to compare and classify the structural safety status of water conservancy and hydropower projects, and to comprehensively assess and classify the current comprehensive risk value. sequentially with the preset graded threshold set Interval matching is performed on each threshold, when When it is determined to be a low-risk state, when When it is determined to be a medium-risk state, when When it is determined to be a high-risk state, when It is determined to be in an extremely high-risk state and outputs the structural safety classification results for real-time early warning and decision support for water conservancy and hydropower projects.
[0023] Example 2: This example provides a water conservancy and hydropower project safety data governance system based on cloud-edge collaborative computing, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the water conservancy and hydropower project safety data governance method based on cloud-edge collaborative computing described in Example 1 above.
[0024] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.
Claims
1. A cloud-edge collaborative computing-based water conservancy and hydropower engineering safety data management method, characterized in that, include: S1. Collect multi-source safety monitoring data of water conservancy and hydropower projects, and use edge nodes to preprocess the multi-source safety monitoring data to generate edge data; S2. Based on edge data, a lightweight convolutional neural network is used to identify abnormal events through video target detection and generate a set of video events. At the same time, the change rate of structural monitoring sensor data is combined with an asymmetric change rate anomaly judgment method based on water level unloading path memory to identify structural state anomalies and generate a set of state events. An edge event data set is generated based on the set of video events and the set of state events. This includes the following steps: S2.6, based on the pre-processed structure monitoring data, construct a continuous sequence under a unified time axis, and calculate the change rate characteristics of each structure monitoring data based on discrete time difference , and based on the change rate characteristics generate a multi-dimensional change feature vector; S2.7 Collect historical health operation data, construct normal fluctuation baselines for various structural monitoring data, and form a reference range for the rate of change; S2.8 Identify structural anomaly events using an asymmetric rate of change anomaly determination method based on water level unloading path memory, based on the rate of change reference interval; S2.81, collect water level monitoring sequence , construct water level change rate based on time difference , represent time step, construct water level change rate based on water level change rate construct path state function, generate path state sequence : the path state function is used for discrete characterization of water level change process, preset water level change rate threshold value; S2.82, Based on path state sequence Constructing exponentially decaying path memory variables , This is the memory decay coefficient; S2.83, Based on exponentially decaying path memory variables Characteristics of the rate of change Perform asymmetric rate of change correction to generate path-dependent asymmetric rate of change. , This is the unloading enhancement factor; S2.84, Based on the rate of change characteristic and water level change rate Construct a directional consistency index between structural response and water level change And based on the directional consistency index Constructing a direction anomaly indicator function ; S2.85, Based on exponentially decaying path memory variables asymmetric rate of change and direction anomaly indicator function Construct a comprehensive decision function Used to identify structural anomalies: In the formula, For the amount of structural monitoring data, As the weight of structural monitoring data, For the first Each structural monitoring data at time The dynamic upper limit threshold for normalization is used for normalization. , This is the water level influence coefficient. The path memory affects the weight. Weights are enhanced to improve directional consistency, where, when It was determined to be a structural anomaly. This is the threshold for structural anomalies; S2.9 Summarize all identified structural anomaly events to form a state event set. ; S2.
10. Based on unified timestamps and spatial region identifiers, analyze video event sets. and state event set After spatiotemporal alignment and fusion, an edge event data set is generated. ; S3. Construct a data priority function based on the edge data and edge event data set, and use the phase space critical manifold to perform topology locking optimization on the priority to generate a cloud data set; This includes the following steps: S3.
1. Perform timestamp alignment and spatial correlation matching on edge data and edge event data sets, and construct a fused data set based on a unified indexing mechanism; S3.
2. Based on the edge event dataset, edge events are classified into different risk levels; S3.3, For each piece of fused data in the fused data set To construct its comprehensive value factors; S3.
4. Generate a data priority function based on a comprehensive value factor and by introducing a resource constraint function; S3.
5. By constructing a phase space critical manifold with change rate characteristics, and implementing priority topology locking on the structural monitoring data within the phase space critical manifold, the data can maintain a stable priority under critical conditions, and an optimized priority can be generated. S3.51, Based on the characteristics of asymmetric rate of change Construct a normalized rate of change vector ; S3.52, Based on normalized rate of change vector The infinite norm is used to reflect the overall anomaly intensity, and a critical state region is constructed. , This is the critical bandwidth parameter; S3.53, Based on the critical state region For structural monitoring data, a locked state variable is constructed. To achieve hysteresis control: ; S3.
54. During the evolution of the phase space region attribution status of structural monitoring data over time, monitor and lock the state variables. Changes; S3.55, Based on locked state variables The changes in the results are used to construct a piecewise priority function to generate the optimized priority. ,when At this time, the data is in a "perpetually critical" state, and the priority remains frozen. , For the first fused data at any time Data priority function value, ,in, To indicate the first The comprehensive value factors of integrated data, As a global risk factor, For risk modulation coefficients, For resource constraint functions, This is a stability correction constant; S3.6 Sort all fused data in descending order according to the optimized priority, generate a scheduling sequence, and build a real-time scheduling queue. Based on scheduling queues Filter out the cloud data set; S4. Based on cloud-based data sets, utilize cloud platforms to perform data fusion and form governance datasets; S5. Construct a safety status vector based on the governance dataset, and build a multi-indicator fusion risk assessment model based on the safety status vector to comprehensively assess and classify the structural safety status of water conservancy and hydropower projects.
2. The method for managing safety data in water conservancy and hydropower projects based on cloud-edge collaborative computing as described in claim 1, characterized in that, In step S1, multi-source safety monitoring data of water conservancy and hydropower projects are collected, and edge nodes are used to preprocess the multi-source safety monitoring data to generate edge data, including the following steps: Collect structural monitoring data of water conservancy and hydropower projects, including at least displacement, seepage pressure value, crack width, and tilt angle data; simultaneously collect on-site visual data to form a multi-source safety monitoring dataset; preprocess the multi-source safety monitoring dataset, and uniformly encapsulate the preprocessed multi-source safety monitoring data to generate edge data.
3. The method for managing safety data in water conservancy and hydropower projects based on cloud-edge collaborative computing as described in claim 1, characterized in that, In step S2, a lightweight convolutional neural network is used based on edge data to identify abnormal events through video object detection and generate a video event set, including the following steps: S2.1 Extract on-site visual data from edge data, decode the on-site visual data at edge nodes, perform frame extraction at fixed time intervals to generate a video frame sequence, and preprocess the video frame sequence. S2.2 Perform target detection inference on the preprocessed video frames based on a lightweight convolutional neural network, and output a candidate target set. ; S2.
3. For target detection inference in video frames, a lightweight target association mechanism is introduced to construct a set of target cross-frame trajectories. And based on the target cross-frame trajectory set Extract the temporal behavioral features of the target to form an enhanced target state representation. ; S2.4, Based on enhanced target state representation Perform anomaly identification and generate anomaly events; S2.
5. Summarize all identified abnormal events to form a video event set. .
4. The method for managing safety data in water conservancy and hydropower projects based on cloud-edge collaborative computing as described in claim 1, characterized in that, In step S4, data fusion is performed using a cloud platform based on the cloud-based data set to form a governance dataset, including the following steps: S4.1 Upload the cloud data collection to the cloud platform to uniformly access cloud data from different edge nodes; S4.2 To address the discrepancy between the timestamp and spatial location of cloud data, a spatiotemporal correction function is constructed. Through time alignment interpolation and spatial coordinate normalization, spatiotemporal consistency correction is performed on the cloud data, and a spatiotemporally consistent dataset is output. S4.
3. Based on the spatiotemporal consistency dataset, a unified governance representation is generated through a weighted fusion mechanism, and the unified governance representation is encapsulated for governance to generate a structured governance dataset.
5. The method for managing safety data in water conservancy and hydropower projects based on cloud-edge collaborative computing according to claim 1, characterized in that, In step S5, a safety status vector is constructed based on the governance dataset, and a multi-indicator fusion risk assessment model is built based on the safety status vector to comprehensively assess and classify the structural safety status of water conservancy and hydropower projects, including the following steps: S5.
1. Based on the governance dataset, feature decoupling and index extraction are performed using a feature separation network to construct a multi-dimensional safety index system, including structural stress indexes. Displacement and deformation index seepage pressure index Crack development indicators Furthermore, physical quantity normalization processing is performed on different indicators to obtain standardized indicator vectors. ; S5.2, Based on Standardized Indicator Vectors Construct structural safety state vector ; S5.3, Based on structural safety state vector Construct a multi-indicator integrated risk assessment model to generate a comprehensive risk value; S5.
4. The comprehensive risk value is compared and divided with the preset classification threshold set to comprehensively assess and classify the structural safety status of water conservancy and hydropower projects, and the structural safety classification results are output for real-time early warning and decision support of water conservancy and hydropower projects.
6. A water conservancy and hydropower engineering safety data governance system based on cloud-edge collaborative computing, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the water conservancy and hydropower engineering safety data governance method based on cloud-edge collaborative computing as described in any one of claims 1-5.
Citation Information
Patent Citations
Grid operation risk early warning method, system and device based on power grid cloud edge cooperation and medium
CN120855654A
Hydraulic structure monitoring and risk control system based on cloud edge collaborative architecture
CN121879156A
