Water conservancy project state diagnosis and prediction method based on multi-dimensional features and space-time integration

Abstract

The application discloses a water conservancy project state diagnosis and prediction method based on multi-dimensional features and space-time integration, relates to the technical field of water conservancy projects, and comprises the following steps: environment quantity monitoring equipment and effect quantity monitoring equipment are arranged in advance at key positions of a water conservancy project structure, environment quantity data and effect quantity data are synchronously collected, a raw monitoring data set with a time stamp is generated, the raw monitoring data set is subjected to data cleaning, and a standardized original time sequence is constructed. The application realizes accurate positioning of abnormalities, automatic identification of fault types and intelligent division of early warning levels by introducing a health index HI quantitative evaluation and graph convolution spatial topology diagnosis, can provide clear structure abnormal positions, deterioration trends and targeted maintenance suggestions for operation and maintenance personnel, realizes early identification of early hidden dangers and slow deterioration trends, realizes online management and control of the whole life cycle of the operation state of the water conservancy project, and effectively improves the engineering safety guarantee capability and the intelligent management level.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering technology, and more specifically, to a method for water conservancy engineering state diagnosis and prediction based on multidimensional features and spatiotemporal integration. Background Technology

[0002] my country's water conservancy projects have long been operating in complex natural environments, subject to water level fluctuations, temperature changes, geological settlement, and the coupled effects of long-term loads. Structures such as dams, dikes, and sluices are prone to safety hazards such as displacement, seepage, and stress variations. Traditional monitoring methods primarily rely on single-effect threshold judgments, making it difficult to eliminate interference from environmental factors and leading to false alarms and missed alarms. Furthermore, they depend on manual inspections and experience-based analysis, failing to predict structural deterioration trends in advance and thus failing to meet the real-time requirements for safety management throughout the entire lifecycle of the project.

[0003] Existing diagnostic technologies for water conservancy projects often focus on time series analysis or single-dimensional mining of spatial distribution features, which has significant limitations: relying solely on time series models cannot identify sensor failures and local structural anomalies, and spatial topological comparisons alone are insufficient to quantify the overall health and evolution trend of the structure. Furthermore, long-term monitoring data suffers from high redundancy and computational complexity, making it difficult for conventional algorithms to achieve efficient online prediction. The lack of spatiotemporal feature fusion mechanisms and standardized health assessment systems leads to inaccurate fault location, ambiguous early warning level classifications, and an inability to provide precise decision support for maintenance and repair.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a method for diagnosing and predicting the state of water conservancy projects based on multidimensional features and spatiotemporal integration, in order to overcome the aforementioned technical problems existing in existing related technologies.

[0006] The technical solution of this invention is implemented as follows:

[0007] A method for diagnosing and predicting the condition of hydraulic engineering projects based on multidimensional features and spatiotemporal integration includes the following steps:

[0008] Environmental and effect monitoring equipment is pre-installed at key locations on the hydraulic engineering structure to simultaneously collect environmental and effect data, generating a timestamped raw monitoring dataset. This raw dataset is then cleaned to construct a standardized raw time series. ;

[0009] Variational Mode Decomposition (VMD) algorithm is used to process the standardized original time series. Adaptive decomposition yields K intrinsic mode components with different center frequencies. This is used to achieve noise separation and feature deconstruction.

[0010] Using each monitoring point on the hydraulic engineering structure as a graph node, a Gaussian kernel adjacency matrix is ​​constructed based on the physical distance between nodes to extract the spatial topological association features of the monitoring points; the consistency between the features of each node and the features of its topological neighboring nodes is compared, and monitoring points that deviate from the threshold range are marked as anomalies, and the spatial diagnosis results are finally output.

[0011] The trend component u obtained from the decomposition trend Input the improved Informer model to complete the long-term time-series operation trend prediction of water conservancy projects, and output the predicted values ​​for future periods. ;

[0012] Based on real-time monitoring values, predicted values, and environmental impact factors, dynamic residuals and health indices are calculated, and online assessment of operational status is achieved according to preset health index thresholds.

[0013] Based on the spatial diagnostic results and health index, the system matches the knowledge base of water conservancy project faults to generate online diagnostic results.

[0014] Furthermore, the environmental data includes at least water level, water temperature, air temperature, precipitation, and seepage pressure; the effect data includes at least dam / structure displacement, settlement, seepage flow, stress, strain, and vibration parameters.

[0015] Furthermore, the noise separation and feature deconstruction are achieved by the following steps:

[0016] The restricted variational function is pre-constructed, and its objective function is expressed as:

[0017] ;

[0018] The calibration constraints are expressed as follows:

[0019] ;

[0020] In the formula, The k-th intrinsic mode component contains the trend component u. trend Periodic component u cycle Noise component u noise ; The center frequency of the k-th modal component; * represents the Dirac function; ∗ represents the convolution operation; is the first-order partial derivative over time; K is the number of mode decompositions;

[0021] Perform component filtering and retain low-frequency trend components u trend With periodic component u cycle Remove high-frequency noise components u noiseThis completes data denoising and effective feature extraction.

[0022] Furthermore, the extraction of spatial topological association features of monitoring points includes the following steps:

[0023] Based on the physical distance between monitoring points Constructing an adjacency matrix , represented as:

[0024] ;

[0025] In the formula, n is the number of monitoring points, and σ is the Gaussian kernel width parameter. Let the connection weight between node i and node j be _____.

[0026] Introducing the self-connection identity matrix I yields the improved adjacency matrix. , represented as:

[0027] ;

[0028] Calculate the improved adjacency matrix degree matrix The diagonal elements satisfy:

[0029] ;

[0030] Computational graph convolution is represented as:

[0031] ;

[0032] In the formula, The feature matrix of the l-th layer graph convolution; Here is the learnable weight matrix for the l-th layer; It is a non-linear activation function; This is the normalized Laplace matrix.

[0033] Furthermore, the improved Informer model employs an encoder-decoder structure; the encoder includes an embedding layer, a ProbSparse self-attention layer, and a feedforward network layer, used to extract long-term temporal features of historical trend components; the decoder includes a masked self-attention layer, a cross-ProbSparse self-attention layer, a feedforward network layer, and a fully connected output layer, used to deduce the model's predicted values ​​for future time periods segment by segment. ;

[0034] Among them, ProbSparse self-attention uses L-log determinant sampling to achieve sparse computation, as follows:

[0035] ;

[0036] In the formula, These are the query, key, and value matrices, respectively. For sparse query matrix; The dimension of the input vector.

[0037] Furthermore, the dynamic residual is calculated and expressed as:

[0038] ;

[0039] In the formula, The measured monitoring value at time t; The model prediction value at time t; Let be the normalized variance of the environmental impact factor at time t; This is the environmental compensation coefficient.

[0040] Furthermore, the health index is calculated and expressed as:

[0041] ;

[0042] In the formula, Parameters for fitting the health index.

[0043] The beneficial effects of this invention are:

[0044] 1. This invention effectively solves the problem of traditional single-threshold judgment being easily affected by operating conditions by using spatiotemporal dual-dimensional fusion diagnosis and adaptive correction of environmental factors. It can accurately distinguish between real structural anomalies, sensor failures, and overall environmental fluctuations, significantly reducing false alarm and false negative rates. At the same time, it uses VMD decomposition combined with an improved Informer model to achieve long-term time-series trend prediction, optimizing the computational complexity to O(LlnL). Minute-level online updates can be completed on ordinary industrial control equipment, greatly improving prediction efficiency and engineering applicability.

[0045] 2. This invention introduces quantitative assessment of the Health Index (HI) and graph convolutional spatial topology diagnosis to achieve precise anomaly location, automatic fault type identification, and intelligent classification of early warning levels. It can provide maintenance personnel with clear structural anomaly locations, deterioration trends, and targeted maintenance suggestions, enabling early identification of hidden dangers and slow deterioration trends. This allows for online management and control of the entire life cycle of water conservancy project operation status, effectively improving the project's safety assurance capabilities and intelligent management level. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the water conservancy project state diagnosis and prediction method based on multidimensional features and spatiotemporal integration according to an embodiment of the present invention. Detailed Implementation

[0048] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0049] According to embodiments of the present invention, a method for diagnosing and predicting the state of water conservancy projects based on multidimensional features and spatiotemporal integration is provided.

[0050] like Figure 1 As shown, the hydraulic engineering condition diagnosis and prediction method based on multidimensional features and spatiotemporal integration according to an embodiment of the present invention includes the following steps:

[0051] Step S1: Collect multi-source monitoring data of water conservancy projects in advance, perform missing value filling, outlier removal and normalization processing to obtain the normalized monitoring time series. ;

[0052] This technical solution deploys environmental and effect quantity monitoring equipment at key structural components of hydraulic engineering projects. Environmental monitoring equipment includes radar water level gauges, temperature and humidity sensors, rain gauges, and piezometers to collect real-time data on water level, water temperature, air temperature, precipitation, and seepage pressure. Effect quantity monitoring equipment includes GNSS displacement stations, hydrostatic levels, seepage flow meters, stress-strain gauges, and vibration sensors to collect real-time data on dam / structure displacement, settlement, seepage flow, stress, strain, and vibration parameters. The equipment is evenly distributed across the structural cross-sections, with increased density at key locations, enabling synchronous and timed data acquisition. Timestamped monitoring data is uploaded to the system via wired or wireless means. The raw data undergoes sequential processing including missing value filling, outlier removal, and normalization to construct a one-dimensional original time series of displacement, settlement, seepage flow, and stress effect quantity data at equal intervals. This data is then used as input for step S2 variational mode decomposition to complete denoising and feature deconstruction.

[0053] Specifically, the collected data undergoes missing value imputation, outlier removal, and normalization, as shown below:

[0054]

[0055] In the formula, x represents the original monitoring data. The minimum value of the data. For the maximum value of the data, The data is normalized and its value range is [0,1].

[0056] Step S2: The variational mode decomposition (VMD) algorithm is used to process the standardized original time series. Adaptive decomposition yields K intrinsic mode components with different center frequencies. This achieves noise separation and feature deconstruction, as detailed below:

[0057] Step S201: Construct a constrained variational problem whose objective function is to minimize the sum of the bandwidths of the demodulated signals of each modal component, expressed as:

[0058] ;

[0059] Its constraint is that the sum of all modal components equals the original signal, expressed as:

[0060] ;

[0061] In the formula, The k-th intrinsic mode component contains the trend component u. trend Periodic component u cycle Noise component u noise ; The center frequency of the k-th modal component; * represents the Dirac function; ∗ represents the convolution operation; is the first-order partial derivative over time; K is the number of mode decompositions, which is taken as 4~8 according to the characteristics of water conservancy monitoring data.

[0062] Step S202: Perform component filtering and retain low-frequency trend components u. trend With periodic component u cycle Remove high-frequency noise components u noise This completes data denoising and effective feature extraction.

[0063] Step S3: Using each monitoring point on the hydraulic engineering structure as a graph node, construct a Gaussian kernel adjacency matrix based on the physical distance between nodes, introduce a self-connected identity matrix and calculate a normalized Laplacian matrix, and extract the spatial topological association features of the monitoring points through an improved graph convolutional neural network; compare the consistency between the features of each node and the features of its topological neighboring nodes, mark the monitoring points that deviate from the threshold range as anomalies, and finally output the spatial diagnosis results.

[0064] Specifically, in step S301, based on the physical distance between monitoring points... Constructing an adjacency matrix , represented as:

[0065] ;

[0066] In the formula, n is the number of monitoring points, and σ is the Gaussian kernel width parameter. Let be the connection weight between node i and node j.

[0067] Step S302: Introduce the self-connected identity matrix I to obtain the improved adjacency matrix. , represented as:

[0068] ;

[0069] Step S303: Calculate the improved adjacency matrix degree matrix The diagonal elements satisfy:

[0070] ;

[0071] Step S304, calculate graph convolution, expressed as:

[0072] ;

[0073] In the formula, The feature matrix of the l-th layer graph convolution; Here is the learnable weight matrix for the l-th layer; The ReLU function is used as the non-linear activation function. This is the normalized Laplace matrix.

[0074] This technical solution, after obtaining the spatial feature vectors of each monitoring node through an improved graph convolutional neural network, performs spatial feature consistency discrimination. Specifically, it calculates the cosine similarity or Euclidean distance between the feature vector of each node and the mean feature vectors of its topological neighbors, and compares this value with a preset discrimination threshold, as detailed below:

[0075] If the calculated similarity is greater than or equal to the preset threshold, it indicates that the change trend of the node is consistent with the overall coordination of the surrounding neighboring nodes, the spatial features match, and the spatial state is judged to be normal.

[0076] If the similarity is less than the preset threshold, or the Euclidean distance exceeds the threshold range, it indicates that the change of the node deviates significantly from the overall pattern of the neighborhood and the spatial characteristics are inconsistent. It is judged as a spatial anomaly and the measuring point is marked as a spatial anomaly point.

[0077] Based on the above comparison and judgment results, the final output of complete spatial diagnosis results includes: the spatial characteristic consistency judgment results of each monitoring point, the spatial anomaly measurement point number and the location of its structural part, the anomaly type identification results, and the overall spatial topological stability assessment results of the engineering structure.

[0078] Step S4, decompose the obtained trend components u trendInput the improved Informer model to complete the long-term time-series operation trend prediction of water conservancy projects, and output the predicted values ​​for future periods. ;

[0079] Specifically, the improved Informer model adopts an encoder-decoder structure; the encoder includes an embedding layer, a probSparse self-attention layer, and a feedforward network layer, used to extract long-term temporal features of historical trend components; the decoder includes a masked self-attention layer, a cross-probSparse self-attention layer, a feedforward network layer, and a fully connected output layer, used to deduce the model's predicted values ​​for future time periods segment by segment. ;

[0080] Among them, ProbSparse self-attention uses L-log determinant sampling to achieve sparse computation, with a complexity of O(n log n). L is the length of the time series;

[0081] Specifically, the calibrated standard self-attention mechanism is represented as:

[0082] ;

[0083] Using L-log determinant sampling to filter sparse query matrices Construct a probabilistic sparse self-attention mechanism, represented as:

[0084] ;

[0085] In the formula, These are query, key, and value matrices, respectively, composed of trend components u. trend Obtained by linear transformation; For sparse query matrix; The dimension of the input vector;

[0086] Step S5: Calculate the dynamic residual and health index based on real-time monitoring values, predicted values ​​and environmental impact factors, and conduct online assessment of the operating status according to the preset health index threshold.

[0087] Among them, the measured values ​​are calculated. Compared with the predicted value dynamic residual , represented as:

[0088] ;

[0089] In the formula, The measured monitoring value at time t; The model prediction value at time t; Let be the normalized variance of the environmental impact factor at time t; This is the environmental compensation coefficient, with a value ranging from 0.1 to 0.5.

[0090] The health index is calculated as follows:

[0091] ;

[0092] In the formula, The parameters for fitting the health index are calibrated using engineering measurement data.

[0093] Specifically, during implementation, a health index threshold is preset, and early warning rules are set based on the preset health index threshold, as follows:

[0094] like This indicates a healthy state;

[0095] like This indicates a sub-healthy state;

[0096] like , indicating an abnormal state.

[0097] Step S6: Based on the health index and spatial diagnosis results, match the knowledge base of water conservancy project faults to automatically generate decision information such as structural anomaly location, fault type determination, and maintenance suggestions, thereby realizing online diagnosis and prediction.

[0098] This technical solution integrates the spatial diagnosis results obtained in step S3 with the health index calculated in step S5 in a spatiotemporal dimension for judgment, matches it with a water conservancy project fault knowledge base, and automatically generates decision information such as structural anomaly location, fault type determination, and maintenance suggestions. The specific judgment and handling rules are as follows:

[0099] If the health status and spatial diagnosis are normal, the structure and monitoring system are both in normal condition, and the system continues to monitor at the regular frequency without any additional warnings.

[0100] If the health status and space diagnosis are abnormal, the structure itself has no safety hazards. The space abnormality is caused by sensor drift, measuring point failure, or data acquisition error. The system marks the measuring point equipment as abnormal and outputs suggestions for sensor calibration, measuring point inspection and replacement, without triggering a structural safety alarm.

[0101] If the sub-health status and space diagnosis are normal, and the problem is caused by fluctuations in environmental load or short-term changes in operating conditions, the system will issue a yellow warning, increase the monitoring frequency, and output suggestions for continuous observation and routine inspection.

[0102] If the sub-healthy state and spatial diagnosis show slight abnormalities, there is an early structural risk. The system will be upgraded to a yellow alert, key areas of concern will be marked, and suggestions for encrypted monitoring, special inspections, and short-term trend tracking will be provided.

[0103] If the abnormal state and spatial diagnosis are normal, the system determines that the structure is currently in an abnormal operating condition, with consistent changes at all monitoring points, no local damage or sensor failure characteristics, and the anomaly is caused by environmental factors such as sudden changes in water level, abnormal temperature, and flood season loads. The system does not determine it as a structural safety accident, but still records it as an abnormal state, triggering an orange alert and performing continuous tracking observation and environmental factor correlation analysis; if the health index continues to deteriorate, the alert level will be further upgraded.

[0104] If the abnormal status and spatial diagnosis are abnormal, it is determined to be a real structural abnormality. The system triggers a red emergency alarm, locates the abnormal area and fault type, and outputs immediate inspection, emergency response, and professional maintenance suggestions.

[0105] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Those skilled in the art, upon considering the disclosure in the specification and embodiments, will readily conceive of other embodiments of this disclosure. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0106] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for diagnosing and predicting the state of hydraulic engineering projects based on multidimensional features and spatiotemporal integration, characterized in that: Includes the following steps: Environmental and effect monitoring equipment is pre-deployed at key locations on the hydraulic engineering structure to simultaneously collect environmental and effect data, generating a timestamped raw monitoring dataset. This raw dataset is then cleaned to construct a standardized raw time series. ; Variational Mode Decomposition (VMD) algorithm is used to process the standardized original time series. Adaptive decomposition yields K intrinsic mode components with different center frequencies. This is used to achieve noise separation and feature deconstruction. Using each monitoring point on the hydraulic engineering structure as a graph node, a Gaussian kernel adjacency matrix is ​​constructed based on the physical distance between nodes to extract the spatial topological association features of the monitoring points. By comparing the consistency of the features of each node with the features of its topological neighboring nodes, the measurement points that deviate from the threshold range are marked as anomalies, and the spatial diagnosis results are finally output. The trend component u obtained from the decomposition trend Input the improved Informer model to complete the long-term time-series operation trend prediction of water conservancy projects, and output the predicted values ​​for future periods. ; Based on real-time monitoring values, predicted values, and environmental impact factors, dynamic residuals and health indices are calculated, and online assessment of operational status is achieved according to preset health index thresholds. Based on the spatial diagnostic results and health index, the system matches the knowledge base of water conservancy project faults to generate online diagnostic results.

2. The method for diagnosing and predicting the state of hydraulic engineering based on multidimensional features and spatiotemporal integration as described in claim 1, characterized in that, The environmental data includes at least water level, water temperature, air temperature, precipitation, and seepage pressure; the effect data includes at least dam / structure displacement, settlement, seepage flow, stress, strain, and vibration parameters.

3. The method for diagnosing and predicting the state of hydraulic engineering based on multidimensional features and spatiotemporal integration as described in claim 1, characterized in that, The process of achieving noise separation and feature deconstruction includes the following steps: The restricted variational function is pre-constructed, and its objective function is expressed as: ； The calibration constraints are expressed as follows: ； In the formula, The k-th intrinsic mode component contains the trend component u. trend Periodic component u cycle Noise component u noise ; The center frequency of the k-th modal component; * represents the Dirac function; ∗ represents the convolution operation; is the first-order partial derivative over time; K is the number of mode decompositions; Perform component filtering and retain low-frequency trend components u trend With periodic component u cycle Remove high-frequency noise components u noise This completes data denoising and effective feature extraction.

4. The method for diagnosing and predicting the state of hydraulic engineering based on multidimensional features and spatiotemporal integration according to claim 1, characterized in that, The extraction of spatial topological association features of monitoring points includes the following steps: Based on the physical distance between monitoring points Constructing an adjacency matrix , is represented as: ； In the formula, n is the number of monitoring points, and σ is the Gaussian kernel width parameter. Let be the connection weight between node i and node j; Introducing the self-connection identity matrix I yields the improved adjacency matrix. , is represented as: ； Calculate the improved adjacency matrix degree matrix The diagonal elements satisfy: ； Computational graph convolution is represented as: ； In the formula, The feature matrix of the l-th layer graph convolution; Here is the learnable weight matrix for the l-th layer; It is a non-linear activation function; This is the normalized Laplace matrix.

5. The method for diagnosing and predicting the state of hydraulic engineering based on multidimensional features and spatiotemporal integration according to claim 1, characterized in that, The improved Informer model adopts an encoder-decoder structure; the encoder includes an embedding layer, a ProbSparse self-attention layer, and a feedforward network layer, which are used to extract long-term temporal features of historical trend components. The decoder consists of a masked self-attention layer, a cross-ProbSparse self-attention layer, a feedforward network layer, and a fully connected output layer, used to deduce model predictions for future time periods segment by segment. ; Among them, ProbSparse self-attention uses L-log determinant sampling to achieve sparse computation, as follows: ； In the formula, These are the query, key, and value matrices, respectively. For sparse query matrix; The dimension of the input vector.

6. The method for diagnosing and predicting the state of hydraulic engineering based on multidimensional features and spatiotemporal integration according to claim 1, characterized in that, The dynamic residual is calculated and expressed as: ； In the formula, The measured monitoring value at time t; The model prediction value at time t; Let be the normalized variance of the environmental impact factor at time t; This is the environmental compensation coefficient.

7. The method for diagnosing and predicting the state of hydraulic engineering based on multidimensional features and spatiotemporal integration as described in claim 6, is characterized in that, The health index is calculated and expressed as follows: ； In the formula, Parameters for fitting the health index.

Images