Water conservancy facility operation and maintenance method based on multi-source data

Abstract

The invention relates to a water conservancy facility operation and maintenance method based on multi-source data, and the method comprises the steps: collecting multi-source heterogeneous data of a water conservancy facility, and obtaining a standardized multi-dimensional data matrix; multi-modal features of the standardized multi-dimensional data matrix are extracted through a graph neural network fusion algorithm, and a fusion feature vector is obtained; performing fault diagnosis on the fusion feature vector through a knowledge graph reasoning algorithm automatically constructed by a domain ontology to obtain fault diagnosis information; based on the fault diagnosis information, performing facility state evolution analysis through a space-time coupling degradation prediction model to obtain a space-time degradation prediction result; according to the space-time degradation prediction result, state evaluation of the water conservancy facilities is carried out through a dual-objective optimized dynamic weight distribution algorithm, and an operation and maintenance state evaluation result is obtained; and on the basis of the operation and maintenance state evaluation result, performing cross-basin cascade fault prevention and control on the water conservancy facilities to obtain a cascade fault prevention and control scheme. The method improves the prediction precision of the degradation of the water conservancy facilities, and enhances the anti-risk capability of the water conservancy infrastructure.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment control, and in particular to a water conservancy facility operation and maintenance method based on multi-source data. Background Art

[0002] Currently, the operation and maintenance management of water conservancy facilities primarily relies on condition monitoring and assessment technology based on individual facilities. This technology deploys sensor networks at each facility to collect operational data, uses machine learning algorithms such as support vector machines and random forests to identify faults, and predicts equipment conditions based on time series analysis methods such as ARIMA models and gray prediction models. Ultimately, maintenance recommendations are generated through expert systems or rule-based decision trees. This technology approach is relatively mature for monitoring and diagnosing individual facilities, enabling basic fault detection and condition assessment capabilities. It has been widely used in key facilities such as pumping stations and gates.

[0003] In recent years, some advanced systems have begun to introduce deep learning technologies such as CNN and LSTM for fault pattern identification, achieve quantitative assessment by building an equipment health indicator system, and establish a digital operation and maintenance platform covering a single river basin or management area.

[0004] However, existing facility status assessment technologies have fundamental limitations: First, existing technologies treat each water conservancy facility as an independent individual for modeling and analysis, ignoring the complex physical connections, functional dependencies and operational synergy between facilities within the basin. They are unable to identify and predict the cascading impact of a facility's status change on related facilities, leading to inaccurate predictions and improper resource allocation when facing regional operation and maintenance issues. Second, traditional prediction models mainly focus on the time series changes of individual facilities, ignoring the spatial correlation and degradation propagation effects between facilities. The prediction accuracy and scope of application are limited, and the assessment results have a low match with actual operation and maintenance needs. At the same time, each management unit adopts an independent assessment system and data standards, lacking a cross-domain coordination mechanism. When making large-scale operation and maintenance decisions involving multiple basins or management areas, it is impossible to achieve globally optimal resource allocation and risk prevention and control, and it is difficult to formulate an effective coordinated response strategy. Summary of the Invention

[0005] The present invention provides a water conservancy facility operation and maintenance method based on multi-source data, so as to solve the defects of the prior art.

[0006] The present invention provides a water conservancy facility operation and maintenance method based on multi-source data, comprising: S1: Collect multi-source heterogeneous data of water conservancy facilities and obtain a standardized multidimensional data matrix; S2: extracting multimodal features of the standardized multidimensional data matrix through a graph neural network fusion algorithm to obtain a fused feature vector; S3: Perform fault diagnosis on the fused feature vector through a knowledge graph reasoning algorithm automatically constructed by domain ontology to obtain fault diagnosis information; S4: Based on the fault diagnosis information, perform facility state evolution analysis using a spatiotemporal coupled degradation prediction model to obtain a spatiotemporal degradation prediction result; S5: Based on the spatiotemporal degradation prediction results, a state assessment of the water conservancy facilities is performed using a dynamic weight allocation algorithm of dual-objective optimization to obtain an operation and maintenance state assessment result; S6: Based on the operation and maintenance status assessment results, cross-basin cascading fault prevention and control is performed on the water conservancy facilities, a cascading fault prevention and control plan is obtained, and the water conservancy facilities are operated and maintained according to the cascading fault prevention and control plan.

[0007] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S1 further includes: S11: Use the sliding window smoothing algorithm to eliminate the noise of the sensor collected data and obtain the denoised time series data; S12: Filling missing values ​​in the denoised time series data to obtain complete time series data; S13: performing standardization processing on the complete time series data to obtain a standardized multidimensional data matrix.

[0008] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S2 further includes: S21: extracting the spatiotemporal features of the standardized multidimensional data matrix through an LSTM-CNN hybrid network to obtain spatiotemporal feature data; S22: Performing cross-modal feature mapping on the spatiotemporal feature data through a graph attention network to obtain high-dimensional feature space data; S23: Perform feature discrimination enhancement on the high-dimensional feature space data through a contrastive learning mechanism to obtain a fused feature vector.

[0009] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, in step S21, the LSTM-CNN hybrid network is optimized and trained through a federated learning framework, specifically comprising: S211: Establish a federated learning client, update the model parameters of the federated learning client through local gradient calculation, and obtain local model parameters; S212: Performing noise perturbation on the local model parameters through a differential privacy mechanism to obtain privacy protection parameters; S213: Based on the privacy protection parameters, global model aggregation is performed through a secure aggregation protocol to obtain an LSTM-CNN hybrid network for federated learning collaborative optimization.

[0010] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S3 further includes: S31: Using the BiLSTM-CRF model, we extract entity relationships from text data in the water conservancy industry and obtain a set of entity relationship triples. S32: vectorizing and embedding the entity relationship triple set to obtain a water conservancy facilities domain ontology framework; S33: Inputting the fused feature vector into the water conservancy facilities domain ontology framework to perform fault pattern recognition through entity linking reasoning to obtain fault diagnosis information.

[0011] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S31 further includes: S311: Select an operation and maintenance report text from a public corpus, perform word segmentation on the operation and maintenance report text using a word segmentation algorithm, and obtain a word sequence; S312: Perform context encoding on the vocabulary sequence through a BiLSTM encoder to obtain a vocabulary vector representation; S313: Identify the entity boundaries represented by the vocabulary vector through the CRF layer to obtain a set of entity relationship triples.

[0012] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S4 further includes: S41: performing degradation field modeling on the fault diagnosis information through a spatiotemporal partial differential equation solver to obtain a degradation propagation matrix; S42: performing spatial discretization on the degradation propagation matrix using a finite element analysis algorithm to obtain a gridded degradation distribution; S43: Calculating facility state evolution using a spatiotemporal coupling prediction algorithm based on the gridded degradation distribution to obtain a spatiotemporal degradation prediction result.

[0013] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S41 further includes: S411: Establish a facility degradation field, describe the degradation propagation mechanism through diffusion equation modeling, and obtain the diffusion control equation; S412: setting boundary conditions of the diffusion control equation to determine the solution domain and obtain a degenerate field solution model; S413: Perform degradation propagation calculation using the degradation field solution model to obtain a degradation propagation matrix.

[0014] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S5 further includes: S51: Based on the spatiotemporal degradation prediction result, cost risk modeling is performed using a multi-objective optimization function to obtain a dual-objective function model; wherein the multi-objective optimization function includes an economic index function and a safety index function; S52: Solving the Pareto optimal solution of the dual-objective function model to obtain a weight distribution coefficient; S53: In combination with the weight distribution coefficient and the water resource scheduling constraint condition, a state quantitative evaluation is performed through a fuzzy comprehensive evaluation algorithm to obtain an operation and maintenance state evaluation result.

[0015] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S6 further includes: S61: Modeling cross-basin dependency relationships based on the operation and maintenance status assessment results to obtain a cascading fault propagation diagram; S62: Select a fault blocking point from the cascade fault propagation graph to obtain a set of fault prevention and control nodes; S63: Based on the fault prevention and control node set, a cross-basin collaborative strategy is generated through a resource allocation optimization algorithm to obtain a cascading fault prevention and control solution.

[0016] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S61 further includes: S611: Collect the connection relationships of multi-basin facilities, construct the network topology through graph theory modeling algorithms, and obtain the cross-basin facility network diagram; S612: Analyze the propagation path of the cross-basin facility network diagram using a fault propagation dynamics model to obtain a fault diffusion matrix; S613: Perform cascading effect prediction based on the fault diffusion matrix to obtain a cascading fault propagation graph.

[0017] According to a water conservancy facility operation and maintenance method based on multi-source data provided by the present invention, step S62 further includes: S621: Perform key node measurement based on the cascading fault propagation graph to obtain a node importance ranking; S622: Based on the node importance ranking, a blocking point selection is performed using a fault blocking effect evaluation algorithm to obtain candidate prevention and control nodes; S623: Determine the optimal control point based on the candidate control nodes using a cost-benefit analysis algorithm to obtain a set of fault control nodes.

[0018] The present invention also provides a water conservancy facility operation and maintenance system based on multi-source data, comprising: Acquisition module: used to collect multi-source heterogeneous data of water conservancy facilities and obtain standardized multi-dimensional data matrix; Extraction module: used to extract multimodal features of the standardized multidimensional data matrix through a graph neural network fusion algorithm to obtain a fused feature vector; Diagnosis module: used to perform fault diagnosis on the fused feature vector through the knowledge graph reasoning algorithm automatically constructed by the domain ontology to obtain fault diagnosis information; Analysis module: used to perform facility state evolution analysis based on the fault diagnosis information through a spatiotemporal coupled degradation prediction model to obtain spatiotemporal degradation prediction results; Evaluation module: used to evaluate the status of water conservancy facilities based on the spatiotemporal degradation prediction results through a dynamic weight allocation algorithm of dual-objective optimization to obtain an operation and maintenance status evaluation result; Operation and maintenance module: used to carry out cross-basin cascading fault prevention and control of water conservancy facilities based on the operation and maintenance status assessment results, obtain a cascading fault prevention and control plan, and operate and maintain water conservancy facilities through the cascading fault prevention and control plan.

[0019] The present invention provides a water conservancy facility operation and maintenance method and system based on multi-source data, which collects and processes multi-source heterogeneous data through a distributed sensor network and combines the coordinated application of a sliding window smoothing algorithm, an interpolation algorithm and anomaly detection algorithm, which can significantly improve data quality and integrity, lay a solid foundation for data analysis, and compared with the traditional single data source collection method, the data reliability is improved, and the problems of data missing and noise interference in water conservancy facility monitoring are effectively solved; secondly, the present invention uses an innovative combination of a graph neural network fusion algorithm and an LSTM-CNN hybrid network, especially through a graph attention network for cross-modal feature mapping and a contrastive learning mechanism for feature discrimination enhancement, so that the efficiency of multimodal data fusion is improved, the feature characterization capability is enhanced, and the technical bottleneck that traditional methods cannot effectively process heterogeneous data is broken through; thirdly, the knowledge graph inference algorithm for automatic construction of the domain ontology of the present invention uses a deep fusion application of entity relationship extraction by a BiLSTM-CRF model and vectorized embedding by a TransE algorithm, which is effective. The dynamic construction and adaptive update of the knowledge graph are realized, and the accuracy of knowledge extraction and reasoning efficiency are improved compared with the traditional expert system, which effectively alleviates the problems of high cost and difficult updating of manual ontology construction; in addition, the collaborative work of the spatiotemporal partial differential equation solver and the finite element analysis algorithm in the spatiotemporal coupled degradation prediction model of the present invention innovatively models facility degradation as a spatiotemporal propagation field, greatly improves the prediction accuracy, and expands the prediction time span, realizing accurate modeling of the propagation path of degradation in the facility network; it also models facility dependencies through a complex network analysis algorithm through a cross-basin cascading fault prevention and control algorithm, combines the key node identification algorithm and the resource allocation optimization algorithm to achieve reduced cascading fault risk and shortened emergency response time; the present invention also uses a deep integration application of the federated learning framework and the differential privacy mechanism throughout, and realizes multi-party collaborative optimization through a secure aggregation protocol, which improves model performance and collaborative efficiency while protecting data privacy, solving the data island and privacy leakage risks of traditional centralized learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 A schematic flow chart of a water conservancy facility operation and maintenance method based on multi-source data provided by an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a water conservancy facility operation and maintenance system based on multi-source data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0023] like Figure 1 As shown, the present invention provides a water conservancy facility operation and maintenance method based on multi-source data, comprising: S1: Collect multi-source heterogeneous data of water conservancy facilities and obtain a standardized multidimensional data matrix.

[0024] Furthermore, in step S1, the present invention conducts all-round monitoring of water conservancy facilities through a distributed sensor network. The distributed sensor network is a sensor cluster deployed at different locations such as pump stations, gates, channels, and reservoirs, including flow sensors, water level sensors, pressure sensors, vibration sensors, temperature and humidity sensors, and other types. The collected multi-source heterogeneous data comes from different data sources and is monitoring data with different data formats, sampling frequencies, and data types. The heterogeneity is reflected in the differences in data structures, such as flow data is a continuous numerical type, equipment status is a discrete label type, and image data is a matrix type. The standardized multidimensional data matrix finally obtained is a multidimensional array structure that converts all heterogeneous data into a unified format, in which each row represents an observation sample with a timestamp, each column represents a feature dimension, and the numerical range of the matrix elements is unified within the interval [0,1] after normalization.

[0025] Wherein, step S1 further includes: S11: Use the sliding window smoothing algorithm to eliminate the noise of the sensor collected data and obtain denoised time series data.

[0026] Furthermore, the sliding window smoothing algorithm slides on the time series data by setting a time window of fixed length. The data points in the window are smoothed by weighted averaging calculation, which is specifically implemented through a Gaussian filter. The final denoised time series data retains the main trend characteristics of the data by eliminating high-frequency noise components and abnormal fluctuations.

[0027] S12: Fill missing values ​​in the denoised time series data to obtain complete time series data.

[0028] Furthermore, the linear interpolation in step S12 is for short-term missing values. Specifically, the present invention uses cubic spline interpolation for long-term missing values. By constructing a piecewise cubic polynomial function for fitting, the complete time series data finally obtained ensures that each time point has a corresponding observation value, thereby ensuring the integrity of the data.

[0029] S13: performing standardization processing on the complete time series data to obtain a standardized multidimensional data matrix.

[0030] Furthermore, the anomaly detection algorithm in step S13 adopts the Z-score statistical method to identify abnormal data points. For the identified data points, the present invention adopts the Min-Max normalization method to perform standardization processing. The processed standardized multidimensional data matrix forms a regular matrix structure by aligning each feature dimension according to the timestamp.

[0031] S2: Extract the multimodal features of the standardized multidimensional data matrix through a graph neural network fusion algorithm to obtain a fusion feature vector.

[0032] Furthermore, in step S2 of the present invention, the multimodal features of the standardized multidimensional data matrix are extracted through a graph neural network fusion algorithm. The graph neural network fusion algorithm refers to the combination of the traditional graph neural network architecture and multimodal data processing technology, which is specifically used to process heterogeneous data from different sensor types in the present invention. Specifically, in step S2, the standardized multidimensional data matrix output by step S1 is first received as input. Each element value in the matrix has been normalized and is located in the interval [0,1]. After the standardized multidimensional data matrix is ​​subjected to multi-layer feature extraction and fusion processing, a high-dimensional vector representation is obtained. The obtained high-dimensional vector representation contains comprehensive information on the operating status of the water conservancy facility.

[0033] Wherein, step S2 further includes: S21: Extracting the spatiotemporal features of the standardized multidimensional data matrix through an LSTM-CNN hybrid network to obtain spatiotemporal feature data.

[0034] Furthermore, in step S21, the present invention extracts the spatiotemporal features of the standardized multidimensional data matrix through an LSTM-CNN hybrid network. The LSTM-CNN hybrid network is a deep learning architecture that combines a long short-term memory network and a convolutional neural network in series, and is used to simultaneously capture the long-term dependencies and local spatial patterns of time series. Among them, the LSTM component first processes the sequence information of the time dimension, and inputs the feature vector of each time step into the LSTM unit. The LSTM unit contains three gating mechanisms: a forget gate, an input gate, and an output gate. The forget gate determines which information to discard from the cell state through a sigmoid activation function, the input gate calculates new candidate information, and the output gate controls the output of the cell state. After processing, the LSTM outputs a hidden state sequence. The CNN component then receives the hidden state sequence output by the LSTM, slides on the time dimension through a one-dimensional convolution kernel, extracts the spatial feature pattern for each local time window, reduces the dimension through a maximum pooling layer after convolution, and finally outputs the spatiotemporal feature data.

[0035] In step S21, the LSTM-CNN hybrid network is optimized and trained through a federated learning framework, specifically including: S211: Establish a federated learning client, update the model parameters of the federated learning client through local gradient calculation, and obtain local model parameters.

[0036] Furthermore, in the process of establishing a federated learning client, the present invention uses each water conservancy management unit as an independent client node to locally store the water conservancy facility monitoring data within the jurisdiction of the unit. After the client is established, the client receives the global model parameters and uses the local data set to first perform forward propagation calculations, and then performs backpropagation calculations. During the backpropagation process, the gradient of the loss function with respect to the parameters of each layer is calculated using the chain rule. Subsequently, a small batch stochastic gradient descent algorithm is used to perform local gradient calculations to perform gradient update formulas. After the client completes multiple rounds of local training, the model parameter update amount is calculated and finally the update amount is passed to the federated server as the local model parameter. The output local model parameters include the numerical changes of all trainable parameters such as the convolutional layer weights, LSTM weight matrix, and bias vector.

[0037] S212: Perform noise perturbation on the local model parameters through a differential privacy mechanism to obtain privacy protection parameters.

[0038] Furthermore, in step S212, the present invention first calculates the L2 norm bounds of the model parameters, limits the parameter updates to a preset threshold through gradient clipping, and then adds Laplace noise to the clipped parameters. Specifically, independent and identically distributed Laplace noise is first added to each model parameter. The noise intensity is dynamically adjusted based on the privacy budget. The smaller the privacy budget, the greater the added noise and the higher the privacy protection level. The resulting privacy-preserving parameters are the set of model parameters after adding noise, while maintaining the dimensional structure of the original parameters.

[0039] S213: Based on the privacy protection parameters, global model aggregation is performed through a secure aggregation protocol to obtain an LSTM-CNN hybrid network for federated learning collaborative optimization.

[0040] Furthermore, in step S213, the present invention first performs identity authentication and survival checks on all clients to ensure that the number of clients participating in the aggregation meets the minimum threshold requirement. After the aggregation server collects all privacy protection parameters, it uses a weighted average method to calculate the global model parameters. During the aggregation process, the present invention uses cryptographic technology to prevent the server from snooping on the parameter information of a single client, and implements the parameter summation operation in the ciphertext state through the homomorphic encryption algorithm. After the aggregation is completed, the server distributes the new global model parameters to each client, completing a round of federated learning training. The LSTM-CNN hybrid network finally obtained by the collaborative optimization of federated learning is the global model that converges after multiple rounds of federated training. The model integrates the local data characteristics of each client while protecting the privacy of the original data.

[0041] S22: Perform cross-modal feature mapping on the spatiotemporal feature data through a graph attention network to obtain high-dimensional feature space data.

[0042] Furthermore, in step S22, the graph attention network first receives the spatiotemporal feature data output from step S21 and then constructs an adjacency matrix between facilities. The elements in the matrix represent the connectivity between the two facilities, including physical pipeline connections, functional dependencies, and geographic proximity. Subsequently, in the graph attention network, the graph attention mechanism calculates the attention coefficient between each pair of nodes. The attention weights are calculated using the dot-product attention mechanism. Cross-modal feature mapping is implemented using a multi-head attention mechanism, in which each head independently calculates attention, and the outputs of all heads are then concatenated to ultimately output a high-dimensional feature space.

[0043] S23: Perform feature discrimination enhancement on the high-dimensional feature space data through a contrastive learning mechanism to obtain a fused feature vector.

[0044] In step S23, data enhancement is first performed on the high-dimensional feature space data input in step S22 to generate different perspective representations of the same facility. The enhancement methods include random masking of feature dimensions, addition of Gaussian noise, random sampling of time steps, etc. During the generation, for each facility node, positive and negative sample pairs are constructed, and then comparative learning is performed using the InfoNCE loss. The calculation formula is:

[0045] in, is the facility node index value, For the The contrastive learning loss value of samples, For the The original feature representation of samples, for The positive sample feature vector after data enhancement, For the The feature vector representation of different facility nodes is used as negative samples. Excluding own samples, is the temperature parameter, is the similarity function, and cosine similarity is used for calculation.

[0046] Feature discrimination enhancement is achieved by minimizing contrast loss, which makes different representations of the same facility closer in the feature space and the representations of different facilities more separated. The optimization process uses the Adam optimizer, the learning rate is set to 0.0001, and the batch size is 256. After training, the output of the feature encoder is extracted as the fused feature vector, where D is the dimension of the final feature representation. The fused feature vector contains the comprehensive operating status information of the water conservancy facilities and the correlation features between facilities.

[0047] S3: The fused feature vector is subjected to fault diagnosis by a knowledge graph reasoning algorithm automatically constructed by a domain ontology to obtain fault diagnosis information.

[0048] Wherein, step S3 further includes: S31: Use the BiLSTM-CRF model to extract entity relationships from text data in the water conservancy industry and obtain a set of entity relationship triples.

[0049] Wherein, step S31 further includes: S311: Select an operation and maintenance report text from a public corpus, perform word segmentation on the operation and maintenance report text using a word segmentation algorithm, and obtain a word sequence.

[0050] Furthermore, the operation and maintenance report text data is extracted from documents such as historical fault records, equipment maintenance reports, and abnormal event descriptions within the water conservancy management system. The text format is a raw string containing unstructured information such as the equipment name, fault symptoms, and treatment measures. In step S311 of the present invention, these raw text strings are received via a word segmentation algorithm and then preprocessed. After processing, a Chinese word segmentation method based on a combination of dictionary matching and statistical models is employed. The dictionary matching portion utilizes a water conservancy professional vocabulary library containing professional terms such as "water pump," "gate," "flow meter," and "pressure sensor." The matching process is performed according to the maximum matching principle, searching for the longest matching word in the dictionary from the beginning of the text. The statistical model portion utilizes a hidden Markov model, determining word segmentation boundaries by calculating transition and emission probabilities between characters. The final word segmentation result is output as a word sequence, comprising multiple segmented word units. The sequence length n varies according to the original text content, while the word sequence maintains its original semantic order.

[0051] S312: Perform context encoding on the vocabulary sequence through a BiLSTM encoder to obtain a vocabulary vector representation.

[0052] Furthermore, the word sequence outputted in step S311 is first received, and then each word is converted into a dense vector representation through a word embedding layer to obtain a word embedding matrix. Each word obtains its corresponding embedding vector through a table lookup operation. Specifically, the BiLSTM network comprises two components: a forward LSTM and a backward LSTM. The forward LSTM processes the word sequence from left to right in chronological order and calculates a forward hidden state sequence. The backward LSTM processes the word sequence from right to left in reverse order and calculates a backward hidden state sequence. The calculation process of the LSTM unit includes a forget gate, an input gate, candidate values, a cell state update, an output gate, and a hidden state. Finally, the BiLSTM concatenates the forward and backward hidden states to obtain a complete contextual representation and outputs a word vector representation sequence containing bidirectional contextual information for the word.

[0053] S313: Identify the entity boundaries represented by the vocabulary vector through the CRF layer to obtain a set of entity relationship triples.

[0054] Furthermore, the CRF conditional random field model is used for sequence labeling tasks, transforming entity recognition into a BIO labeling problem, where B represents the entity start position, I represents the internal position of the entity, and O represents the non-entity position. The CRF layer first maps the hidden state to the label space through a linear transformation, calculates the emission score matrix, and then stores the transition probabilities between different labels through the transition score matrix. The subsequent decoding process uses the Viterbi algorithm to find the optimal label sequence, and finally decodes and outputs a BIO label sequence. Entity boundaries are then extracted according to the BIO labeling rules, and continuous BIO sequences are identified as complete entities. The output entity relationship triple set is constructed through entity co-occurrence and dependency syntax analysis, including water conservancy facility entities, fault type entities, treatment measure entities, and their relationships.

[0055] S32: Vectorize and embed the entity relationship triple set to obtain an ontology framework for the water conservancy facilities domain.

[0056] Furthermore, the vectorized embedding process in step S32 aims to convert the symbolized triples into continuous vector representations using a knowledge graph embedding algorithm. Specifically, an entity vocabulary and a relationship vocabulary are first constructed. The entity vocabulary contains all entity names that appear, and the relationship vocabulary contains all relationship types. Then, the entities and relationships are mapped to the same vector space through the TransE algorithm. The core idea of ​​TransE is h+r≈t, where h is the head entity vector, r is the relationship vector, and t is the tail entity vector. Its loss function calculates the distance difference between positive triples and negative triples. Its training process optimizes the entity and relationship vectors through stochastic gradient descent, so that the h+r of the positive triple is as close to t as possible, and the h'+r of the negative triple is far away from t', where h' is the head entity vector of the negative sampling, and t' is the tail entity vector of the negative sampling. The final ontology framework for the water conservancy facilities field contains an entity vector matrix and a relationship vector matrix. Each row corresponds to a vector representation of an entity or relationship. The ontology framework also contains an entity-relationship adjacency matrix, which records the connection relationship and relationship type between entities. The matrix element value is the vector index of the corresponding relationship.

[0057] S33: Inputting the fused feature vector into the water conservancy facilities domain ontology framework to perform fault pattern recognition through entity linking reasoning to obtain fault diagnosis information.

[0058] Specifically, the fused feature vector is first input, and then the water conservancy facility domain ontology framework calculates the similarity between the fused feature vector and the entity vector in the ontology framework to obtain a similarity matrix, in which each element represents the degree of similarity between the facility and the entity. After obtaining the similarity matrix, the entity linking process is first performed. The candidate entity set corresponding to each facility is determined by threshold screening. In this embodiment, the threshold is set to 0.8, and entities with similarity greater than the threshold are selected as candidates. Secondly, graph reasoning is performed. The fault propagation path analysis is performed based on the entity-relationship graph structure in the ontology framework. The graph neural network is used for multi-hop reasoning. The reasoning process iteratively updates the node representation, aggregates neighbor node information to identify the fault mode, and the fault mode recognition is realized through the classifier. The final node representation is input into the fully connected layer, and the fault category probability distribution is output. The fault diagnosis information contains structured data such as fault type, severity, and impact range. The format is {facility ID: fault type, confidence: value, related facilities: list}. The confidence value range is [0,1]. The related facility list contains the associated facility numbers identified during the reasoning process.

[0059] S4: Based on the fault diagnosis information, a facility state evolution analysis is performed through a spatiotemporal coupled degradation prediction model to obtain a spatiotemporal degradation prediction result.

[0060] Wherein, step S4 further includes: S41: performing degradation field modeling on the fault diagnosis information through a spatiotemporal partial differential equation solver to obtain a degradation propagation matrix.

[0061] Wherein, step S41 further includes: S411: Establish a facility degradation field, describe the degradation propagation mechanism through diffusion equation modeling, and obtain the diffusion control equation.

[0062] In step S411, the present invention first converts the fault diagnosis information into a numerical degradation intensity matrix, where each element represents the initial degradation intensity value of the facility, which is calculated based on the fault type and confidence level. The calculation formula is: , is the weight value, determined by the fault type (0.2 for minor faults, 0.5 for moderate faults, and 1.0 for severe faults), is the confidence value; then the facility degradation field is established. During the establishment of the facility degradation field, the water conservancy facility network is mapped into a two-dimensional spatial domain. Each facility occupies a point position in space, and the physical connection relationship between facilities is represented by the spatial distance function; then the diffusion equation modeling is performed, and the parabolic partial differential equation is used to describe the propagation process of degradation in time and space. The basic form is:

[0063] in Indicates spatial location In time The degradation concentration of is the diffusion coefficient matrix, is the Laplace operator, is the source term function. The diffusion coefficient in the diffusion coefficient matrix is ​​determined according to the connection strength between facilities. The connection strength is comprehensively calculated through factors such as pipeline diameter, flow rate, and functional dependence. The source term function describes the generation and disappearance of degradation. A positive source term is set at the location of the faulty facility, and a negative source term is set at the location of the healthy facility. Diffusion control describes the change law of degradation concentration over time and space.

[0064] S412: Setting the boundary conditions of the diffusion control equation to determine the solution domain and obtain a degenerate field solution model.

[0065] Furthermore, the boundary conditions set include Dirichlet boundary conditions and Neumann boundary conditions. Dirichlet boundary conditions directly specify the function value on the boundary of the solution domain, which is set as ,in represents the boundary of the solution domain, The preset degradation concentration value on the boundary is determined according to the actual status of the boundary facilities. , fault boundary setting ; Neumann boundary conditions specify the value of the normal derivative on the boundary, set to ,in is the boundary normal vector, is the boundary flow density, which indicates the rate at which degradation flows into or out of the boundary. It is determined as a rectangular area containing all water conservancy facilities. The size of the area is determined according to the distribution range of the facilities. The initial conditions are set as The degradation concentration distribution at the moment is converted into a continuous initial concentration field through the interpolation method. The radial basis function method is used for interpolation. The degradation field solution model includes four components: partial differential equation body, boundary conditions, initial conditions and parameter settings, forming a complete initial boundary value problem. The mathematical expression is:

[0066] in is the prediction time window, represents the prediction time, For spatial location The original concentration, To find the solution domain, Indicates that in the solution domain Within and time interval Inside.

[0067] S413: Perform degradation propagation calculation using the degradation field solution model to obtain a degradation propagation matrix.

[0068] In step S413, the present invention performs a degenerate propagation calculation on the degenerate field solution model to solve it. The solution method selected is the finite difference method. The finite difference method discretizes the continuous space-time domain during the processing process, using a central difference format in the spatial direction and a forward Euler format in the temporal direction. Specifically, spatial discretization divides the solution domain into a regular grid, and temporal discretization divides the time interval into multiple time steps. The discretized partial differential equations are converted into a system of difference equations through central differences and forward differences of time derivatives. The specific form of the difference equations is:

[0069] in 、 are the x and y components of the diffusion coefficient matrix, is the spatial step length in the x direction, is the spatial step length in the y direction, is the grid point At the moment The concentration value of other similar parameter values ​​is the same as quite, is the source term at the grid point At the moment The value of the grid point In the The intensity of the degradation source at each time step represents the rate of degradation generation or elimination at that location. After differencing, the solution is performed through the time steps, starting from the initial conditions and gradually calculating the value of each time step. Finally, the degradation propagation matrix is ​​obtained, in which the elements represent the degradation concentration value at the spatial location at the time step. The matrix records the evolution of degradation in the entire space-time domain.

[0070] S42: spatially discretizing the degradation propagation matrix using a finite element analysis algorithm to obtain a gridded degradation distribution.

[0071] Furthermore, in step S42, the present invention divides the solution domain into a number of non-overlapping triangular or quadrilateral units using the finite element method. The degradation concentration in each unit is represented by interpolation of the node value and the shape function. The concentration distribution in the unit is expressed as:

[0072] in is the shape function, For nodes The concentration value is then solved by solving the linear equations Obtain the gridded degradation distribution, where is the inertia matrix, which is used to describe the capacity effect of the accumulated degradation concentration at each node and reflect the impedance characteristics of the degradation concentration changing with time. is the diffusion conduction matrix, which is used to describe the impedance effect of degradation propagation between adjacent nodes and reflects the resistance of degradation propagation in space. is the node concentration vector, is the concentration time derivative vector, For the load vector, including the contribution of source terms and boundary conditions, gridded degradation distribution data are generated, including the node coordinate matrix and the node concentration matrix, which records the degradation concentration value of each node at each time.

[0073] S43: Calculating facility state evolution using a spatiotemporal coupling prediction algorithm based on the gridded degradation distribution to obtain a spatiotemporal degradation prediction result.

[0074] In step S43, the gridded degradation distribution data is used as input to calculate the facility state evolution using a spatiotemporal coupled prediction algorithm. This algorithm combines time series prediction with spatial interpolation methods. It first performs trend analysis on the concentration time series of each facility node. It then uses the ARIMA model to predict the trend component in time series. Simultaneously, a Kriging interpolation algorithm performs spatial prediction based on a semivariogram model. The final facility state evolution calculation integrates the temporal prediction results and spatial interpolation results, dynamically adjusting them based on the prediction error. The output spatiotemporal degradation prediction results are formatted as a facility state matrix. The matrix elements represent the degradation level of the facility at the prediction time, with values ​​ranging from [0, 1], where 0 represents complete health and 1 represents complete failure.

[0075] S5: Based on the spatiotemporal degradation prediction results, a status assessment of the water conservancy facilities is performed using a dynamic weight allocation algorithm of dual-objective optimization to obtain an operation and maintenance status assessment result.

[0076] Wherein, step S5 further includes: S51: Based on the spatiotemporal degradation prediction result, cost risk modeling is performed through a multi-objective optimization function to obtain a dual-objective function model; wherein the multi-objective optimization function includes an economic index function and a safety index function.

[0077] Furthermore, in step S51, the spatiotemporal degradation prediction result of the above steps is used as input data, where each element Display facilities At the time of prediction The degradation degree value. The economic index function construction process adopts multi-level nonlinear coupling modeling. First, the degradation propagation cost between facilities is calculated, considering the amplification effect of cascading failures on economic losses. The formula is:

[0078] in, and All are facility index values. For facilities At the moment The cascading degradation cost of For facilities Basic maintenance costs, is the degradation cost index coefficient, For facilities Neighborhood facilities collection, For facilities and The mutual influence coefficient between For facilities At the moment The degradation degree value of is the steepness parameter of the sigmoid function, is the degradation threshold parameter.

[0079] Secondly, the dynamic maintenance cost of the present invention takes into account the time window effect and resource competition, adopts a time-varying price model, and is expressed as follows:

[0080] in, For facilities At the moment Dynamic maintenance cost, For facilities The unit maintenance cost, is the seasonal price fluctuation coefficient, is the price cycle, For facilities At the moment The intensity of maintenance requirements, is the total maintenance resource, is the resource competition index, is the degradation complexity coefficient.

[0081] Downtime penalty takes into account network topology and time-dependent service demands:

[0082] in For facilities At the moment downtime losses, is the service area index value, is the number of service areas, For the region At the moment service needs, For facilities For the region Service contribution, For facilities To area The set of facilities on the service path, is the repair efficiency parameter, is the repair time function.

[0083] The economic index function combines the above and includes the cascade degradation cost , dynamic maintenance cost , downtime losses The multiple cost components of the , and introduce time discounting and risk adjustment:

[0084] in, is the economic index function, is the prediction time window, is the discount rate, is the total number of facilities, is the risk adjustment factor, For the moment The variance of the degradation degree of all facilities, For the moment The mean degradation level of all facilities.

[0085] The safety index function adopts a multi-dimensional risk coupling model, taking into account the spatiotemporal correlation and environmental factors of facility failure:

[0086] in, For facilities At the moment safety indicators, For facilities The repair efficiency parameter, is the environmental sensitivity coefficient, For the moment The comprehensive environmental factors For facilities Sensitivity to changes in environmental factors, For facilities About facilities The impact weight of the failure probability.

[0087] The severity of accident consequences takes into account multiple impact paths and time cumulative effects:

[0088] in, For facilities At the moment The severity index of accident consequences, is the impact type index value, including personnel safety type, environmental pollution type, and economic loss type. is the number of impact types, For the The weight of the impact, For facilities For the first The basic severity of the impact, For facilities and facilities The influence propagation coefficient between For facilities and facilities The distance between For recovery time, is the time decay parameter, For the Effects over time intensity function.

[0089] The safety index function integrates the failure probability and consequence severity, and considers system resilience:

[0090] in, is a safety indicator, is the system resilience function, is the system vulnerability amplification factor.

[0091] The dual-objective function model also introduces constraint penalties to deal with capacity limitations, budget constraints, and technical constraints:

[0092] in is the inequality constraint index value, is the number of inequality constraints, is the equality constraint index value, is the number of equality constraints, and are the first penalty coefficient and the second penalty coefficient, and are the first constraint function and the second constraint function.

[0093] The final dual objective function model is .

[0094] S52: Solving the Pareto optimal solution for the dual-objective function model to obtain a weight distribution coefficient.

[0095] The dual-objective function model is solved using a Pareto optimal solution algorithm. A Pareto optimal solution is a set of solutions that cannot simultaneously improve all objective function values ​​in a multi-objective optimization scenario. Specifically, the NSGA-II genetic algorithm is used as the solution. First, a population P is initialized with a population size of 100, with each individual representing a set of decision variable values. During the solution, individuals are encoded using real numbers. Decision variables include maintenance time, maintenance intensity, and resource allocation ratios, with a coding length of 3 × N. After initialization, the objective function value for each individual is calculated through fitness evaluation.

[0096] When calculating the objective function value, the non-dominated sorting classifies the population into Pareto levels, where level 1 contains all non-dominated solutions, level 2 contains non-dominated solutions dominated by level 1, and so on; and the crowding distance calculates the distribution density of individuals in the same level, and then the selection operation is performed based on the non-dominated sorting and crowding distance, giving priority to individuals in lower levels, and selecting individuals with larger crowding distances in the same level. For the crossover operation, the present invention uses SBX to simulate binary crossover, the crossover probability is set to 0.9, and the mutation operation uses polynomial mutation. After multiple generations of evolution, the algorithm converges to the Pareto front, which contains multiple non-dominated solutions. Finally, the knee point detection method is used to identify the key solutions on the Pareto front, and the weight distribution coefficient obtained is determined according to the knee point position. The calculation formula is:

[0097]

[0098] in, is the economic weight coefficient, is the safety weight coefficient, is the economic objective function value at the knee point, is the safety objective function value at the knee point.

[0099] S53: Combining the weight distribution coefficient and the water resource scheduling constraint condition, a state quantitative evaluation is performed through a fuzzy comprehensive evaluation algorithm to obtain an operation and maintenance state evaluation result.

[0100] In step S53, the present invention combines the weight distribution coefficient obtained in step S52 with the water resource scheduling constraint conditions to perform state quantitative evaluation. The water resource scheduling constraint conditions include flow balance constraint, capacity limit constraint, time constraint, etc.

[0101] During the evaluation, the present invention processes the comprehensive evaluation of multiple evaluation indicators through a fuzzy comprehensive evaluation algorithm. The evaluation indicator set includes degradation degree index, safety risk index, economic cost index, and operation efficiency index. The comment set includes five levels: excellent, good, average, poor, and very poor.

[0102] The single factor evaluation matrix is ​​calculated by the membership function, which adopts trapezoidal distribution. The membership function of the excellent grade is:

[0103] in, and is the threshold parameter, For evaluation parameters.

[0104] Furthermore, the weight vector is determined by the hierarchical analysis method and combined with the weight distribution coefficient adjustment. After obtaining the above parameters, the fuzzy comprehensive evaluation calculation is performed through the fuzzy synthesis factor. All comment levels are traversed and the comprehensive membership of each level is calculated to obtain the comprehensive evaluation vector. Then, the center of gravity method is used to convert the fuzzy evaluation results into precise values ​​and output the operation and maintenance status assessment score of the facility.

[0105] After obtaining the score, the present invention also determines the risk level based on the assessment score. The criteria are: a score ≥ 90 is low risk, 70 ≤ < 90 is medium risk, and a score < 70 is high risk. After determining the risk level, the present invention also outputs a maintenance priority, taking into account the risk level and the importance of the facility. The importance is determined by factors such as the facility's service population, criticality, and substitutability.

[0106] Finally, all calculation results are integrated into the operation and maintenance status assessment result matrix. The matrix contains the assessment information of all facilities. Each row contains four columns of data: facility ID (string type), assessment score (floating point number, range [55,95]), risk level (string type, "high risk" / "medium risk" / "low risk"), and maintenance priority (integer type, range [1,9]). The matrix is ​​sorted from high to low by maintenance priority. If the priorities are the same, it is sorted from low to high by assessment score. The final operation and maintenance status assessment result is output.

[0107] S6: Based on the operation and maintenance status assessment results, cross-basin cascading fault prevention and control is performed on the water conservancy facilities, a cascading fault prevention and control plan is obtained, and the water conservancy facilities are operated and maintained according to the cascading fault prevention and control plan.

[0108] Wherein, step S6 further includes: S61: Based on the operation and maintenance status assessment result, cross-basin dependency modeling is performed to obtain a cascading fault propagation diagram.

[0109] Wherein, step S61 further includes: S611: Collect the connection relationships of multi-basin facilities, construct the network topology through graph theory modeling algorithms, and obtain the cross-basin facility network diagram.

[0110] Furthermore, in step S611, the present invention collects connection relationship data of multi-basin facilities, including physical connections (pipelines, channels, water flow paths), functional dependencies (water supply links, control logic) and geographical proximity relationships, integrates these relationships into a unified network topology structure through graph theory algorithms, and establishes an adjacency matrix between facilities. When establishing the adjacency matrix, a weighted summation method is used to calculate the comprehensive connection strength, with weights of 0.5 for physical connection, 0.3 for functional dependency, and 0.2 for geographical proximity. Finally, a cross-basin facility network diagram containing characteristic parameters such as node degree and clustering coefficient is generated.

[0111] S612: Analyze the propagation path of the cross-basin facility network diagram through a fault propagation dynamics model to obtain a fault diffusion matrix.

[0112] In step S612, the present invention establishes a fault propagation dynamics model based on infection propagation theory, categorizing facility states into three types: healthy, infected, and failed. When analyzing propagation paths, the state is first initialized based on the operation and maintenance assessment results: scores below 60 are set to infected, scores between 60 and 80 indicate healthy but with a high probability of infection, and scores above 80 indicate strong resistance to infection. The infection rate is then calculated using the formula: base infection rate × connection strength × fault severity factor. Monte Carlo simulation is then performed 1000 times with a time step of 1 hour and a total simulation time of 168 hours (7 days). Each time, an initial faulty facility is randomly selected, and the infection and recovery probabilities are calculated. The propagation path tree is then recorded. The fault diffusion probability matrix and average infection delay are then calculated. The different propagation characteristics of the three fault types, mechanical, electrical, and water quality, are considered to ultimately establish a categorized diffusion matrix.

[0113] S613: Perform cascading effect prediction based on the fault diffusion matrix to obtain a cascading fault propagation graph.

[0114] In step S613, the present invention first establishes a cascading effect model based on the load-capacity relationship, defines the facility load (current workload) and capacity (maximum carrying capacity), and then adopts a load redistribution strategy based on distance and capacity, that is, the load of the faulty facility is redistributed according to the connection strength and available capacity ratio of the neighboring facilities, and then the redistribution process is iterated through the algorithm until the network is balanced.

[0115] After obtaining the weight and cascade effect model, a cascade propagation tree is constructed to record the propagation hierarchy. The root node is the initial fault, and the child nodes are the affected facilities. The failure time and cause are recorded, and finally a cascade fault propagation graph is generated, which includes edge attributes such as propagation probability, delay, and impact intensity.

[0116] S62: Select a fault blocking point from the cascade fault propagation graph to obtain a set of fault prevention and control nodes.

[0117] In step S62, the present invention uses five dimensions to evaluate key nodes, including node degree centrality (direct connection capability), betweenness centrality (bridge function, calculated using the Brandes algorithm), closeness centrality (average distance), eigenvector centrality (neighbor importance, solved using power iteration), and cascade influence coefficient. After evaluation, a comprehensive importance score is calculated, i.e., the weighted sum of each indicator is set to 0.2, 0.3, 0.2, 0.1, and 0.2, respectively. Then, after simulating the removal of nodes, the differences in indicators such as connectivity change, average propagation distance, and propagation coverage are calculated. The network robustness is analyzed to evaluate the blocking effect. After analysis, an importance threshold of 0.7 and a blocking effect threshold of 0.6 are set for two rounds of screening. Finally, the optimal combination is solved using integer linear programming. The objective function of the optimal combination is to maximize the blocking effect, generating multiple prevention and control nodes.

[0118] S63: Based on the fault prevention and control node set, a cross-basin collaborative strategy is generated through a resource allocation optimization algorithm to obtain a cascading fault prevention and control solution.

[0119] Furthermore, in step S63, the present invention analyzes four types of resource requirements during resource allocation: human resources (operation and maintenance personnel, technical experts, and management personnel) to establish a personnel demand matrix; material resources (spare parts, repair tools, and monitoring equipment) to establish a material demand matrix; technical resources determined based on the complexity of prevention and control measures; and information resources, including historical data and expert knowledge. The resource supply capacity of each river basin is calculated to establish a supply matrix. A multi-objective optimization model is then constructed to minimize allocation costs (transportation + storage + opportunity costs), minimize response time (preparation + transportation + deployment time), and maximize resource utilization. Constraints are also established: resource balance constraints to ensure demand is met, supply capacity constraints to prevent overshoot, time window constraints to ensure timeliness, and fairness constraints (Gini coefficient) to balance the burdens of each river basin. An optimization algorithm is then used to generate an optimal allocation plan, enabling coordinated multi-basin prevention and control.

[0120] Taking a large-scale water conservancy system as an example, it contains 150 facilities distributed in 5 river basins. The operation and maintenance status assessment results show that 15 of them have an assessment score below 60 points and are marked as high risk. During the S611 stage processing, the present invention extracts 230 pipeline connection records and 85 channel connection records from the engineering drawings. When calculating the connection strength through the pipeline diameter data, it is found that the maximum pipeline diameter is 2.5 meters. The diameter of a pipeline connecting two pumping stations is 1.8 meters, and its connection strength is calculated to be 0.72; the geographic coordinate data shows two reservoirs 12 kilometers apart. The actual distance is calculated using the spherical distance formula and the proximity is obtained by the Gaussian attenuation function as 0.527. After integrating the three types of connection relationships, the comprehensive connection strength of this pair of facilities in the adjacency matrix is ​​0.369, exceeding the threshold of 0.1 and retaining the connection.

[0121] In the Monte Carlo simulation of stage S612, a pumping station with an evaluation score of 45 was selected as the initial fault facility. Its infection rate was calculated to be 0.074, corresponding to a probability of infecting neighboring facilities within 1 hour of 0.057. After 1,000 simulations, statistics showed that when the pumping station fails, there is a 67% probability of affecting the downstream water treatment plant within 24 hours. The output constructed fault diffusion matrix records the diffusion probability relationship between all pairs of facilities.

[0122] During the load redistribution in stage S613, the current load of the water treatment plant with a processing capacity of 8,000 cubic meters per hour is 6,400 cubic meters per hour, and the load-to-capacity ratio is 0.8. When the two upstream pumping stations fail at the same time and need to bear an additional load of 2,400 cubic meters per hour, the new load reaches 8,800 cubic meters per hour, exceeding the capacity threshold and triggering a cascade failure.

[0123] During the S62 phase of the critical node assessment, the degree centrality of a dispatch center controlling five downstream facilities was 0.034, and the betweenness centrality calculated using the Brandes algorithm was 0.127. The node's overall importance score reached 0.754, exceeding the screening threshold. A blocking effectiveness assessment showed that removing this node reduced the network's maximum connected components from 150 to 132, a 12% decrease in connectivity. The average transmission distance increased from 3.2 hops to 4.7 hops, and the transmission coverage rate decreased from 89% to 74%. The overall blocking effectiveness score was 0.681.

[0124] In the S63 stage resource allocation, the control node set includes 12 key facilities that require 48 operation and maintenance personnel and 120 sets of monitoring equipment. Basin A can provide 15 personnel and 40 sets of equipment, and Basin B can provide 20 personnel and 35 sets of equipment. The transportation distance and response time are taken into account when allocating personnel through the multi-objective optimization algorithm. The final allocation plan allocates 12 personnel from Basin A to the four control nodes that are closer, and 18 personnel from Basin B to the remaining eight control nodes. The total allocation cost is controlled within the budget and the response time does not exceed 6 hours.

[0125] like Figure 2 As shown, the present invention also provides a water conservancy facility operation and maintenance system based on multi-source data, comprising: Acquisition module 100: used to collect multi-source heterogeneous data of water conservancy facilities and obtain a standardized multi-dimensional data matrix; Extraction module 200: used to extract multimodal features of the standardized multidimensional data matrix through a graph neural network fusion algorithm to obtain a fused feature vector; Diagnosis module 300: used to perform fault diagnosis on the fused feature vector through the knowledge graph reasoning algorithm automatically constructed by the domain ontology to obtain fault diagnosis information; Analysis module 400: used to perform facility state evolution analysis based on the fault diagnosis information using a spatiotemporal coupled degradation prediction model to obtain spatiotemporal degradation prediction results; Evaluation module 500: for performing a status evaluation of the water conservancy facilities based on the spatiotemporal degradation prediction results by using a dynamic weight allocation algorithm of dual-objective optimization to obtain an operation and maintenance status evaluation result; Operation and maintenance module 600: used to perform cross-basin cascading fault prevention and control on water conservancy facilities based on the operation and maintenance status assessment results, obtain a cascading fault prevention and control plan, and operate and maintain water conservancy facilities through the cascading fault prevention and control plan.

[0126] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0127] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus the necessary general hardware platform, or of course, by means of hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute a water conservancy facility operation and maintenance method based on multi-source data as described in various embodiments or certain parts of the embodiments.

[0128] This invention significantly improves the accuracy and timeliness of water conservancy facility operation and maintenance through multi-source data fusion and intelligent analysis technology. It can quickly identify hidden dangers in facilities. At the same time, the spatiotemporal coupled degradation model predicts equipment degradation trends in advance, and cooperates with dual-objective optimization evaluation to achieve a dynamic balance between economy and safety. Overall, this invention can effectively block the fault propagation chain, reduce the risk of chain accidents, increase the fault detection rate of water conservancy facilities, reduce operation and maintenance costs, and extend equipment life. It is particularly suitable for the full life cycle management of key facilities such as reservoirs, dams, and pumping stations, greatly enhancing the risk resistance and comprehensive benefits of water conservancy infrastructure, and providing intelligent protection for water resource scheduling safety.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A water conservancy facility operation and maintenance method based on multi-source data, characterized in that: include: S1: Collect multi-source heterogeneous data of water conservancy facilities and obtain a standardized multidimensional data matrix; S2: extracting multimodal features of the standardized multidimensional data matrix through a graph neural network fusion algorithm to obtain a fused feature vector; S3: Perform fault diagnosis on the fused feature vector through a knowledge graph reasoning algorithm automatically constructed by domain ontology to obtain fault diagnosis information; S4: Based on the fault diagnosis information, perform facility state evolution analysis using a spatiotemporal coupled degradation prediction model to obtain a spatiotemporal degradation prediction result; Wherein, step S4 further includes: S41: performing degradation field modeling on the fault diagnosis information through a spatiotemporal partial differential equation solver to obtain a degradation propagation matrix; S42: performing spatial discretization on the degradation propagation matrix through a finite element analysis algorithm to obtain a gridded degradation distribution; S43: performing facility state evolution calculation based on the gridded degradation distribution through a spatiotemporal coupling prediction algorithm to obtain a spatiotemporal degradation prediction result; S5: Based on the spatiotemporal degradation prediction results, a state assessment of the water conservancy facilities is performed using a dynamic weight allocation algorithm of dual-objective optimization to obtain an operation and maintenance state assessment result; S6: Based on the operation and maintenance status assessment results, cross-basin cascading fault prevention and control is performed on the water conservancy facilities, a cascading fault prevention and control plan is obtained, and the water conservancy facilities are operated and maintained according to the cascading fault prevention and control plan.

2. The water conservancy facility operation and maintenance method based on multi-source data according to claim 1, characterized in that: Step S2 further comprises: S21: extracting the spatiotemporal features of the standardized multidimensional data matrix through an LSTM-CNN hybrid network to obtain spatiotemporal feature data; S22: Performing cross-modal feature mapping on the spatiotemporal feature data through a graph attention network to obtain high-dimensional feature space data; S23: Perform feature discrimination enhancement on the high-dimensional feature space data through a contrastive learning mechanism to obtain a fused feature vector.

3. The water conservancy facility operation and maintenance method based on multi-source data according to claim 2, characterized in that: In step S21, the LSTM-CNN hybrid network is optimized and trained through a federated learning framework, specifically including: S211: Establish a federated learning client, update the model parameters of the federated learning client through local gradient calculation, and obtain local model parameters; S212: Performing noise perturbation on the local model parameters through a differential privacy mechanism to obtain privacy protection parameters; S213: Based on the privacy protection parameters, global model aggregation is performed through a secure aggregation protocol to obtain an LSTM-CNN hybrid network for federated learning collaborative optimization.

4. The water conservancy facility operation and maintenance method based on multi-source data according to claim 1, characterized in that: Step S3 further comprises: S31: Using the BiLSTM-CRF model, we extract entity relationships from text data in the water conservancy industry and obtain a set of entity relationship triples. S32: vectorizing and embedding the entity relationship triple set to obtain a water conservancy facilities domain ontology framework; S33: Inputting the fused feature vector into the water conservancy facilities domain ontology framework to perform fault pattern recognition through entity linking reasoning to obtain fault diagnosis information.

5. The water conservancy facility operation and maintenance method based on multi-source data according to claim 4 is characterized in that: Step S31 further includes: S311: Select an operation and maintenance report text from a public corpus, perform word segmentation on the operation and maintenance report text using a word segmentation algorithm, and obtain a word sequence; S312: Perform context encoding on the vocabulary sequence through a BiLSTM encoder to obtain a vocabulary vector representation; S313: Identify the entity boundaries represented by the vocabulary vector through the CRF layer to obtain a set of entity relationship triples.

6. The water conservancy facility operation and maintenance method based on multi-source data according to claim 1, characterized in that: Step S41 further includes: S411: Establish a facility degradation field, describe the degradation propagation mechanism through diffusion equation modeling, and obtain the diffusion control equation; S412: setting boundary conditions of the diffusion control equation to determine the solution domain and obtain a degenerate field solution model; S413: Perform degradation propagation calculation using the degradation field solution model to obtain a degradation propagation matrix.

7. The water conservancy facility operation and maintenance method based on multi-source data according to claim 1, characterized in that: Step S5 further comprises: S51: Based on the spatiotemporal degradation prediction result, cost risk modeling is performed using a multi-objective optimization function to obtain a dual-objective function model; wherein the multi-objective optimization function includes an economic index function and a safety index function; S52: Solving the Pareto optimal solution of the dual-objective function model to obtain a weight distribution coefficient; S53: In combination with the weight distribution coefficient and the water resource scheduling constraint condition, a state quantitative evaluation is performed through a fuzzy comprehensive evaluation algorithm to obtain an operation and maintenance state evaluation result.

8. The water conservancy facility operation and maintenance method based on multi-source data according to claim 1, characterized in that: Step S6 further comprises: S61: Modeling cross-basin dependency relationships based on the operation and maintenance status assessment results to obtain a cascading fault propagation diagram; S62: Select a fault blocking point from the cascade fault propagation graph to obtain a set of fault prevention and control nodes; S63: Based on the fault prevention and control node set, a cross-basin collaborative strategy is generated through a resource allocation optimization algorithm to obtain a cascading fault prevention and control solution.

9. The water conservancy facility operation and maintenance method based on multi-source data according to claim 8, characterized in that: Step S61 further includes: S611: Collect the connection relationships of multi-basin facilities, construct the network topology through graph theory modeling algorithms, and obtain the cross-basin facility network diagram; S612: Analyze the propagation path of the cross-basin facility network diagram using a fault propagation dynamics model to obtain a fault diffusion matrix; S613: Perform cascading effect prediction based on the fault diffusion matrix to obtain a cascading fault propagation graph.

Images