Gravity dam safety risk management and control method based on knowledge graph and large model

Abstract

The invention discloses a gravity dam safety risk management and control method based on a knowledge graph and a large model, and the method comprises the steps: constructing a gravity dam safety risk space-time knowledge graph which comprises an entity layer, a relation layer, an attribute layer and a time-space mapping layer, and achieving the full-life-cycle multi-source data structured association; constructing a knowledge enhanced large model, and embedding the space-time knowledge graph into a reasoning process; real-time and historical data of the gravity dam are collected and preprocessed, and the large model is input; calling a map association relationship and an evolution rule through a large model to complete risk anomaly identification, cause positioning and evaluation; historical cases and industry specifications in the atlas are retrieved, and a management and control decision scheme matched with the current working condition is generated; and collecting scheme execution data, extracting new knowledge and updating the graph. According to the invention, a closed-loop management and control mode is constructed, and the intelligent and refined level of risk management and control is improved.

Description

Technical Field

[0001] This invention relates to the field of water conservancy project safety management and control technology, and in particular to a method for managing the safety risks of gravity dams based on knowledge graphs and large models. Background Technology

[0002] Gravity dams, as massive and crucial water-retaining structures in water conservancy projects, are widely used in major infrastructure construction for flood control, power generation, and irrigation. Their safe and stable operation is directly related to the safety of people's lives and property, the sustainability of the ecological environment, and the stable development of the regional economy and society. With the extension of their service life, gravity dams face the impact of natural factors such as aging of the dam concrete, evolution of foundation geological conditions, and water erosion. Simultaneously, the frequent occurrence of extreme weather events and fluctuations in operating loads, combined with human and environmental factors, lead to the continuous accumulation of safety risks and hidden dangers such as abnormal seepage flow, excessive dam displacement, and crack propagation. Against this backdrop, the multi-source heterogeneous data generated throughout the entire life cycle of gravity dams is experiencing explosive growth, encompassing structural parameters from the design phase, process records from the construction phase, real-time monitoring data from the operation phase, historical incident cases, maintenance records, and industry standard texts. Traditional control models relying on manual experience and single monitoring indicator analysis are no longer sufficient to meet the high demands of modern safety management for data integration, risk prediction, and scientific decision-making. There is an urgent need to introduce intelligent and systematic technical means to upgrade risk management.

[0003] There are currently several key issues in the field of gravity dam safety risk management that urgently need to be addressed: First, multi-source data lacks effective structured correlation and spatiotemporal mapping. Information such as design parameters, real-time monitoring data, and historical cases are stored in a scattered and heterogeneous format, making it impossible to establish logical links such as causal relationships and spatiotemporal correlations between entities. This results in a lack of comprehensive and coherent data support for risk analysis, making it difficult to accurately capture the evolutionary patterns of risks. Second, the accuracy and adaptability of risk causation identification and management decisions are insufficient. Traditional methods often rely on single threshold judgments to identify risks, making it difficult to uncover the key links and root causes of risk propagation. Moreover, decision-making solutions often simply copy general experience or industry standards, failing to be personalized and optimized based on the specific working conditions such as the current dam structure characteristics and operating status. At the same time, the knowledge system cannot dynamically iterate with management practices, resulting in limited decision-making effectiveness. Summary of the Invention

[0004] The technical problem to be solved by this invention is the lack of structured association of multi-source data, inaccurate identification of risk causes, poor adaptability of control decisions, and inability of knowledge system to dynamically iterate in the safety risk management of gravity dams. To this end, a gravity dam safety risk management method based on knowledge graphs and large models is provided to solve the above problems.

[0005] To address the aforementioned technical problems, the technical solution of this invention is: a gravity dam safety risk management method based on knowledge graphs and large models, comprising the following steps:

[0006] S1: Construct a spatiotemporal knowledge graph of gravity dam safety risks. The spatiotemporal knowledge graph includes an entity layer, a relation layer, and an attribute layer to realize the structured association and spatiotemporal dimension mapping of multi-source data throughout the entire life cycle of gravity dams.

[0007] S2: Based on the retrieval-enhanced generation architecture and domain-adaptive training, a knowledge-enhanced large model is constructed, and the spatiotemporal knowledge graph is embedded as a structured knowledge source into the reasoning process of the large model;

[0008] S3: Collect real-time operation data and full life-cycle historical data of gravity dams, and after data cleaning, standardization and feature extraction processing, input them into the knowledge-enhanced large model;

[0009] S4: By calling the entity relationship and spatiotemporal evolution rules of the spatiotemporal knowledge graph through the knowledge-enhanced large model, risk anomaly identification, correlation path mining and root cause location are completed, and risk assessment results containing traceability evidence are generated.

[0010] S5: Based on the risk assessment results, the knowledge-enhanced big model retrieves historical cases, industry standards and control rules from the spatiotemporal knowledge graph to generate interpretable risk control decision-making schemes that are adapted to the current working conditions;

[0011] S6: Collect the execution data and effect evaluation results of the risk management decision-making scheme, extract new entities, relationships and rules through the knowledge-enhanced big data model, and dynamically update the spatiotemporal knowledge graph.

[0012] Preferably, in step S1, the entity layer includes dam structure entity, monitoring facility entity, risk factor entity, historical case entity, control measure entity, and industry standard entity;

[0013] The relationship layer includes causal relationships, spatiotemporal relationships, subordinate relationships, and constraint relationships between entities; the attribute layer includes static attributes and dynamic attributes of entities. Static attributes include design parameters, structural features, and specification thresholds, while dynamic attributes include real-time monitoring data, operating status parameters, and timestamp information.

[0014] Preferably, in step S1, the process of constructing the spatiotemporal knowledge graph includes:

[0015] S11: Collect multi-source heterogeneous data on gravity dams, including design documents, construction records, monitoring data, historical incident cases, maintenance records, and industry standard texts;

[0016] S12: Using natural language processing technology and graph neural network algorithms, entity, relation and attribute information is extracted from the data, and after verification and correction, it is formed into structured knowledge units;

[0017] S13: Based on the OGC standard, a spatiotemporal correlation model is constructed, which binds structured knowledge units with spatial location information and time series information to form spatiotemporal knowledge triples;

[0018] S14: Use a graph database to store spatiotemporal knowledge triples to construct a spatiotemporal knowledge graph of gravity dam safety risks with spatiotemporal query and path traversal capabilities.

[0019] Preferably, in step S2, the construction process of the knowledge-enhanced large model includes:

[0020] S21: Use a knowledge embedding algorithm to convert entities, relations and attributes in the spatiotemporal knowledge graph into low-dimensional vectors and integrate them into the word embedding layer of the large model;

[0021] S22: Construct a retrieval-enhanced generative architecture to establish a real-time connection between the spatiotemporal knowledge graph and the large model, enabling the large model to dynamically retrieve relevant structured knowledge during reasoning;

[0022] S23: Collect labeled data in the field of gravity dam safety management and conduct domain-adaptive training on the large model.

[0023] Preferably, step S4, the risk tracing process specifically includes the following steps:

[0024] S41: The knowledge-enhanced large model compares the preprocessed data with the normal operation threshold in the spatiotemporal knowledge graph to identify risk and abnormal signals and corresponding related entities;

[0025] S42: Traverse the direct and indirect relationship paths of related entities in the spatiotemporal knowledge graph to uncover the key links of risk propagation;

[0026] S43: Analyze the influence weight of each entity in the critical link based on the reasoning ability of the large model, and locate the root cause of the risk by combining the spatiotemporal evolution rules;

[0027] S44: Connect professional evidence in the spatiotemporal knowledge graph to generate interpretable risk assessment results that include risk causes, evolution paths, and judgment criteria.

[0028] Preferably, in step S5, the process of generating the control decision scheme includes:

[0029] S51: The knowledge-enhanced big model retrieves matching historical success cases and relevant industry standards from the spatiotemporal knowledge graph based on the root causes of risks and the characteristics of current working conditions.

[0030] S52: Optimize the retrieved historical control measures to form a set of candidate measures that conform to the current structural characteristics and operating status of the gravity dam;

[0031] S53: Based on the scope of risk impact, control objectives, and implementation conditions, prioritize candidate measures and generate a complete decision-making plan that includes control objectives, specific measures, implementation steps, expected effects, and professional basis.

[0032] Preferably, in step S6, the dynamic update process of the spatiotemporal knowledge graph specifically includes the following steps:

[0033] S61: Collect data on the execution process of the decision-making plan, on-site feedback data, and effect verification data;

[0034] S62: Analyze the deviation between the execution effect and the expected effect through knowledge-enhanced large model, and extract new risk factors, control measures and correlations;

[0035] S63: After verification, the newly extracted knowledge units are integrated into the spatiotemporal knowledge graph, updating entity attributes, relationship paths, and evolution rules to achieve the autonomous evolution of the knowledge graph.

[0036] Preferably, the data preprocessing process in step S3 specifically includes the following steps:

[0037] An outlier removal algorithm is used to process outliers in the monitoring data. Data units are standardized or normalized using a standardization or normalization algorithm. Text features in unstructured documents are extracted using optical character recognition technology. Morphological feature parameters in monitoring images are extracted using an image feature extraction algorithm.

[0038] Preferably, the knowledge embedding algorithm in step S21 specifically adopts one or more combinations of TransE, TransH, TransR or ComplEx algorithms;

[0039] The domain adaptation training includes pre-training fine-tuning based on domain corpus, cue engineering optimization, and supervised learning training with a small number of labeled samples.

[0040] Preferably, the retrieval process in step S51 adopts a dual retrieval mechanism of semantic similarity matching and working condition feature vector comparison. The priority ranking criteria in step S53 include the implementation cost of the measures, the speed of risk mitigation, the construction complexity, and the degree of impact on the normal operation of the gravity dam.

[0041] The beneficial effects of this invention are:

[0042] 1. This invention constructs a spatiotemporal knowledge graph of gravity dam safety risks, comprising entity, relation, attribute, and spatiotemporal mapping layers. It structurally integrates multi-source heterogeneous data from the entire lifecycle, including design documents, construction records, monitoring data, and historical cases. By leveraging natural language processing technology and graph neural network algorithms to extract entity, relation, and attribute information, and combining OGC standards to establish a spatiotemporal correlation model, this invention solves the problems of data dispersion, heterogeneous formats, and lack of logical connections in traditional management and control. It transforms previously isolated data into a structured knowledge system with causal relationships and spatiotemporal correlations, providing comprehensive and coherent data support for risk analysis and accurately capturing the spatiotemporal patterns of risk evolution.

[0043] 2. This invention relies on the collaborative reasoning capabilities of a knowledge-enhanced large model and a spatiotemporal knowledge graph. It identifies anomalies by comparing multi-source preprocessed data with standardized thresholds in the graph, utilizes depth-first search to uncover key links in risk propagation, and combines the large model's self-attention mechanism to analyze entity influence weights, accurately locating the root cause. Simultaneously, it uses industry standards and historical cases as judgment criteria to generate risk assessment results that include causes, evolution paths, and professional evidence, solving the problems of vague risk tracing and lack of scientific support in traditional methods.

[0044] 3. The knowledge-enhanced large model proposed in this invention matches historical successful cases and industry standards through a dual retrieval mechanism. It calculates the adaptability by combining the current dam structure characteristics and operating status, optimizes historical control measures to form a candidate set, and then ranks them based on indicators such as risk impact scope and implementation feasibility to generate personalized decision-making schemes. At the same time, by collecting decision execution data and effect evaluation results, it extracts new risk factors, control measures and correlations, integrates them into the knowledge graph after verification, and realizes the autonomous evolution of the knowledge system. It solves the problems of traditional decision-making copying experience, poor adaptability and lack of knowledge iteration, and continuously improves the scientificity and timeliness of control decisions.

[0045] In summary, this invention, through an innovative architecture combining spatiotemporal knowledge graphs with a knowledge-enhanced large-scale model, systematically addresses the core pain points in gravity dam safety risk management, such as difficulty in integrating multi-source data, inaccurate risk positioning, poor decision adaptability, and rigid knowledge systems. It achieves full-process optimization of structured data association, precise risk tracing, personalized decision adaptation, and dynamic knowledge iteration, constructing a closed-loop management model of knowledge construction, reasoning application, and knowledge updating, significantly improving the intelligence and precision of gravity dam safety risk management. Attached Figure Description

[0046] Figure 1 This is a closed-loop block diagram of the overall technical process proposed in this invention;

[0047] Figure 2 This is a block diagram of the knowledge-enhanced large model architecture proposed in this invention. Detailed Implementation

[0048] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0049] Please refer to the reference. Figures 1-2 A gravity dam safety risk management method based on knowledge graphs and large models includes the following steps:

[0050] S1: Construct a spatiotemporal knowledge graph of gravity dam safety risks. The spatiotemporal knowledge graph includes an entity layer, a relation layer, and an attribute layer, realizing the structured association and spatiotemporal dimension mapping of multi-source data throughout the entire life cycle of gravity dams.

[0051] The entity set E is denoted as the entity layer, which includes the dam structure entity, monitoring facility entity, risk factor entity, historical case entity, control measure entity, and industry standard entity. The relationship set R is denoted as the relationship layer, which includes the causal relationship, spatiotemporal correlation relationship, subordinate relationship, and constraint relationship. The attribute set A is denoted as the static attribute (design parameters, structural features, standard thresholds) and dynamic attribute (real-time monitoring data, operating status parameters, timestamp information) of the attribute layer. This realizes the structured association and spatiotemporal dimension mapping of multi-source data throughout the entire life cycle of gravity dams.

[0052] The construction process of the spatiotemporal knowledge graph includes:

[0053] S11: Collect multi-source heterogeneous data on gravity dams, including design documents, construction records, monitoring data, historical incident cases, maintenance records, and industry standard texts. Design documents include dam structural parameters (dam height, dam width, concrete strength grade) and foundation geological parameters (rock compressive strength, permeability coefficient). Construction records include construction process parameters (pouring thickness, curing period, quality inspection parameters (pouring density, initial crack width). Monitoring data includes real-time monitoring parameters (seepage flow, dam horizontal displacement, vertical displacement, stress, crack propagation width). Historical incident case parameters include incident type, occurrence time, impact range, and treatment effect. Maintenance records include maintenance measures, implementation time, and maintenance period. Industry standard texts include standard number, threshold standards, and constraints.

[0054] S12: Using natural language processing technology and graph neural network algorithms, entity, relation, and attribute information is extracted from the data. After verification and correction, structured knowledge units are formed, specifically including:

[0055] Entity extraction employs BERT combined with Conditional Random Fields to extract text-like data from multi-source heterogeneous data and convert it into text sequences. , For the i-th text), and based on the entity categories predefined in the field of gravity dam safety management, a reference sequence used to label entities in the text sequence is used as the entity label set ( (K is the number of entity categories); define the text sequence and the entity label set as input parameters; construct the objective function, expressed as: ;in, For the set of model parameters, The number of data samples. This represents the maximum time step of the text sequence. For tags arrive The transition probability, Output for BERT model Corresponding tags The probability of launch, Normalization factor;

[0056] Relation extraction employs PCNN combined with the Attention algorithm, and sentence-level relation features are represented as follows: ;in, For the characteristic projection matrix ( For relational feature dimensions, (output feature dimension for PCNN) This represents the local feature extraction results of PiecewiseCNN on the input text x. Here is the Attention weight vector. This is the final output relation feature vector;

[0057] Attribute extraction uses the BiLSTM algorithm, and the hidden layer feature output is: , , ;in, For the first The input text vector at each time step and These are the first and second LSTMs, respectively. Step hidden layer output, and These represent the historical hidden layer states of the forward and backward LSTMs, respectively. This is a vector concatenation operation. For the first The final attribute feature vector of the step;

[0058] S13: Construct a spatiotemporal correlation model based on the OGC standard, binding structured knowledge units with spatial location information and time series information to form spatiotemporal knowledge triples; specifically including:

[0059] The spatiotemporal correlation parameter definition includes spatial location information parameters such as the spatial coordinates of the dam section (longitude lon, latitude lat, elevation alt) and the deployment location of monitoring facilities (coordinates). The time series information parameter includes the data acquisition start time. End time Update time ;

[0060] The spatiotemporal correlation model is constructed based on the OGC standard (EPSG:4326 spatial reference system, ISO8601 time format), and a spatiotemporal mapping function is built to associate structured knowledge units with spatiotemporal parameters.

[0061] Binding and triple formation: Structured knowledge units Binding with spatial information S and temporal information T, a spatiotemporal knowledge triple is formed, expressed as follows: ,in For the head entity, It is a related relationship. For tail entities, For attribute information, , ;

[0062] S14: A spatiotemporal knowledge triplet is stored using a graph database to construct a spatiotemporal knowledge graph of gravity dam safety risks, capable of spatiotemporal querying and path traversal; specifically including:

[0063] The graph database storage configuration uses Neo4j graph database, which maps entities to graph nodes, relationships to edges between nodes, and attributes and spatiotemporal information to attributes of node edges.

[0064] Spatiotemporal index construction: Based on R-tree spatial index and B-tree temporal index, a spatiotemporal composite index is constructed to support efficient retrieval in the spatiotemporal dimension;

[0065] Spatiotemporal query and path traversal are implemented. Spatiotemporal query supports conditional retrieval based on spatial range (e.g., "monitoring points within 50m of the dam section") and time interval (e.g., "seepage data in month xx of year xxxx"). Path traversal uses the depth-first search (DFS) algorithm to mine risk links, and the expression is: ( For the starting entity node, For the set of relations to be traversed, The target entity node is defined as follows: The gravity dam safety risk-related triples stored in the Neo4j graph database, where entities are mapped to graph nodes, relationships are mapped to edges between nodes, and attributes and spatiotemporal information are mapped to node or edge attributes, are denoted as the set of spatiotemporal knowledge positive triples. ;

[0066] S2: Based on the retrieval-enhanced generation architecture and domain-adaptive training, a knowledge-enhanced large model is constructed, embedding the spatiotemporal knowledge graph from step S1 as a structured knowledge source into the large model's reasoning process; the construction process of the knowledge-enhanced large model includes:

[0067] S21: A knowledge embedding algorithm is used to convert entities, relations, and attributes in the spatiotemporal knowledge graph into low-dimensional dense vectors that can be recognized by large models, achieving a unified representation of structured knowledge and textual semantics. The input data is directly taken from the entity set E, relation set R, and attribute set A in the spatiotemporal knowledge graph constructed in step S1. This step uses one or more combinations of TransE, TransH, TransR, and ComplEx algorithms to complete the knowledge embedding, and the implementation logic is as follows:

[0068] When using the TransE algorithm, based on the assumption that the sum of the head entity vector and the relation vector is approximately equal to the tail entity vector, vector optimization is achieved by minimizing the distance difference between positive and negative sample triples. Its loss function expression is as follows: ,in, Let E represent the set of negative sample triples generated by randomly replacing the head or tail entities of positive samples, where h and t are the low-dimensional vectors corresponding to the head and tail entities in E, respectively, and r is the low-dimensional vector corresponding to the relation in R. It is the L2 norm. For the boundary margin parameter;

[0069] When using the TransH algorithm, a dedicated hyperplane is defined for each relationship. Entity vectors are projected onto the hyperplane before association modeling, thus resolving the entity vector confusion problem in multi-relationship scenarios. The formula for calculating the projected vector is as follows: , The loss function is ,in Let r be the normal vector of the hyperplane corresponding to relation r (satisfying) ), , The projection vector of the entity onto the hyperplane;

[0070] When using the TransR algorithm, entities and relations are mapped to entity space and relation space respectively. Spatial transformation is achieved through projection matrices, improving the ability to represent complex relations. The mapping formula is as follows: , The loss function is consistent with TransE, where Let r be the projection matrix corresponding to relation r (with dimensions d×k, where k is the dimension of the relation space, usually d≤k);

[0071] The ComplEx algorithm embeds vectors into complex space and utilizes the conjugate symmetry of complex numbers to model bidirectional associations between entities and attributes, and between entities and relations. Its objective function is... Where h, t, and r are complex space vectors, For the complex conjugate of t, For the inner product of a complex space, To perform the real part operation, The Sigmoid activation function is used. After processing by the above algorithm, entities, relations, and attributes are all transformed into low-dimensional vectors of a unified dimension, which are directly integrated into the word embedding layer of the large model to achieve a unified representation of structured knowledge and text semantics.

[0072] S22: Construct a retrieval-enhanced generative architecture to establish a real-time connection between the spatiotemporal knowledge graph and the large model, enabling the large model to dynamically retrieve relevant structured knowledge during reasoning; this retrieval-enhanced generative architecture specifically includes:

[0073] Using the R-tree spatial index and B-tree temporal index constructed in step S14, vector indexes and semantic indexes are established for spatiotemporal knowledge triples. The vector index is constructed using a KD-tree, which supports fast vector similarity matching. The semantic index is built based on entity name, relation type, and attribute keywords to construct an inverted index, which supports text semantic matching.

[0074] During large-scale model inference, a query vector q or query text Q is generated based on the current task (such as risk anomaly identification or cause localization). Relevant knowledge is acquired through a dual retrieval mechanism. Vector retrieval calculates the cosine similarity between the query vector and the knowledge graph vector index (the formula is...). ), where q is the query vector generated by the large model, and v is the entity / relationship / attribute vector. Similarity value (range [-1, 1]) Semantic retrieval extracts keywords from the query text and matches them with the semantic index; then the retrieved spatiotemporal triples, entity attributes, and relational paths are converted into a text format that the large model can understand, concatenated with the original input data, and input into the large model to achieve collaborative reasoning between retrieved knowledge and input data;

[0075] S23: Collect labeled data in the field of gravity dam safety management, conduct domain-adaptation training on the large model, and improve the model's ability to understand hydraulic engineering terminology and risk evolution patterns; specifically including:

[0076] The training data is determined by acquiring the labeled dataset D in the field of gravity dam safety management and the unlabeled corpus in the field, which are recorded as the training data for field adaptation training. Among them, the labeled dataset D contains three core samples: risk identification samples, cause location samples, and decision scheme samples. The unlabeled corpus in the field consists of gravity dam design documents, construction records, and industry standard texts.

[0077] Training fine-tuning involves further pre-training of the large model based on unlabeled domain corpora. The masked language model (MLM) loss function is used to optimize the large model parameters. The expression for the MLM loss function is as follows: ;in, The number of words that are masked. These are professional terms used in the field of gravity dam safety management that have been masked. The unmasked context text. For the parameter set of a large model, Predict the probability of masked words for large models;

[0078] The system provides suggestions for engineering optimization, including pre-set specific suggestion templates for gravity dam safety management. These templates guide the large model to output results that conform to hydraulic engineering standards. An example of the suggestion template is: "Given the current working conditions of the gravity dam: horizontal displacement of the dam body = 3.2mm (standard threshold ≤ 2.5mm), seepage flow = 12L / s (standard threshold ≤ 8L / s); combining historical cases in the spatiotemporal knowledge graph and industry standards, the system identifies the types of risk anomalies, locates the root causes, and generates management measures with a set priority order. Each measure must explain the implementation steps and professional basis."

[0079] Supervised learning with a small number of labeled samples uses a labeled dataset D to supervise the training of a large model, employing the cross-entropy loss function as the objective function. The expression for the cross-entropy loss function is as follows: ;in, C represents the number of labeled samples, and C represents the number of task categories. Let be the true label of the i-th sample, represented by independent encoding. Predict the probability that the i-th sample belongs to the c-th class for a large model;

[0080] During the domain adaptation training process, the following strategies are employed to optimize training effectiveness and stability:

[0081] Using cosine annealing learning rate, the initial learning rate is set to... The decay period is 1 / 3 of the number of training rounds; the batch size is set, with a value range of 8-32; an early stopping mechanism is also set, which stops training when the task accuracy on the validation set does not improve for a set number of consecutive rounds; the task accuracy refers to the proportion of samples whose prediction results are consistent with the true labels of the samples after inputting the validation set samples divided from the domain-labeled dataset D into the trained large model and obtaining the prediction results of the model for tasks such as risk identification, cause localization, and decision scheme generation.

[0082] S3: Collect real-time operation data and full life-cycle historical data of gravity dams, and after data cleaning, standardization and feature extraction processing, input them into the knowledge-enhanced large model;

[0083] S4: By calling the entity relationship and spatiotemporal evolution rules of the spatiotemporal knowledge graph through the knowledge-enhanced large model, risk anomaly identification, correlation path mining and root cause location are completed, and risk assessment results containing traceability evidence are generated.

[0084] The risk tracing process specifically includes the following steps:

[0085] S41: The knowledge-enhanced large model compares the preprocessed data with the normal operation thresholds in the spatiotemporal knowledge graph to identify risk anomaly signals and corresponding associated entities; specifically including:

[0086] Using the preprocessed data output from step S3 as input, and relying on the relationship between the knowledge-enhanced large model and the spatiotemporal knowledge graph, the preliminary identification of risk anomalies is completed:

[0087] For each monitoring data point in the preprocessed data (such as dam horizontal displacement, seepage flow, etc.), the normal operation threshold range of the corresponding entity is retrieved from the attribute set A of the spatiotemporal knowledge graph (from step S1). (Belonging to the static attribute "standard threshold" in A), calculate the relative deviation of this monitoring data from the mean threshold. ,in is the threshold mean, and x is a single monitoring data point after preprocessing;

[0088] relative deviation Compared with the preset anomaly detection threshold If compared, If the signal corresponding to the monitoring data is determined to be a risk anomaly signal, the entity to which the monitoring data belongs (such as the "dam structure entity" corresponding to the horizontal displacement of the dam body) is determined to be an associated entity. The associated entity is taken from the entity set E constructed in step S1.

[0089] S42: Traverse the direct and indirect relationship paths of related entities in the spatiotemporal knowledge graph to uncover key links in risk propagation; specifically including:

[0090] Starting with the related entity And obtain the relation set R, which includes causal relationships, spatiotemporal relationships, etc.; use the depth-first search (DFS) algorithm to traverse the direct and indirect relationships of the starting node, and the traversal expression is: ;in, Let R be the set of relations. This involves identifying risk-related entity nodes (such as "crack propagation entities") within the spatiotemporal knowledge graph. During traversal, the entity nodes and relationship types contained in each path are recorded, forming several risk propagation paths. Subsequently, the link strength of each path is calculated using the following formula: ,in, The influence weight of relation r in the path; link strength With preset key threshold If compared, If so, then this path is identified as the key link in the risk transmission;

[0091] S43: Based on the reasoning capabilities of large models, analyze the influence weights of each entity in the critical links, and combine this with spatiotemporal evolution rules to pinpoint the root causes of risks; specifically including:

[0092] By leveraging the self-attention mechanism of knowledge-enhanced large models, the attention weights of each entity node in the critical link are calculated. ( (This indicates the semantic importance of the entity in the link).

[0093] Retrieve the dynamic attribute data of each entity in the critical link, and calculate the time-series rate of change of the entity attributes. The formula is as follows: ,in, The attribute data of entity i at the current moment. The attribute data of entity i at the previous moment;

[0094] The overall influence weight of an entity is calculated based on attention weight and temporal change rate, using the following formula: ;

[0095] Weight the combined impact of each entity Sort by size from largest to smallest, the entity with the highest weight corresponds to the attribute anomaly, which is the direct cause of the risk.

[0096] By combining the spatiotemporal evolution rules in the spatiotemporal knowledge graph (derived from the causal relationships in the relation set R, such as the temporal causal rule of "abnormal seepage flow → increased dam permeability coefficient → increased porosity of geological layer"), the upstream entity anomaly corresponding to the direct cause can be traced, and the upstream entity anomaly is the root cause of the risk.

[0097] S44: Based on professional evidence from the spatiotemporal knowledge graph, generate interpretable risk assessment results containing risk causes, evolution paths, and judgment criteria; specifically including:

[0098] The system retrieves the attribute "clause content" (such as relevant clauses of the "Code for Design of Concrete Gravity Dams" GB50199-2012) corresponding to the "industry standard entity" in the spatiotemporal knowledge graph entity set E, and associates and matches the root cause located in S43 and the key risk propagation link discovered in S42 with the corresponding industry standard clauses. Through the process from S41 to S44, the system completes the full-link tracing of risk from anomaly identification to root cause location, and generates interpretable assessment results with professional basis, providing accurate support for the generation of subsequent control and management decision-making schemes.

[0099] S5: Based on the risk assessment results, the knowledge-enhanced big model retrieves historical cases, industry standards and control rules from the spatiotemporal knowledge graph to generate interpretable risk control decision-making schemes that are adapted to the current working conditions;

[0100] S6: Collect execution data and effect evaluation results of risk management decision-making schemes, extract new entities, relationships and rules through knowledge-enhanced large models, dynamically update the spatiotemporal knowledge graph, and form a closed-loop management of "knowledge construction - reasoning application - knowledge iteration".

[0101] In this application, step S5, the process of generating the control decision scheme, includes:

[0102] S51: The knowledge-enhanced large model retrieves matching historical success stories and relevant industry standards from the spatiotemporal knowledge graph based on the root causes of risks and the characteristics of current operating conditions; specifically including:

[0103] The input consists of the root causes of the risk identified in step S43 (e.g., "concrete compaction is less than 95%)" and the current working condition characteristics. These current working condition characteristics include dam structural parameters (dam height, material strength, attributes of the "dam structural entity" in the entity set E of step S1), real-time operating parameters (current water level, seepage rate, data from the preprocessed data in step S3), and environmental parameters (temperature, rainfall, results from preprocessed monitoring data). The knowledge-enhanced large model converts this input into a retrieval vector. The retrieval enhancement architecture constructed in step S22 is invoked to perform vector retrieval on the "historical success case entities" and "industry standard entities" within the entity set E in the spatiotemporal knowledge graph. The retrieval similarity is calculated using cosine similarity, with the formula: ;in, For low-dimensional vectors corresponding to "historical success case entities" or "industry standard entities"; similarity Historical control measures and industry standard clauses corresponding to entities that exceed the preset similarity threshold (which is determined according to the case matching standard of the "Construction Specification for Risk Management and Control Case Database of Water Conservancy Projects") are identified as matching search results.

[0104] S52: Optimize the retrieved historical control measures to form a set of candidate measures that conform to the current structural characteristics and operational status of the gravity dam; specifically including:

[0105] Calculate the fit of each historical measure. The formula is ;

[0106] in, The structural adaptation coefficient (valued as the cosine similarity between the current dam structure parameters and the dam structure parameters of historical cases; all parameters are taken from the attributes of the entity set E in step S1). The running status adaptation coefficient (the value is the cosine similarity between the current running parameters and the running parameters of historical cases, and the parameter is taken from the preprocessed data in step S3). The effect fit coefficient (valued as the governance efficiency of historical measures for similar risks, taken from the "governance efficiency" attribute of "historical successful case entity"). , , These are the weights of each coefficient;

[0107] If compatibility If the value exceeds the preset adaptation threshold, the historical measure will be optimized in detail (e.g., the "grouting pressure 0.5MPa" in the historical case will be adjusted to "grouting pressure 0.4MPa" to adapt to the current dam strength); all optimized measures will be summarized to form a set of candidate measures.

[0108] S53: Based on the scope of risk impact, control objectives, and implementation conditions, prioritize candidate measures and generate a complete decision-making plan that includes control objectives, specific measures, implementation steps, expected results, and professional basis; specifically including:

[0109] Three ranking evaluation indicators and their corresponding calculation methods were determined:

[0110] The risk impact scope indicator is the number of entity nodes involved in the risk. This indicates that the normalized value is calculated. ( The maximum number of entity nodes involved in historical risk cases in the spatiotemporal knowledge graph (taken from the attribute of "historical successful case entity").

[0111] The indicator of alignment between control and management objectives is the degree to which candidate measures cover two types of objectives: "short-term risk containment" and "long-term structural stability." (Specifically determined by reasoning based on the knowledge-enhanced large model and the "Technical Standard for Safety Management of Concrete Gravity Dams");

[0112] Feasibility indicators for implementation include the matching degree between the implementation cost, construction period, and on-site resources of the candidate measures. (The specific details are determined based on real-time data such as on-site equipment configuration and personnel numbers.)

[0113] Calculate the overall priority score for each candidate measure using the following formula: ;in , , These represent the weighting coefficients corresponding to the risk impact scope indicator, the control objective alignment indicator, and the implementation condition feasibility indicator, respectively.

[0114] Candidate measures by Sort by high to low and select the measures based on the previously set number of items;

[0115] The sorted measures are associated with the industry standard clauses retrieved from S51 to generate a complete control decision plan. The plan includes: control objectives, specific measures, implementation steps, expected results, and professional basis.

[0116] Through the process from S51 to S53, a control and decision-making scheme that is adapted to the current working conditions of the gravity dam, has feasibility and professional basis is generated, providing clear guidance for the implementation of subsequent risk control measures.

[0117] In this application, step S6, the dynamic update process of the spatiotemporal knowledge graph, specifically includes the following steps:

[0118] S61: Collect data on the execution process of the decision-making plan, on-site feedback data, and effect verification data;

[0119] The execution process data includes key construction parameters (such as grouting pressure, grouting volume, and construction time) and equipment operation parameters (such as the measurement accuracy of testing equipment and the operating load of construction machinery) during the implementation of control measures. The data is collected synchronously through on-site sensors and electronic records of construction logs and transmitted to the data storage module in real time.

[0120] On-site feedback data is recorded by on-site technicians using standardized forms, covering abnormal situations that occur during the implementation of measures (such as increased seepage in local dam bodies, problems with the adaptability of construction technology), environmental interference factors (such as sudden rainfall, minor changes in geological structure), personnel operation adjustment records, etc. The form fields are designed in strict accordance with the requirements of the "Regulations for Quality Inspection and Evaluation of Water Conservancy Project Construction".

[0121] The effect verification data is obtained by collecting core monitoring data related to risks according to a preset monitoring cycle after the implementation of control measures, including the horizontal displacement, vertical displacement, seepage flow rate, concrete stress, etc. of the dam body. The data collection equipment is the same as that in step S3;

[0122] All collected data is accompanied by a timestamp (using the ISO8601 time format) and spatial coordinates (lon, lat, alt) (using the EPSG:4326 spatial reference system), which are consistent with the spatio-temporal standards of the spatio-temporal knowledge graph, providing spatio-temporal dimension support for subsequent knowledge extraction;

[0123] S62: Analyze the deviation between the implementation effect and the expected effect through a knowledge-enhanced large model, and extract new risk factors, control measures and their relationships; the specific process is as follows:

[0124] Calculate the deviation degree between the implementation effect and the expected effect of the control measures. Taking the core monitoring indicators as the evaluation objects, the deviation degree calculation formula is: ; where is the number of core monitoring indicators (for example, if there are 3 core indicators such as displacement, seepage flow rate, and stress, then n = 3), is the actual monitoring value of the i-th indicator, is the expected effect value of the i-th indicator, is the weight of the i-th indicator (determined based on the impact degree of the indicator on risk control, such as seepage flow rate weight, displacement weight, stress weight, which is reviewed and determined by the Water Conservancy Project Risk Assessment Expert Committee);

[0125] The knowledge-enhanced large model combines the numerical characteristics of the deviation degree Err to analyze the causes of the deviation:

[0126] If Err < 0.1, it is determined that the effect meets the standard, and focus on extracting the optimization experience suitable for the current working condition in the plan; if 0.1 ≤ Err < Err ≥ 0.3, it is determined that there is a slight deviation, and analyze the surface causes such as construction process adjustment and environmental factor interference; if, it is determined that there is a significant deviation, and deeply explore the potential risk factors, control measure defects or relationships not covered by the existing knowledge graph;

[0127] Based on the analysis results of the deviation causes, extract new knowledge units:

[0128] New risk factors, such as "abnormal setting rate of grouting materials caused by extreme temperature fluctuations" and "incomplete plugging of concealed water seepage channels in geological faults", and clarify their names, influence ranges and triggering conditions;

[0129] New control measures, such as the "directional grouting plugging technology" for "concealed water seepage channels", and clarify their construction parameters, applicable scenarios and implementation conditions;

[0130] New relationships, such as the causal relationship of "extreme temperature fluctuation (risk factor) - decrease in grouting solidification rate (intermediate state) - attenuation of reinforcement effect (risk outcome)", and the constraint relationship of "directional grouting sealing technology (control measures) - hidden seepage channels (risk factor)", ensure that the relationship type is consistent with the relationship set R in step S1;

[0131] S63: After verification, the newly extracted knowledge units are integrated into the spatiotemporal knowledge graph, updating entity attributes, relationship paths, and evolution rules to achieve the autonomous evolution of the knowledge graph; the specific process includes:

[0132] First, perform double verification on the new knowledge unit:

[0133] For the knowledge units extracted from S62, such as new risk factors, new control measures, and new relationships, an effectiveness verification mechanism combining expert review and data consistency verification was used.

[0134] The expert review mechanism involves organizing no fewer than a number of senior engineers with experience in gravity dam safety management to review the scientific validity (such as the triggering logic of new risk factors), feasibility (such as the adaptability of new control measures to construction conditions), and logical validity (such as the causal correspondence of new relationships) of knowledge units. Only knowledge units that receive "pass" comments from all reviewers can proceed to the integration stage.

[0135] The data consistency verification mechanism compares the attribute data of the new knowledge unit (such as the parameter threshold of the new control measures) with the same attributes (such as the parameter range of the existing control measures) and industry standards (such as the "Technical Specification for Safety Monitoring of Concrete Gravity Dams") in the attribute set A constructed in step S1. If there are numerical conflicts or logical contradictions, the process returns to S62 for re-analysis. If the attribute dimension is missing, it is supplemented and improved according to the attribute definition standard (static attribute / dynamic attribute classification) in step S1.

[0136] Further integration and multi-dimensional updates of the knowledge unit graph:

[0137] Specifically, the validated knowledge units will be integrated according to the structural specifications of the spatiotemporal knowledge graph, completing updates in the following dimensions:

[0138] Entity attributes are updated. If the new knowledge unit is a new entity (such as a new risk factor or a new control measure), a corresponding entity node is added, a unique identifier ID is assigned, and its spatial coordinates (using EPSG:4326 spatial reference system) and time range (using ISO8601 time format) attributes are associated. The entity category is assigned to the category corresponding to the entity set E in step S1 (such as a new risk factor being assigned to "risk factor entity").

[0139] If the new knowledge unit is supplementary information for an existing entity, then the attributes of that entity are updated: static attributes (such as the applicable dam height range of control measures) are supplemented with new parameter ranges, and dynamic attributes (such as the real-time monitoring threshold of risk factors) are supplemented with new data sequences and corresponding timestamps;

[0140] The relationship path is updated by mapping the new relationship to the graph in the form of edges, connecting the corresponding entity nodes. The relationship type is consistent with the relationship set R (causal relationship, spatiotemporal relationship, etc.) in step S1, and the relationship is assigned the corresponding influence weight (determined based on expert review opinions) and spatiotemporal constraints (such as the time interval and spatial range of the relationship's effectiveness).

[0141] Evolutionary rules are updated by optimizing the evolutionary rules of the spatiotemporal knowledge graph based on new relationships. This includes risk propagation rules (such as adding a propagation link of "new risk factor → specific entity anomaly") and control measure effectiveness rules (such as the timeliness of the effect of new control measures on target risks), ensuring the logical consistency between rules and entities and relationships.

[0142] Next, we will optimize the graph index synchronization, and update the retrieval index of the spatiotemporal knowledge graph synchronously to ensure the query and traversal efficiency of the updated graph:

[0143] The spatiotemporal composite index is optimized by integrating the spatiotemporal information (spatial coordinates and time range) of newly added entities and relationships into the R-tree spatial index and B-tree temporal index constructed in step S14, ensuring the retrieval accuracy of the spatiotemporal dimension.

[0144] Vector index optimization employs the knowledge embedding algorithm from step S21 to convert the attributes of newly added entities and relationships into low-dimensional vectors, which are then integrated into the vector index of the graph (KD tree structure) to ensure that the retrieval enhancement architecture of the knowledge-enhanced large model can efficiently match the newly integrated knowledge units.

[0145] Through the S63 process, the spatiotemporal knowledge graph completes the updating of entities, attributes, relationships, and rules, realizing the autonomous expansion and iteration of the knowledge system, and providing more accurate and comprehensive structured knowledge support for subsequent gravity dam safety risk management.

[0146] In this application, the data preprocessing process in step S3 specifically includes the following steps:

[0147] An outlier removal algorithm is used to process outliers in the monitoring data. Data units are standardized or normalized using a standardization or normalization algorithm. Text features in unstructured documents are extracted using optical character recognition technology. Morphological feature parameters in monitoring images are extracted using an image feature extraction algorithm.

[0148] Domain-specific training includes pre-training fine-tuning based on domain corpora, cue engineering optimization, and supervised learning training with a small number of labeled samples.

[0149] In this application, the retrieval process in step S51 specifically adopts a dual retrieval mechanism of semantic similarity matching and working condition feature vector comparison. The priority ranking criteria in step S53 include the cost of implementing the measures, the speed of risk mitigation, the construction complexity, and the degree of impact on the normal operation of the gravity dam.

[0150] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A gravity dam safety risk management method based on knowledge graphs and large models, characterized in that, Includes the following steps: S1: Construct a spatiotemporal knowledge graph of gravity dam safety risks. The spatiotemporal knowledge graph includes an entity layer, a relation layer, and an attribute layer to realize the structured association and spatiotemporal dimension mapping of multi-source data throughout the entire life cycle of gravity dams. S2: Based on the retrieval-enhanced generation architecture and domain-adaptive training, a knowledge-enhanced large model is constructed, and the spatiotemporal knowledge graph is embedded as a structured knowledge source into the reasoning process of the large model; S3: Collect real-time operation data and full life-cycle historical data of gravity dams, and after data cleaning, standardization and feature extraction processing, input them into the knowledge-enhanced large model; S4: By calling the entity relationship and spatiotemporal evolution rules of the spatiotemporal knowledge graph through the knowledge-enhanced large model, risk anomaly identification, correlation path mining and root cause location are completed, and risk assessment results containing traceability evidence are generated. S5: Based on the risk assessment results, the knowledge-enhanced big model retrieves historical cases, industry standards and control rules from the spatiotemporal knowledge graph to generate interpretable risk control decision-making schemes that are adapted to the current working conditions; S6: Collect the execution data and effect evaluation results of the risk management decision-making scheme, extract new entities, relationships and rules through the knowledge-enhanced big data model, and dynamically update the spatiotemporal knowledge graph.

2. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 1, characterized in that, In step S1, the entity layer includes dam structure entity, monitoring facility entity, risk factor entity, historical case entity, control measure entity, and industry standard entity; The relationship layer includes causal relationships, spatiotemporal relationships, subordinate relationships, and constraint relationships between entities; the attribute layer includes static attributes and dynamic attributes of entities. Static attributes include design parameters, structural features, and specification thresholds, while dynamic attributes include real-time monitoring data, operating status parameters, and timestamp information.

3. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 1, characterized in that, In step S1, the construction process of the spatiotemporal knowledge graph includes: S11: Collect multi-source heterogeneous data on gravity dams, including design documents, construction records, monitoring data, historical incident cases, maintenance records, and industry standard texts; S12: Using natural language processing technology and graph neural network algorithms, entity, relation and attribute information is extracted from the data, and after verification and correction, it is formed into structured knowledge units; S13: Based on the OGC standard, a spatiotemporal correlation model is constructed, which binds structured knowledge units with spatial location information and time series information to form spatiotemporal knowledge triples; S14: Use a graph database to store spatiotemporal knowledge triples to construct a spatiotemporal knowledge graph of gravity dam safety risks with spatiotemporal query and path traversal capabilities.

4. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 1, characterized in that, In step S2, the construction process of the knowledge-enhanced large model includes: S21: Use a knowledge embedding algorithm to convert entities, relations and attributes in the spatiotemporal knowledge graph into low-dimensional vectors and integrate them into the word embedding layer of the large model; S22: Construct a retrieval-enhanced generative architecture to establish a real-time connection between the spatiotemporal knowledge graph and the large model, enabling the large model to dynamically retrieve relevant structured knowledge during reasoning; S23: Collect labeled data in the field of gravity dam safety management and conduct domain-adaptive training on the large model.

5. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 1, characterized in that, In step S4, the risk tracing process specifically includes the following steps: S41: The knowledge-enhanced large model compares the preprocessed data with the normal operation threshold in the spatiotemporal knowledge graph to identify risk and abnormal signals and corresponding related entities; S42: Traverse the direct and indirect relationship paths of related entities in the spatiotemporal knowledge graph to uncover the key links of risk propagation; S43: Analyze the influence weight of each entity in the critical link based on the reasoning ability of the large model, and locate the root cause of the risk by combining the spatiotemporal evolution rules; S44: Connect professional evidence in the spatiotemporal knowledge graph to generate interpretable risk assessment results that include risk causes, evolution paths, and judgment criteria.

6. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 1, characterized in that, In step S5, the process of generating the control decision plan includes: S51: The knowledge-enhanced big model retrieves matching historical success cases and relevant industry standards from the spatiotemporal knowledge graph based on the root causes of risks and the characteristics of current working conditions. S52: Optimize the retrieved historical control measures to form a set of candidate measures that conform to the current structural characteristics and operating status of the gravity dam; S53: Based on the scope of risk impact, control objectives, and implementation conditions, prioritize candidate measures and generate a complete decision-making plan that includes control objectives, specific measures, implementation steps, expected effects, and professional basis.

7. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 1, characterized in that, In step S6, the dynamic update process of the spatiotemporal knowledge graph specifically includes the following steps: S61: Collect data on the execution process of the decision-making plan, on-site feedback data, and effect verification data; S62: Analyze the deviation between the execution effect and the expected effect through knowledge-enhanced large model, and extract new risk factors, control measures and correlations; S63: After verification, the newly extracted knowledge units are integrated into the spatiotemporal knowledge graph, updating entity attributes, relationship paths, and evolution rules to achieve the autonomous evolution of the knowledge graph.

8. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 1, characterized in that, The data preprocessing process in step S3 specifically includes the following steps: An outlier removal algorithm is used to process outliers in the monitoring data. Data units are standardized or normalized using a standardization or normalization algorithm. Text features in unstructured documents are extracted using optical character recognition technology. Morphological feature parameters in monitoring images are extracted using an image feature extraction algorithm.

9. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 4, characterized in that, The knowledge embedding algorithm in step S21 specifically adopts one or more combinations of TransE, TransH, TransR, or ComplEx algorithms; The domain adaptation training includes pre-training fine-tuning based on domain corpus, cue engineering optimization, and supervised learning training with a small number of labeled samples.

10. The gravity dam safety risk management method based on knowledge graphs and large models according to claim 6, characterized in that, The retrieval process in step S51 specifically adopts a dual retrieval mechanism of semantic similarity matching and working condition feature vector comparison. The priority ranking criteria in step S53 include the implementation cost of the measures, the speed of risk mitigation, the construction complexity, and the degree of impact on the normal operation of the gravity dam.

Images