A knowledge graph driven landslide risk assessment method and related device

Abstract

The application relates to the technical field of landslide risk assessment, and provides a landslide risk assessment method driven by a knowledge graph and related equipment, which comprises the following steps: acquiring a plurality of landslide case texts, and performing knowledge extraction on all the landslide case texts to obtain a landslide knowledge graph; mapping a to-be-evaluated landslide task into the landslide knowledge graph to generate a task vector of the to-be-evaluated landslide task; based on the task vector, performing model screening from the landslide knowledge graph to obtain a final evaluation model of the to-be-evaluated landslide task; based on the final evaluation model, performing data screening from the landslide knowledge graph to obtain final landslide data corresponding to the to-be-evaluated landslide task; and performing risk evaluation on the to-be-evaluated landslide task by using the final evaluation model according to the final landslide data to obtain a landslide risk evaluation result. The method can improve the reliability and efficiency of landslide risk evaluation.

Description

Technical Field

[0001] This application relates to the field of landslide risk assessment technology, and in particular to a knowledge graph-driven landslide risk assessment method and related equipment. Background Technology

[0002] Landslides are among the most widespread and destructive natural disasters globally, and the accuracy of their risk assessment directly impacts the effectiveness of disaster prevention and mitigation efforts. In recent years, with the deep integration of spatiotemporal big data and artificial intelligence technologies, landslide risk assessment research is gradually shifting from traditional numerical simulation to a new knowledge-driven paradigm.

[0003] Landslide risk assessment is a core component of geological disaster prevention and control research, aiming to quantitatively or qualitatively evaluate the probability of landslide occurrence and the potential losses through scientific methodologies. With the rapid development of Earth observation technology, big data, and artificial intelligence, landslide risk assessment has gradually evolved from traditional empirical qualitative evaluation and simple mathematical statistical analysis into a complex systems engineering approach integrating multiple disciplines and technologies. Currently, scholars both domestically and internationally have raised higher requirements for assessment accuracy, response speed, and automation, driving profound changes in assessment systems in terms of model-driven approaches, technical pathways, evaluation dimensions, and granularity, showing a trend towards model diversification, technological automation, multi-source data, and refined scale.

[0004] With the rapid development of Earth observation technology and the advancement of information technology in disaster prevention and mitigation, landslide risk assessment research has accumulated massive amounts of monitoring data covering multiple sources and spatiotemporal scales, leading to the emergence of numerous risk assessment models. Existing landslide risk assessment models can be categorized into three types: physical models, data-driven models, and knowledge-driven models. Physical models are primarily based on the principles of classical mechanics, hydrology, and soil mechanics, simulating the physical processes of landslides through mathematical equations. These models typically focus on the stress equilibrium state of the slope and use slope stability coefficients to quantitatively assess risk. Data-driven models do not require explicit physical evolution equations but rather predict landslide risks by mining the nonlinear correlations between historical landslide samples and environmental impact factors. These models include empirical statistical models, machine learning models, and deep learning models. Knowledge-driven models rely heavily on the prior knowledge and experience of industry experts, assigning different weights to various influencing factors to qualitatively or semi-quantitatively assess landslide risk. Their core models include the analytic hierarchy process (AHP), fuzzy comprehensive evaluation, and expert scoring methods. The landslide risk assessment data accumulated through research can be divided into multi-source heterogeneous data, including topographic data, geological environment data, hydrological and meteorological data, socio-economic data, land cover data, engineering activity data, and disaster history data.

[0005] However, faced with massive databases and disorganized model libraries, practical applications still rely on the traditional experience-driven model of "humans finding data and selecting models." Semantic gaps and mismatches exist between multi-source heterogeneous data and multi-dimensional models, leading to a high dependence on human intervention in risk assessment data selection and model application. Therefore, this model presents a severe challenge in matching models and data. On the one hand, due to the lack of unified matching standards, manual screening often struggles to quickly and accurately determine suitable assessment models and data when facing different regions and types of landslides. On the other hand, this model, heavily reliant on expert experience, results in a lengthy response chain and limited assessment efficiency, leading to poor reliability and efficiency in landslide risk assessment. Summary of the Invention

[0006] This application provides a knowledge graph-driven landslide risk assessment method and related equipment, which can solve the problems of poor reliability and efficiency in landslide risk assessment.

[0007] In a first aspect, embodiments of this application provide a knowledge graph-driven landslide risk assessment method, which includes: Multiple landslide case texts are obtained, and knowledge is extracted from all landslide case texts to obtain a landslide knowledge graph. The landslide knowledge graph includes multiple landslide cases, as well as the evaluation model and landslide data corresponding to each landslide case. Map the landslide task to be evaluated to the landslide knowledge graph to generate the task vector of the landslide task to be evaluated; Based on task vectors, models are selected from the landslide knowledge graph to obtain the final evaluation model for the landslide task to be evaluated. Based on the final evaluation model, data is filtered from the landslide knowledge graph to obtain the final landslide data corresponding to the landslide task to be evaluated. Based on the final landslide data, the final assessment model is used to conduct a risk assessment of the landslide task to be assessed, and the landslide risk assessment results are obtained.

[0008] Secondly, embodiments of this application provide a knowledge graph-driven landslide risk assessment device, comprising: The acquisition module is used to acquire multiple landslide case texts and extract knowledge from all landslide case texts to obtain a landslide knowledge graph. The landslide knowledge graph includes multiple landslide cases, as well as the evaluation model and landslide data corresponding to each landslide case. The generation module is used to map the landslide task to be evaluated to the landslide knowledge graph and generate the task vector of the landslide task to be evaluated. The model selection module is used to select models from the landslide knowledge graph based on task vectors to obtain the final evaluation model for the landslide task to be evaluated. The data filtering module is used to filter data from the landslide knowledge graph based on the final evaluation model to obtain the final landslide data corresponding to the landslide task to be evaluated. The assessment module is used to conduct a risk assessment of the landslide task to be assessed based on the final landslide data and the final assessment model, and obtain the landslide risk assessment results.

[0009] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the knowledge graph-driven landslide risk assessment method described above.

[0010] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the knowledge graph-driven landslide risk assessment method described above.

[0011] The above-mentioned solution in this application has the following beneficial effects: In the embodiments of this application, multiple landslide case texts are acquired, and knowledge extraction is performed on all landslide case texts to obtain a landslide knowledge graph. Then, the landslide task to be assessed is mapped onto the landslide knowledge graph, generating a task vector for the landslide task to be assessed. Based on the task vector, model selection is performed from the landslide knowledge graph to obtain the final assessment model for the landslide task to be assessed. Then, based on the final assessment model, data selection is performed from the landslide knowledge graph to obtain the final landslide data corresponding to the landslide task to be assessed. Finally, based on the final landslide data, the final assessment model is used to conduct a risk assessment of the landslide task to be assessed, obtaining the landslide risk assessment result. The construction of the landslide knowledge graph accumulates knowledge of landslide cases, forming a structured knowledge system that provides data support for landslide risk assessment. Model and data selection from the landslide knowledge graph enables autonomous model and data scheduling, improving the relevance of the model and data to the landslide task to be assessed, and effectively improving the reliability and efficiency of landslide risk assessment.

[0012] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1A flowchart of a knowledge graph-driven landslide risk assessment method provided in an embodiment of this application; Figure 2 This is a schematic diagram of a knowledge extraction ontology provided in an embodiment of this application; Figure 3 A schematic diagram of the structure of a knowledge graph-driven landslide risk assessment device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0015] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0016] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0017] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0018] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0019] To address the issues of poor reliability and efficiency in existing landslide risk assessment methods, this application provides a knowledge graph-driven landslide risk assessment method. This method constructs a landslide knowledge graph, accumulates knowledge of landslide cases, and forms a structured knowledge system, providing data support for landslide risk assessment. It also performs model and data screening from the landslide knowledge graph, enabling autonomous model and data scheduling, improving the relevance of models and data to the landslide assessment task, and effectively enhancing the reliability and efficiency of landslide risk assessment.

[0020] The following is an illustrative example of the knowledge graph-driven landslide risk assessment method provided in this application.

[0021] like Figure 1 As shown, the knowledge graph-driven landslide risk assessment method provided in this application includes the following steps: Step 11: Obtain multiple landslide case texts and extract knowledge from all landslide case texts to obtain a landslide knowledge graph.

[0022] The aforementioned landslide knowledge graph includes multiple landslide cases, along with corresponding assessment models and landslide data for each case. These landslide cases can be actual landslide accident cases, landslide analysis cases, landslide assessment cases, etc. The assessment models are models applicable to landslide risk assessment, such as physics-driven models (e.g., Fast Shallow Landslide Assessment Model, FSLAM), empirical statistical models (e.g., logistic regression), machine learning models (e.g., support vector machines), deep learning models (e.g., fully connected neural network (FCNN), and coupled models (e.g., convolutional neural network-deep mixture (CNN-MD)). The landslide data includes landslide-related data, such as topographic data, geological environment data, and hydrological and meteorological data for the landslide area. The aforementioned landslide case texts can be news articles, scientific and technological literature, etc., related to landslides, including the time, location, and scale of the landslide, as well as information such as environmental and topographical factors that affect the landslide. A large language model can be used to extract knowledge from the landslide case texts, integrate landslide-related information into landslide cases, and match them with an assessment model that can be used for analysis, thus integrating the data in the texts into landslide data.

[0023] In some embodiments of this application, landslide case texts can be obtained by accessing publicly available academic databases, and knowledge extraction can be performed on all landslide case texts using a large language model to obtain a landslide knowledge graph.

[0024] For example, after obtaining the text of landslide case studies, it can be preprocessed by cleaning, such as: First, a search strategy was developed for mainstream academic databases, using "landslide risk assessment" and "landslide risk evaluation" as core keywords. This strategy combined manual screening with automated crawling to acquire scientific and technological literature resources covering Chinese core journals and above. After collection, the downloaded landslide-related scientific and technological literature was deduplicated, its quality assessed, and metadata extracted to construct a landslide scientific and technological literature database in original portable document format (PDF).

[0025] Secondly, to address the complex multi-column layout, nested formulas, and mixed text and image formats in scientific and technological literature, multimodal conversion technology is employed. Image conversion technology is used to uniformly convert PDF-formatted scientific and technological literature into high-fidelity image sequences at high resolution. Furthermore, a deep learning-based layout analysis algorithm is introduced to physically analyze the imaged pages, accurately identifying and labeling the text area, title area, figure and table area, formula area, and reference area.

[0026] Finally, deep optical character recognition (OCR) and high-precision structure extraction are performed. Automated OCR tools, including physical analysis, graph recognition, and semantic segmentation, are used to deeply analyze and clean unstructured content such as text, abstracts, figures, and conclusions in the documents, forming a standardized corpus of scientific and technological literature.

[0027] For example, knowledge extraction using large models can be performed through the following process: First, an extraction ontology and prompt word design for landslide risk assessment are conducted. A multi-dimensional knowledge extraction ontology is constructed, encompassing "landslide case studies, assessment models, and suitable data." The landslide knowledge extraction ontology covers a range of knowledge related to landslide case studies used in risk assessment within scientific literature. The model knowledge extraction ontology covers a range of knowledge related to model algorithms, inputs and outputs, and performance evaluations used in landslide risk assessment analysis. The data knowledge extraction ontology covers a range of knowledge related to the data primarily used in model operation, as well as its semantics and constraints. Based on these knowledge extraction ontologs, a set of extraction prompt word instructions with strong content constraints is designed. Using a thought chain technique, the extraction content, target entity type, relational logic specifications, and output format are explicitly defined in the prompt words to guide the large language model to achieve accurate knowledge extraction with zero samples.

[0028] Secondly, leveraging the powerful semantic understanding capabilities of Large Language Models (LLM), knowledge extraction tasks are performed under explicit cue word constraints. This process encompasses trigger word identification and further performs entity and relation extraction. For trigger word identification, semantic matching algorithms are used to identify trigger words describing landslide events, model usage, or data usage in scientific literature, thus pinpointing knowledge-intensive areas. For entity and attribute extraction, landslide case entities, evaluation model entities, and adaptation data entities are automatically identified and extracted from the structured corpus parsed by OCR, while simultaneously identifying and extracting relevant attribute information. For relation extraction, the inherent relationships between entities are identified and extracted to ensure the logical coherence of the knowledge.

[0029] Finally, the extracted fragmented information is mapped into a standardized triple structure of "entity-relationship-attribute / entity", which accurately captures the core attributes and internal logical connections of the models and data used in landslide risk assessment and analysis, realizing the transformation from massive text to structured knowledge.

[0030] For the extracted multi-source heterogeneous triples, knowledge fusion is performed using techniques such as entity alignment, attribute merging, and relation resolution to eliminate semantic ambiguity and data redundancy.

[0031] Subsequently, the fused high-quality knowledge will be imported into the Neo4j graph database for storage, ultimately constructing a landslide knowledge graph with a complex network of relationships for landslide risk tasks.

[0032] It should be noted that the landslide knowledge graph categorizes stored data into multiple types, including model classification and data classification. Then, it analyzes the compatibility between landslide cases and assessment models, as well as the resource support relationship between assessment models and landslide data, thereby establishing a logical connection between "task-model-data". For example: For model classification: First, based on the task requirements of landslide risk assessment, a multi-dimensional model semantic classification framework is predefined, defining multi-dimensional model classification criteria from the dimensions of mechanism, function, and scale. At the mechanism level, models can be categorized into physics-driven models, empirical statistical models, machine learning models, deep learning models, and coupled models. At the function level, since landslide risk assessment involves multiple aspects, it can be divided into susceptibility assessment, hazard assessment, vulnerability assessment, and risk assessment models. At the scale level, models are categorized into single landslide analysis models, small-area landslide analysis models, and large-area landslide analysis models.

[0033] Building upon this foundation, structured queries are performed on the model subgraph within the landslide knowledge graph constructed in Step 1 to obtain multi-source semantic information of model instances, including attributes such as model name, model type, applicable landslide type, and applicable spatial scale. Subsequently, this semantic information is transformed into a unified multi-dimensional feature representation, constructing feature mapping rules oriented towards model classification criteria. Using a deep semantic matching operator, the mapping from model instances to the classification criterion space is achieved by calculating the vector similarity between the feature vectors of model instances and the classification criterion vectors. Multi-dimensional joint discrimination is then performed based on the similarity magnitude, thereby completing the automatic classification and logical categorization of models across multiple dimensions, including mechanism, function, and scale.

[0034] Subsequently, we will conduct refined modeling of the model's computational attributes and capabilities, defining the model's input constraints (such as essential data items and input parameters), output semantics (such as the physical meaning of the model's output results), and performance metrics (such as model complexity, model accuracy, and model usability).

[0035] Finally, after completing the automated model classification, a top-down hierarchical inheritance architecture is constructed. Based on the hierarchical structure, non-hierarchical horizontal relationships are introduced. Model hierarchy and relationship design are achieved by calculating the hierarchical design of "parent class-child class-concrete model" and the horizontal relationships of evolution or combination between models. For example, "landslide risk assessment model" is defined as the top-level parent class, under which the subclass "machine learning model" differentiates, ultimately mapping to specific model instances such as "support vector machine" at the end nodes. This ultimately constructs a rigorously structured model knowledge classification system.

[0036] For data classification: First, based on the task requirements of landslide geological hazard risk assessment, a multi-dimensional data semantic classification framework is predefined, defining multi-dimensional data classification criteria from the dimensions of data attributes, evaluation factors, and spatiotemporal dimensions. In the data attribute dimension, data is divided into topographic data, geological environment data, hydrological and meteorological data, socioeconomic data, land cover data, engineering activity data, and historical disaster data. In the evaluation factor dimension, data is divided into disaster-inducing environment data, triggering factor data, and disaster-bearing body data. In the spatiotemporal dimension, data can be divided into time and spatial dimension data. Based on this, the data subgraphs are queried and relevant data knowledge is extracted from the landslide knowledge graph constructed in step one. Combined with knowledge such as data instance names, a multi-dimensional semantic feature vector of data instances and data classification criteria is constructed. A deep semantic matching operator is used to calculate the spatial mapping similarity between data instances and classification criteria, achieving automatic cross-dimensional data discrimination and logical classification.

[0037] Subsequently, refined modeling of data semantic attributes and relationships was conducted. This involved defining the data's physical attributes (such as intrinsic attributes like accuracy and resolution) and indicator attributes (such as maximum precipitation and slope range). The relationships between data were then established. Specifically, through the semantic chain of "source data—calculated data," combined with pre-defined attribute inheritance and constraint rules, knowledge reasoning was performed to map key attributes such as resolution and spatial range of the source data to the calculated data, thus automatically establishing the relationships between data. For example, the slope data calculated from Digital Elevation Model (DEM) data should maintain a consistent resolution with the DEM source data. Ultimately, a rigorously structured data classification system was constructed, providing a logical foundation for the subsequent automated selection of data.

[0038] The logical coordination mechanism among the three is defined through formal language: first, a "task-model" demand-driven chain is constructed, clarifying the specific evaluation task ( ) on the candidate model set ( The selection and adaptation relationship of the model; secondly, to construct a "model-data" resource support chain to characterize the model ( ) run on heterogeneous datasets ( The process involves three key steps: first, matching input dependencies and constraints; second, constructing a "data-task" outcome evolution chain, defining how the task's output dataset is fed back and enriches the evaluation system. Through the synergy of these three chains, a semantic closed loop is achieved throughout the entire risk assessment process. For example, in the "task-model" demand-driven chain, specifically for the "large-scale rainfall-induced landslide risk assessment task" (…),… This task can be combined with large-area landslide analysis models and landslide risk assessment models. , Establish connections; in the "model-data" resource support chain, the selected model ( It needs to match multi-source datasets such as rainfall data, DEM data, and land cover data. , , The model outputs risk zoning data, which can be used to update the understanding of landslide susceptibility in the region and serve as the prior data basis for subsequent similar tasks, while also meeting the constraints of resolution and time scale.

[0039] Based on the aforementioned multi-dimensional model and data classification system and the "task-model-data" relationship logic, a formal expression is carried out using ontology modeling tools. A five-tuple model is defined, consisting of: Concepts (including model classification systems, data classification systems, and landslide case knowledge systems), Relations (including relationships between demand-driven chains, resource support chains, and outcome evolution chains; relationships between model computational attributes and capabilities; model hierarchy and relationships; and data semantic attributes and relationships between data), Instances (including specific landslide case instances, specific model instances, and specific data instances), Axioms (including model constraint rules and data constraint rules), and Values ​​(including attribute values ​​of landslide case instances, model instances, and data instances). Through computational ontology modeling, the model classification system, data classification system, demand-driven chain, resource support chain, and outcome evolution chain are mapped to specific ontology components, ultimately forming a standardized, machine-readable computational ontology for landslide risk assessment tasks. Finally, guided by the constructed computational ontology, landslide case knowledge, assessment model knowledge, and matching data knowledge are re-characterized, and the landslide knowledge base is systematically and standardizedly expressed through this unified organizational framework.

[0040] Step 12: Map the landslide task to be evaluated to the landslide knowledge graph to generate the task vector of the landslide task to be evaluated.

[0041] The landslide assessment task described above is text data, such as descriptive text of the landslide area requiring risk assessment and text outlining the purpose of the risk assessment. The task vector includes features of the assessment task across multiple feature dimensions.

[0042] In some embodiments of this application, the step of mapping the landslide task to be evaluated to a landslide knowledge graph and generating a task vector for the landslide task to be evaluated includes: The first step is to perform text cleaning and extraction for the landslide assessment task, resulting in multiple discrete feature words.

[0043] For example, a pre-trained deep language model (such as the Bidirectional Encoder Representations from Transformers (BERT) model) can be used to perform global dependency analysis and semantic segmentation on the input natural language task text, and combined with Named Entity Recognition (NER) technology to clean and extract text for the landslide assessment task, resulting in multiple discrete feature words.

[0044] The second step is to map all discrete feature words into a landslide knowledge graph. Based on the mapped landslide knowledge graph, the task to be evaluated is analyzed from multiple feature dimensions to generate a task vector for the task to be evaluated.

[0045] For example, multiple feature dimensions include spatial features, physical mechanism features, and model and data requirement description features. For spatial features, based on the spatial administrative divisions or geometric ranges extracted from discrete feature words, the corresponding spatial multi-level inclusion relationships are retrieved in the landslide knowledge graph. Then, a spatial scale generalization matrix is ​​constructed. The rows and columns of this matrix map the spatial and administrative levels in the graph, respectively, and the matrix elements represent the generalization association weights between different levels. The extracted spatial feature words are used as input vectors, and matrix multiplication is performed with this generalization matrix to calculate the confidence score of the text description at each spatial scale (which can be achieved using Natural Language Processing (NLP)). The system calculates confidence scores for text using models such as Processing, and uses the highest-scoring level as the automatically determined task space scale range. For physical mechanism features, semantic similarity (e.g., cosine similarity) is used to align the inducing factors described by discrete feature words with landslide types in the landslide knowledge graph, identifying the type of landslide disaster. For model and data requirement description features, based on the extracted discrete feature words related to model and data requirements, a path traversal query is performed along pre-defined semantic relationship edges in the graph (e.g., "applicable model" and "data constraint"), directly mapping the abstract features in the task description to specific model applicability nodes and data applicability nodes in the knowledge graph. These graph entity attributes are then converted into corresponding feature codes, achieving automatic association from the task objective described by discrete feature words to the model and data applicability features in the knowledge graph. Through this multi-dimensional mapping, single text labels are transformed into feature space coordinates supported by deep attributes. Finally, features from all feature dimensions are integrated into a single vector to obtain the task vector.

[0046] Step 13: Based on the task vector, the model is screened from the landslide knowledge graph to obtain the final evaluation model for the landslide task to be evaluated.

[0047] The correlation between the final evaluation model and the task vector is greater than that between all other evaluation models.

[0048] In some embodiments of this application, the steps described above for selecting models from a landslide knowledge graph based on task vectors to obtain the final evaluation model for the landslide task to be evaluated include: The first step is to generate the feature vector corresponding to each landslide case in the landslide knowledge graph.

[0049] For example, the feature vector corresponding to each landslide case can be generated according to the method of generating task vectors in step 12. That is, the landslide case is analyzed from each feature dimension to obtain the features under each feature dimension, and all features are integrated into a vector to obtain the feature vector corresponding to the landslide case.

[0050] The second step is to calculate the similarity between the task vector and each feature vector.

[0051] Specifically, through the formula:

[0052] Calculate the task vector for the landslide assessment task. With the Feature vector of a landslide case similarity between .

[0053] in, Indicates the number of feature dimensions. Indicates the first Weights of each feature dimension, Represents the task vector The first in One characteristic, Representing the eigenvector The first in One characteristic, , This represents the number of landslide cases in the landslide knowledge graph. Weights can be set empirically or learned based on historical data. For categorical features, a semantic matching function is used for similarity mapping; for numerical features, a normalized distance metric is used for unified expression.

[0054] The third step is to sort all similarities from largest to smallest, and then select the landslide cases corresponding to the top few similarities in the sorting results as candidate landslide cases.

[0055] For example, the number of candidate landslide cases is a preset number.

[0056] The fourth step is to conduct multi-dimensional model evaluation of all candidate landslide models and determine the final evaluation model for the landslide task to be evaluated from all candidate landslide models.

[0057] For example, path search and constraint consistency detection can be performed on a knowledge graph. Taking the landslide type and spatial scale of the task as the initial boundary, a path search is performed in the landslide model knowledge graph to activate all connectable candidate landslide models. The initial confidence of the candidate models is initialized based on the similarity of landslide risk assessment cases. Then, the integrated model type, model applicability conditions, and model performance evaluation are transformed into a set of hard constraints. Constraint consistency detection is performed on the searched paths to dynamically eliminate failed models that do not meet the boundary constraints due to physical mechanisms or data input. Multi-criteria cross-reasoning and utility weighting are then performed on the candidate models that pass the detection. Finally, by comprehensively balancing the spatial scale fit, model generalization performance, and multi-model integration gain, a global decision on the optimal path is achieved. Ultimately, the evaluation model with the highest fit with the current landslide task environment is dynamically identified and locked as the final evaluation model.

[0058] Step 14: Based on the final evaluation model, data is filtered from the landslide knowledge graph to obtain the final landslide data corresponding to the landslide task to be evaluated.

[0059] The semantic similarity between the final landslide data and the required data for the final evaluation model is greater than that between all other landslide data.

[0060] In some embodiments of this application, the steps described above, based on the final evaluation model, to filter data from the landslide knowledge graph to obtain the final landslide data corresponding to the landslide task to be evaluated, include: The first step is to analyze the landslide cases corresponding to the final evaluation model to obtain the data input constraint tensor of the final evaluation model.

[0061] For example, the data knowledge management agent retrieves the model input knowledge of the final evaluation model from the landslide knowledge graph, extracts the hard boundary indicators of the model in multiple dimensions such as data type, spatial resolution and data accuracy, and uses feature encoding technology to convert the extracted structured constraint indicators of each dimension into numerical feature vectors in a unified format. Then, according to the preset dimension matrix specification, these feature vectors are spliced ​​and aligned in the channel dimension to finally construct a data input constraint tensor that represents all input constraints of the final evaluation model.

[0062] The second step is to transform the data input constraint tensor into a data requirement feature vector.

[0063] The aforementioned data requirement feature vector is used to describe the types, semantics, and other information of the multiple data required for the final evaluation model.

[0064] For example, natural language processing and semantic mapping techniques (such as the Latent Dirichlet Allocation (LDA) algorithm) can be used to transform the data input requirements of the final evaluation model into a standardized data requirement feature vector.

[0065] The third step is to generate the semantic feature vector for each landslide data in the landslide knowledge graph.

[0066] For example, semantic feature vectors for each landslide data point in a landslide knowledge graph can be generated using word vector models (Word2Vec, Word to Vector).

[0067] The fourth step is to calculate the semantic similarity between each data requirement feature vector and each semantic feature vector.

[0068] Specifically, through the formula:

[0069] Calculate the feature vector of data requirements With the Semantic feature vector of landslide data semantic similarity between .

[0070] in, This represents the length of the demand feature vector. This indicates the number of the last feature in the demand feature vector. Represents the feature vector of data requirements The first in One characteristic, Represents semantic feature vector The first in One characteristic, , This indicates the number of landslide data points in the landslide knowledge graph.

[0071] The fifth step is to use the landslide data with the highest semantic similarity as the final landslide data for the landslide task to be evaluated.

[0072] Step 15: Based on the final landslide data, use the final assessment model to conduct a risk assessment of the landslide task to be assessed, and obtain the landslide risk assessment results.

[0073] The landslide risk assessment results mentioned above are used to describe the risk status of the landslide task to be assessed, and can be classified labels, such as high landslide risk, low landslide risk, etc.

[0074] Specifically, the final landslide data is adapted and corrected based on the final assessment model; the corrected final landslide data is used as input data for the final assessment model, and the final assessment model is used to conduct a risk assessment of the landslide task to be assessed, so as to obtain the landslide risk assessment results.

[0075] For example, to address potential anomalies in landslide risk assessment caused by data accuracy mismatches or differences in physical dimensions, a full-process adaptation and consistency check is performed. A closed-loop conflict resolution logic is constructed for the negotiation mechanism: when an incompatibility risk is detected between the operating conditions of the final assessment model and the currently matched final landslide data, a collaborative negotiation process among multiple agents is automatically triggered. The landslide risk assessment agent simultaneously sends conflict messages to the model knowledge management agent and the data knowledge management agent. The model knowledge management agent prioritizes evaluating the adjustment threshold of the model's internal control parameters. If the conflict is within the threshold (e.g., the model defaults to requiring hourly rainfall data while only 12-hour data is available), the model knowledge management agent performs adaptive parameter adjustment, dynamically correcting the differential iteration step size from 1 hour to 12 hours for backward compatibility. If the conflict exceeds the threshold (e.g., the model mandates a 5-meter resolution DEM while only 30-meter data is available), the negotiation decision is transferred to the data knowledge management agent, which calls backup onboard point cloud data or a higher-order interpolation algorithm to perform super-resolution dynamic reconstruction and incremental completion of the landslide data. The two intelligent agents negotiate through multiple rounds of "model parameter fine-tuning - data feature reconstruction" until the feature differences converge to within the consistency threshold, thus achieving dynamic correction of the final landslide data.

[0076] It is worth mentioning that the construction of the landslide knowledge graph has accumulated knowledge of landslide cases and formed a structured knowledge system, providing data support for landslide risk assessment. Model and data screening can be carried out from the landslide knowledge graph to achieve autonomous model and data scheduling, improve the relevance of models and data to the landslide tasks to be assessed, and effectively improve the reliability and efficiency of landslide risk assessment.

[0077] Furthermore, this application has the following advantages: This application leverages the powerful natural language processing capabilities and deep semantic understanding of large-scale models to extract landslide knowledge, model knowledge, and data knowledge used in landslide risk assessment and analysis from massive amounts of unstructured literature. It then constructs a knowledge graph for structured storage, forming a structured knowledge system. This addresses the problem of insufficient "model-data" knowledge modeling in the multidimensional semantic association of landslides, providing a structured data system and decision support for my country's landslide geological disaster prevention and control system.

[0078] This application constructs a computational ontology of "task-model-data" to fully characterize model knowledge and data knowledge. By building a landslide risk assessment intelligent agent, a model intelligent agent, and a data intelligent agent, it transforms the landslide risk assessment model from "inefficient manual adaptation" to "intelligent autonomous scheduling." This breaks through the limitations of the traditional "human-finds-data, human-selects-model" model, which leads to delayed early warnings when facing clustered and sudden landslide disasters. It effectively improves the level of intelligence in geological disaster prevention and control, and provides core support and a scientific path for proactive prevention and control of smart geological disasters.

[0079] This application introduces a knowledge graph-driven and multi-agent collaborative landslide risk assessment model and data intelligent adaptation technology, solving the problem of insufficient "model-data" collaborative adaptation capability in landslide risk assessment application scenarios. This automated and intelligent adaptation and computing mode wins valuable time windows for accurate pre-disaster early warning and post-disaster emergency command, helping to improve the timeliness and accuracy of response to sudden disasters.

[0080] The method of this application will be illustrated below with a specific example.

[0081] In this embodiment, the search time range for mainstream Chinese academic databases is set from 1996 to 2025 (the last thirty years). Keywords such as "landslide risk assessment" and "landslide risk evaluation" are used as search strategies. A combination of manual screening and automated crawling is employed to obtain scientific and technological literature resources covering core Chinese journals and above. The downloaded landslide-related scientific and technological literature is deduplicated, its quality is assessed, and metadata is extracted to construct a full-text database of original PDFs of landslide risk assessment research.

[0082] To address the complex multi-column layouts, nested formulas, and mixed text and images found in academic literature, this embodiment employs multimodal conversion technology. In the image conversion stage, an image conversion algorithm decomposes the PDF document into a series of JPG images, storing each document in its corresponding independent folder. Based on this, a deep learning layout analysis algorithm is used to extract deep features and segment regions from these JPG images, precisely defining the physical areas for different functions such as text, titles, figures, formulas, and references, thus achieving a structured definition of complex layouts.

[0083] For the structured JPG images, unstructured content is categorized and deeply analyzed: for the main text, semantically coherent text content is extracted using streaming text reconstruction technology; for embedded mathematical formulas in the literature, the formula recognition module of Hunyuan-OCR accurately converts them into standardized mathematical expressions in LaTeX format; for table content, the table structure restoration algorithm identifies the row and column attributes and alignment relationships of cells, achieving structured parsing of the tables. Furthermore, to improve processing efficiency and ensure corpus purity, this process presets region filtering rules, automatically locating and skipping reference areas. Finally, the system logically arranges and aligns the parsed titles, abstracts, main text, LaTeX formulas, and structured tables using vLLM, automatically generating structured Markdown format files, thus constructing a standardized foundation for the landslide science literature corpus.

[0084] Based on the assessment task, the types of knowledge to be extracted are clearly defined, including landslide case knowledge, assessment model knowledge, and adaptation data knowledge. Landslide case knowledge encompasses a range of knowledge about the landslide cases themselves used in risk assessments within scientific and technological literature; assessment model knowledge encompasses a range of knowledge including the model algorithms used in landslide risk assessment analysis, their inputs and outputs, and related descriptions of performance evaluation; adaptation data knowledge encompasses a range of knowledge including the main data used in model operation and its semantics and constraints. Based on this, landslide knowledge extraction ontology, model knowledge extraction ontology, and data knowledge extraction ontology are designed.

[0085] The constructed landslide knowledge extraction ontology comprises three aspects: First, landslide metadata, primarily extracting landslide case names, landslide types, and landslide scales. Second, triggering factors, mainly including natural and human factors. For example, landslides triggered by rainfall can be extracted as natural triggers, while landslides triggered by construction can be extracted as human triggers. Third, spatiotemporal information, primarily extracting the time and location of landslide cases.

[0086] The constructed model knowledge extraction ontology comprises four aspects: First, the model meta-information class, primarily extracting the model name and type. Second, the model input class, indicating the variables the model relies on for calculation, divided into internal and external inputs. Internal inputs include the parameters and coefficients that the model needs to set, while external inputs extract the model's data requirements. Third, the model output class, primarily extracting model results and interpretive knowledge, such as extracting the risk conclusions calculated by the model as output knowledge, and factor importance and parameter contribution relationships as interpretive output knowledge. Fourth, the model performance class, representing the model's effectiveness, directly impacting subsequent model decisions and selections. This mainly extracts performance evaluation and usage conditions. Performance evaluation can extract model accuracy indicators, stability evaluations, and model comparison results; applicability conditions can extract the model's applicable area, applicable landslide types, and applicable data conditions.

[0087] The constructed data knowledge extraction ontology comprises three aspects: First, a data metadata class, primarily extracting data name, data type, and data source. Second, a data attribute class, extracting attribute knowledge such as resolution, time scale, update frequency, data range, and data format of the data used. Third, a data processing class, which mainly extracts knowledge related to the processed data that requires preprocessing or calculation, including its original data name, calculated data name, data processing method, and preprocessing method, such as slope data calculated from DEM data.

[0088] Therefore, an extraction ontology structure covering the needs of the landslide risk assessment task in this example was constructed. The knowledge extraction ontology design is as follows: Figure 2 As shown, the core comprises three parts: landslide case knowledge, assessment model knowledge, and adaptive data knowledge. Landslide case knowledge explains what a landslide is, why it occurs, where it occurs, and what its impact is. Landslide knowledge determines the risk mechanism and includes landslide metadata, triggering factors, and spatiotemporal information, covering landslide type, scale, natural causes, human causes, location, and time of occurrence. Assessment model knowledge describes the methods used to describe the landslide formation mechanism and perform risk calculations. Model knowledge enables mechanism calculations and depends on and is constrained by data knowledge. It includes model metadata, model input, model output, and model performance, covering model name, model type, internal input, external input, output results, model interpretation, performance evaluation, and usage conditions. Adaptive data knowledge clarifies what data is needed for model operation, and what the semantics and constraints of this data are. Data knowledge provides the computational carrier and includes data metadata, data attributes, and data processing, covering data name, data source, resolution, time, update, calculated data name, original data name, data method, and preprocessing.

[0089] The Qwen-max large-scale model is invoked to perform a deep extraction task on the previously constructed Markdown structured corpus. This task conducts multi-dimensional analysis under the strong constraints of the extraction prompt word instruction set. In the trigger word recognition stage, the semantic parsing technology of the Qwen-max large-scale model is used to locate key nodes in the documents describing landslide events, algorithm models, and data usage, thereby identifying knowledge-intensive areas. Subsequently, in the entity and attribute extraction stage, the model automatically identifies and extracts core entities such as landslide elements, evaluation models, and adaptation data, as well as their associated attributes, from the structured corpus. Finally, the inherent relationships between entities and between entities and attributes are identified through relation extraction logic, thereby achieving a deep transformation from structured documents to domain knowledge. The extracted fragmented knowledge is expressed as a standardized triple structure of "entity-relationship-attribute / entity," accurately capturing the core attributes and inherent logical relationships of the models and data used in landslide risk assessment analysis, realizing the transformation from massive text to structured knowledge.

[0090] The large model extracts a large number of structured triples related to landslide cases, model usage, and data usage. In this embodiment, a knowledge fusion procedure is first executed, effectively eliminating semantic ambiguity and data redundancy across document corpora through operators such as entity alignment, attribute merging, and conflict resolution. Subsequently, the processed structured triple knowledge is imported into the Neo4j graph database for storage. This process ultimately constructs a landslide domain knowledge graph with a multi-level relational network, which not only achieves a systematic representation of landslide cases, model usage, and data usage knowledge, but also provides a solid underlying data index for subsequent "model-data" computational ontology modeling for landslide assessment tasks through the relational characteristics of the graph structure.

[0091] In this embodiment, a landslide case subgraph query is performed based on the constructed landslide knowledge graph to extract relevant knowledge of landslide cases and conduct spatiotemporal feature mining. Spatially, the location of landslide cases is obtained through landslide case subgraph retrieval. Geographic coordinate mapping and GIS spatial analysis techniques are used to reveal the spatial distribution patterns and clustering characteristics of landslide disasters, and regional heat analysis is conducted for landslide risk assessment. Temporally, the occurrence time of landslide cases is obtained through landslide case subgraph retrieval, and the model name used is obtained through model usage subgraph retrieval. Time series evolution analysis is used to mine the periodicity and suddenness of landslide disasters and analyze the evolutionary patterns of landslide risk assessment models.

[0092] This embodiment defines a multi-dimensional model semantic classification framework, defining multi-dimensional model classification criteria from the dimensions of mechanism, function, and scale. At the mechanism level, models can be categorized into physics-driven models, empirical statistical models, machine learning models, deep learning models, and coupled models. At the function level, since landslide risk assessment involves multiple assessment aspects, it can be divided into susceptibility assessment, hazard assessment, vulnerability assessment, and risk assessment models. At the scale level, models are categorized into single landslide analysis models, small-area landslide analysis models, and large-area landslide analysis models.

[0093] Based on this, the model subgraphs are queried and relevant model knowledge is extracted from the constructed landslide knowledge graph. Combined with knowledge such as model instance names and applicability, a multi-dimensional semantic feature vector of model instances and model classification criteria is constructed. The spatial mapping similarity between model instances and multi-dimensional classification criteria is calculated using a deep semantic matching operator. Cross-dimensional automatic model discrimination and logical classification are achieved through similarity matching.

[0094] Subsequently, the input constraint system of model input knowledge (including required data items, input parameter range and format, etc.), the output semantic standard of model output knowledge (including clarifying the physical dimension and geographical meaning of the output results), and the multi-dimensional performance indicators of model performance knowledge (covering model applicability, model complexity, prediction accuracy and practicality evaluation indicators) are formally defined to achieve refined modeling of model computational attributes and capabilities.

[0095] Finally, after completing the automated model classification, a top-down hierarchical inheritance architecture is constructed from the mechanistic dimension of the classification system. This scheme sets a clear logical gradient. For example, "landslide risk assessment model" is defined as the top-level parent class, under which the subclass "machine learning model" differentiates, and finally at the end node, it maps to specific model instances such as "Support Vector Machine (SVM)". This structure ensures the downward inheritance of model attributes and the hierarchical refinement of functional logic. Based on the hierarchical structure, non-hierarchical horizontal relationships are introduced. By identifying the coupling relationships between models, nonlinear evolution or heterogeneous combination characteristics are characterized. For example, for the combined model Logistic Regression Weighted Support Vector Machine (LR-WSVM), a relationship can be established between Logistic Regression (LR) (an empirical statistical model) and Weighted Support Vector Machine (WSVM) (a machine learning model). By calculating the hierarchical design of "parent class-subclass-specific model" and the horizontal relationships of evolution or combination between models, the model hierarchy and relationship design are realized. Ultimately, a rigorous model knowledge classification system was constructed, providing a logical foundation for the subsequent automated selection of models.

[0096] In accordance with the requirements of landslide geological hazard risk assessment, this application defines a multi-dimensional data semantic classification framework, defining multi-dimensional data classification criteria from the dimensions of data attributes, evaluation factors, and spatiotemporal dimensions. In the dimension of data attributes, data is divided into topographic and geomorphological data, geological environment data, hydrological and meteorological data, socioeconomic data, land cover data, engineering activity data, and historical disaster data. In the dimension of evaluation factors, data is divided into disaster-inducing environment data, triggering factor data, and disaster-bearing body data. In the spatiotemporal dimension, data can be divided into time and spatial dimension data.

[0097] Subsequently, refined modeling of data semantic attributes and relationships was conducted. The physical attributes (such as data format and resolution) and indicator attributes (such as maximum precipitation and slope range) of data attribute classes were formally defined, and semantic relationships between data were established. For example, slope data is calculated from DEM data, so the resolution and range attributes of the slope data are consistent with those of the source DEM data. Ultimately, a rigorous data classification system was constructed, providing a logical foundation for subsequent automated data selection.

[0098] Then, a deep semantic relationship chain of "task-model-data" is constructed. By defining constraints such as data format, precision, and feature terms, precise matching of model computing power and data resources is achieved. Third, a "data-task" outcome evolution chain is constructed, defining how the result datasets produced by the task are fed back and enrich the evaluation system. This is achieved through semantic parsing of the result set, enabling feedback and enhancement of existing evaluation logic. Through the synergy of these three chains, a semantic closed loop is realized throughout the entire risk assessment process.

[0099] Based on the multi-dimensional model classification system, multi-dimensional data classification system, and the constructed "task-model-data" association logic, this embodiment defines a quintuple model consisting of class (C), relation (R), instance (I), constraint rule (A), and attribute value (V). Task-oriented model and data computation ontology modeling is conducted. Classes represent the model classification system, data classification system, and landslide case knowledge system. Relationships represent the relationships between demand-driven chains, resource support chains, and outcome evolution chains; the relationships between model computational attributes and capabilities; model hierarchical and horizontal relationships; and data semantic attributes and relationships between data. Instances represent specific landslide case instances, specific model instances, and specific data instances. Constraint rules represent model constraint rules and data constraint rules. Attribute values ​​represent the attribute values ​​of landslide case instances, model instances, and data instances. After formalizing the representation using ontology modeling tools, the constructed five-tuple model forms a standardized, machine-readable computational ontology for landslide risk assessment tasks. Finally, guided by the constructed computational ontology, landslide case knowledge, assessment model knowledge, and matching data knowledge are re-characterized. This unified organizational framework is used to systematically and standardizedly express the constructed landslide knowledge base.

[0100] After inputting a landslide risk assessment task, the first step is to perform task requirement analysis based on deep semantic awareness. A pre-trained deep language model is used to perform global dependency analysis and semantic segmentation on the input natural language task text. Combined with Named Entity Recognition (NER) technology, task keywords are dynamically captured using a pre-defined task ontology tag set. This achieves preliminary cleaning and extraction of discrete feature entities from a jumbled text description for the assessment task. For example, if we receive a task such as "A certain city is located in the southwestern mountainous area, and landslides occur frequently due to rainfall, requiring a landslide risk assessment," the extracted keywords would be "a certain city," "southwestern mountainous area," and "affected by rainfall."

[0101] Subsequently, a knowledge-enhanced multidimensional task feature space mapping was implemented. A scalable feature mapping architecture was used to standardize and normalize features through semantic alignment, dynamically mapping features from multiple dimensions based on the specific attributes of the evaluation task. For spatial features, a spatial scale generalization matrix was established to automatically determine the spatial scale range of the task; for example, mapping the spatial range of "a certain city" to "large regional scale," and the geographical location to "southwest mountainous area." For landslide mechanism features, semantic similarity calculation was used to align the inducing factors described in the task with the landslide type, thereby identifying the type of landslide disaster; for example, mapping "affected by rainfall" to "rainfall-induced landslide." For model and data requirement description features, the model applicability and data applicability features of the task objective were automatically associated.

[0102] Finally, the analytical results are transformed into a set of standardized, machine-readable task constraint parameter vectors, enabling the automatic generation of structured task constraint tensors. For example, in this case, three sets of task constraint tensors are obtained: "City → Large-scale region", "Affected by rainfall → Rainfall-induced landslide", and "City → Southwest mountainous area". Through the above operations, the fuzzy landslide risk assessment task requirements are transformed into landslide assessment tasks with clear requirements, serving as the decision entry point for subsequent model selection.

[0103] Upon receiving the parsed structured task constraint tensor, the model knowledge management agent initiates the intelligent model constraint parsing operator to perform deep parsing of task constraints and initial model selection. By invoking the model knowledge subgraph in the landslide knowledge graph, the agent uses the explicit task requirements in the task vector as hard constraints to initially screen a set of candidate models that are applicable in terms of model type and performance. For example, in this case, logical filtering is performed in the model classification system using "large-area scale," and semantic matching is performed on the model's performance evaluation and applicable conditions knowledge using "rainfall-type landslide," thus selecting a set of preliminary candidate models suitable for both large-area scale and rainfall-type landslides.

[0104] Next, for complex decision-making scenarios where multiple highly similar candidate models exist in the initial candidate model set, the model knowledge management agent performs case-based reasoning-based similarity analysis. Multi-dimensional vector comparisons are performed between the multi-dimensional features of the landslide risk assessment task and the relevant knowledge of historical landslide cases stored in the landslide knowledge base to calculate the similarity between the landslide assessment task requirements and the landslide risk assessment cases. For example, in this case, the task feature "southwest mountainous area" is compared with the landslide occurrence location in the landslide case knowledge. The performance evaluation and applicability conditions of the models used in historical cases are used as reference gains to prioritize and weight the current candidate models, effectively resolving the semantic ambiguity and logical conflicts in model selection.

[0105] Finally, the model knowledge management agent integrates core decision-making logic, including model type, applicable conditions, performance evaluation, spatial scale of risk assessment, landslide type, and similarity to landslide risk assessment cases. Utilizing knowledge reasoning techniques and a constructed model selection operator, it conducts a comprehensive utility evaluation of candidate models, thereby achieving knowledge reasoning and final model selection based on integrated decision-making logic. In this example, the integrated model type is a large-scale model, the applicable landslide type is rainfall-induced landslides, the landslide case with the highest similarity to the southwestern mountainous area is selected for risk assessment, and the model has the best performance evaluation. By performing path search and constraint consistency detection on the knowledge graph, it dynamically identifies and locks the evaluation model with the highest fit to the current complex task environment, providing precise algorithmic input for subsequent data intelligent adaptation.

[0106] After the model knowledge management agent determines the final evaluation model, it accesses the data knowledge subgraph in the landslide knowledge graph to perform semantic parsing of data requirements based on the model input. The agent deeply analyzes and constructs the input constraint tensor of the final evaluation model, and uses natural language processing and semantic mapping techniques to transform the data input requirements of the final evaluation model into standardized data discovery retrieval expressions, providing a semantic basis for the accurate selection of heterogeneous data. Next, the agent performs a multi-source heterogeneous data adaptability evaluation based on knowledge reasoning. It invokes built-in data selection operators to perform large-scale searches in the massive heterogeneous data knowledge subgraph. By executing knowledge reasoning and adaptability evaluation algorithms, candidate datasets are quantitatively scored from dimensions such as coverage, spatiotemporal consistency, and data integrity. The agent automatically determines the logical fit between each data source and the model input items, ultimately selecting the optimal adapted dataset that meets the operating conditions of the final evaluation model.

[0107] The landslide risk assessment agent, as the core of the system, is responsible for coordinating the overall adaptation task. Its design primarily focuses on the agent's scheduling, adaptation, and negotiation mechanisms. First, a global task scheduling mechanism centered on the landslide risk assessment agent is constructed. In this mechanism, the landslide risk assessment agent receives landslide knowledge and structured task constraint tensors after computational ontology modeling, dynamically activating the model knowledge management agent and the data knowledge management agent. Next, a knowledge-driven model and data integration adaptation mechanism is built. In this mechanism, the landslide risk assessment agent receives in real-time the final assessment model determined by the model knowledge management agent and the best-fit dataset selected by the data knowledge management agent, integrating the relevant knowledge of the final assessment model and the best-fit dataset to form a complete "model-data" adaptation scheme for the input landslide assessment task. Finally, a conflict detection mechanism based on multi-round collaborative negotiation is established. For potential anomalies caused by data precision mismatch or differences in physical dimensions in the "model-data" adaptation scheme, the landslide risk assessment agent is responsible for performing end-to-end adaptation consistency checks. Regarding the negotiation mechanism, the landslide risk assessment agent constructs a closed-loop conflict judgment logic: when it detects that the operating conditions of the candidate model are incompatible with the features of the currently matched dataset, it automatically triggers a collaborative negotiation process among multiple agents. This agent guides the model knowledge management agent to adaptively adjust parameters, or instructs the data knowledge management agent to perform dynamic reconstruction and incremental completion of the adapted dataset, achieving dynamic correction of the adaptation scheme through multiple rounds of iterative negotiation.

[0108] The landslide risk assessment agent, model knowledge management agent, and data knowledge management agent constructed using the above steps formalize the entire adaptation computation process into four layers: input layer, scheduling layer, knowledge matching layer, and adaptation computation layer. The input layer receives the parsed landslide risk assessment task, i.e., the structured task constraint tensor, and also inputs the standardized landslide knowledge base after computational ontology modeling and reconstruction. The scheduling layer, through the landslide risk assessment agent, schedules the model knowledge management agent and the data knowledge management agent. The knowledge matching layer determines the final assessment model through the model knowledge management agent and the optimal adaptation dataset through the data knowledge management agent. The adaptation computation layer feeds back the final assessment model and the optimal adaptation dataset to the landslide risk assessment agent, performing the final "model-data" intelligent adaptation and negotiation to calculate the final adaptation scheme. Through this hierarchical automated chain, knowledge graph-driven intelligent adaptation of the landslide risk assessment model and data is achieved, significantly improving the scientific rigor and timeliness of disaster assessment.

[0109] The following is an exemplary description of the knowledge graph-driven landslide risk assessment device provided in this application.

[0110] like Figure 3As shown, this application embodiment provides a knowledge graph-driven landslide risk assessment device. The knowledge graph-driven landslide risk assessment device 300 includes: The acquisition module 301 is used to acquire multiple landslide case texts and extract knowledge from all landslide case texts to obtain a landslide knowledge graph. The landslide knowledge graph includes multiple landslide cases, as well as the evaluation model and landslide data corresponding to each landslide case. The generation module 302 is used to map the landslide task to be evaluated to the landslide knowledge graph and generate the task vector of the landslide task to be evaluated. The model selection module 303 is used to select models from the landslide knowledge graph based on the task vector to obtain the final evaluation model for the landslide task to be evaluated. The data filtering module 304 is used to filter data from the landslide knowledge graph based on the final evaluation model to obtain the final landslide data corresponding to the landslide task to be evaluated. The assessment module 305 is used to conduct a risk assessment of the landslide task to be assessed based on the final landslide data and the final assessment model, and obtain the landslide risk assessment result.

[0111] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0113] like Figure 4 As shown, an embodiment of this application provides a terminal device, wherein the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 4The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0114] Specifically, when the processor D100 executes the computer program D102, it accumulates knowledge of landslide cases by constructing a landslide knowledge graph, forming a structured knowledge system that provides data support for landslide risk assessment. It also performs model and data screening from the landslide knowledge graph, enabling autonomous model and data scheduling, improving the relevance of models and data to the landslide assessment task, and effectively improving the reliability and efficiency of landslide risk assessment.

[0115] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0116] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0117] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0118] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0119] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a knowledge graph-driven landslide risk assessment method device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, such as a USB flash drive, a portable hard drive, a magnetic disk, or an optical disk.

[0120] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0121] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0122] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A knowledge graph driven landslide risk assessment method, characterized in that, include: Multiple landslide case texts were obtained, and knowledge was extracted from all landslide case texts to obtain a landslide knowledge graph; The landslide knowledge graph includes multiple landslide cases, as well as the evaluation model and landslide data corresponding to each landslide case; The landslide task to be evaluated is mapped to the landslide knowledge graph to generate a task vector for the landslide task to be evaluated. Based on the task vector, models are selected from the landslide knowledge graph to obtain the final evaluation model for the landslide task to be evaluated; Based on the final evaluation model, data is filtered from the landslide knowledge graph to obtain the final landslide data corresponding to the landslide task to be evaluated. Based on the final landslide data, the final assessment model is used to conduct a risk assessment on the landslide task to be assessed, and the landslide risk assessment result is obtained. The step of filtering data from the landslide knowledge graph based on the final evaluation model to obtain the final landslide data corresponding to the landslide task to be evaluated includes: The landslide cases corresponding to the final evaluation model are analyzed to obtain the data input constraint tensor of the final evaluation model; The data input constraint tensor is transformed into a data requirement feature vector; Generate a semantic feature vector for each landslide data point in the landslide knowledge graph; Calculate the semantic similarity between each data requirement feature vector and each semantic feature vector; The landslide data with the highest semantic similarity is taken as the final landslide data for the landslide task to be evaluated.

2. The landslide risk assessment method according to claim 1, characterized in that, The step of mapping the landslide task to be evaluated onto the landslide knowledge graph to generate a task vector for the landslide task to be evaluated includes: Text cleaning and extraction were performed on the landslide task to be evaluated to obtain multiple discrete feature words; All discrete feature words are mapped to the landslide knowledge graph. Based on the mapped landslide knowledge graph, the landslide task to be evaluated is analyzed from multiple feature dimensions to generate the task vector of the landslide task to be evaluated.

3. The landslide risk assessment method according to claim 1, characterized in that, The step of selecting models from the landslide knowledge graph based on the task vector to obtain the final evaluation model for the landslide task to be evaluated includes: Generate the feature vector corresponding to each landslide case in the landslide knowledge graph; Calculate the similarity between the task vector and each of the feature vectors; All similarities are sorted from largest to smallest, and the landslide cases corresponding to the top few similarities in the sorting results are all selected as candidate landslide cases. A multi-dimensional model evaluation is performed on all candidate landslide models to determine the final evaluation model for the landslide task to be evaluated.

4. The landslide risk assessment method according to claim 3, characterized in that, The calculation of the similarity between the task vector and each of the feature vectors includes: Through the formula: Calculate the task vector for the landslide assessment task. With the Feature vector of a landslide case similarity between ; in, Indicates the number of feature dimensions. Indicates the first Weights of each feature dimension, Represents the task vector The first in One characteristic, Represents the feature vector The first in One characteristic, , This indicates the number of landslide cases in the landslide knowledge graph.

5. The landslide risk assessment method according to claim 1, characterized in that, The calculation of the semantic similarity between each data requirement feature vector and each semantic feature vector includes: Through the formula: Calculate the feature vector of data requirements With the Semantic feature vector of landslide data semantic similarity between ; in, This represents the length of the demand feature vector. This indicates the number of the last feature in the demand feature vector. Represents the feature vector of data requirements The first in One characteristic, Represents semantic feature vector The first in One characteristic, , This indicates the number of landslide data points in the landslide knowledge graph.

6. The landslide risk assessment method according to claim 1, characterized in that, The step of conducting a risk assessment of the landslide task to be assessed based on the final landslide data and using the final assessment model to obtain the landslide risk assessment result includes: The final landslide data is adapted and corrected based on the final evaluation model. The corrected final landslide data is used as input data for the final evaluation model. The final evaluation model is then used to conduct a risk assessment on the landslide task to be evaluated, and the landslide risk assessment result is obtained.

7. A knowledge graph-driven landslide risk assessment device, characterized in that, include: The acquisition module is used to acquire multiple landslide case texts and extract knowledge from all landslide case texts to obtain a landslide knowledge graph. The landslide knowledge graph includes multiple landslide cases, as well as the evaluation model and landslide data corresponding to each landslide case; The generation module is used to map the landslide task to be evaluated to the landslide knowledge graph and generate the task vector of the landslide task to be evaluated. The model filtering module is used to filter models from the landslide knowledge graph based on the task vector to obtain the final evaluation model for the landslide task to be evaluated. The data filtering module is used to filter data from the landslide knowledge graph based on the final evaluation model to obtain the final landslide data corresponding to the landslide task to be evaluated. The assessment module is used to conduct a risk assessment on the landslide task to be assessed based on the final landslide data and the final assessment model, and obtain the landslide risk assessment result. Specifically, the data filtering module is used to implement: The landslide cases corresponding to the final evaluation model are analyzed to obtain the data input constraint tensor of the final evaluation model; The data input constraint tensor is transformed into a data requirement feature vector; Generate a semantic feature vector for each landslide data point in the landslide knowledge graph; Calculate the semantic similarity between each data requirement feature vector and each semantic feature vector; The landslide data with the highest semantic similarity is taken as the final landslide data for the landslide task to be evaluated.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the knowledge graph-driven landslide risk assessment method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the knowledge graph-driven landslide risk assessment method as described in any one of claims 1 to 6.

Images