Landslide control auxiliary decision-making method fusing knowledge graph and case-based reasoning

Abstract

The invention provides a knowledge graph and case-based reasoning integrated landslide control aid decision-making method, which is based on a knowledge graph and case-based reasoning integrated landslide control aid decision-making model KGCBR, and is characterized in that the decision-making model comprehensively considers geographic spatial characteristics of landslide attributes and adaptive distribution of weights and introduces a case correction strategy; the construction process of the decision model comprises the steps of landslide treatment knowledge graph construction, case-based reasoning solving process determination and adaptive weight optimization distribution based on a grey wolf optimization algorithm. According to the method, treatment scheme matching and recommendation under the guidance of the knowledge graph can be realized, and a refined engineering treatment decision basis is provided for landslide disasters with homogeneous characteristics.

Description

Technical Field

[0001] The present invention belongs to the field of information technology, and in particular to a landslide management decision-making assistance method integrating knowledge graph and case reasoning. Background Art

[0002] Among all geological disasters, landslides occur most frequently and are the most common disaster in daily life. According to statistics, there were 5,659 geological disasters in my country in 2022, of which 3,919 were landslides, accounting for more than 60% of the total number of geological disasters in my country. Therefore, landslides have become the focus of geological disaster prevention and control.

[0003] Faced with the severe landslide disaster prevention and mitigation situation, it is crucial to improve the comprehensive landslide disaster management capabilities, determine practical engineering management plans, and reduce landslide disaster risks and hidden dangers.

[0004] At present, the selection of landslide control measures is mainly achieved through empirical judgment, numerical simulation, hierarchical analysis, fuzzy comprehensive evaluation and case reasoning.

[0005] The empirical judgment method is usually used by professionals to select appropriate control measures through qualitative analysis based on on-site conditions. It is a commonly used method in actual work, but it requires extremely high knowledge reserves and practical experience of professional designers, and a large amount of work requires manual participation, which increases the time cost and uncertainty of decision-making.

[0006] The numerical simulation method uses a mechanical mechanism model to analyze and calculate the stability of the landslide, and determines the corresponding treatment plan based on the actual project. It is relatively scientific and accurate, but it relies too much on high-precision monitoring data and has a high cost for parameter acquisition. The hierarchical analysis method calculates the comprehensive score of the schemes by building an evaluation index system in layers to rank and optimize the schemes. This method is systematic, scientific, concise and practical, but it often needs to consider multiple indicator factors when selecting measures, which increases the data statistics and makes it difficult to determine the indicator weights, resulting in problems such as low decision-making efficiency and unreasonable scheme selection.

[0007] The fuzzy comprehensive evaluation method establishes fuzzy logical relationships between multiple factors and then judges the pros and cons of decision-making plans based on a comprehensive evaluation based on a fuzzy mathematical model. This method can effectively handle fuzzy and uncertain decision information to ensure the reliability of the results. However, when there are too many decision-making indicators, the calculation complexity will be greatly increased, and the subjectivity is too strong when determining the weights. Therefore, it is difficult to be promoted in the application of governance engineering design.

[0008] Case Based Reasoning (CBR) uses the design experience of similar projects in the past to solve current engineering problems. This method is in line with human cognitive activities and psychology, and has good interpretability and scalability. In the long-term landslide control process, a large number of successful landslide control cases have been accumulated, which provides a basis for CBR to achieve optimal control schemes, but there are still some key issues to be studied. First, the landslide attributes show a spatial distribution law in the region, which conforms to the corresponding geographic spatial characteristics. There is a certain geographical similarity between the attributes of different landslides, but the existing methods ignore this point, which makes the model have obvious limitations and low robustness in complex and heterogeneous environments; secondly, it is difficult to adaptively generate the optimal weight value of the landslide attribute, which hinders the reasonable allocation of weights, which will reduce the accuracy of the decision-making results. Finally, the existing CBR method directly takes the solution of the most similar case as the recommendation result, but ignores the reasoning contingency caused by the interference of false similar cases with case reuse.

[0009] Knowledge graph is one of the core technologies of artificial intelligence. As an emerging knowledge representation method, knowledge graph contains a large amount of prior knowledge, which associates and deeply integrates multi-source heterogeneous data in the form of structured entity and relationship triples. Among them, the knowledge representation learning method can project the entities and relationships in the knowledge base into a low-dimensional vector space and perform numerical representation, so as to efficiently calculate the complex and implicit semantic associations between entities and relationships and realize the direct calculation of symbolic representation vectorization. For example, Jiao et al. obtained the vector embedding of each entity in the earthquake disaster knowledge graph through triple representation learning, which can better calculate the similarity between the building status and material and each target decision, and infer the best rescue measures for specific disaster scenarios; Chen et al.

[26] converted the landslide prediction task into a link prediction task and realized landslide sensitivity prediction through the knowledge embedding model, which proved the potential of knowledge graph embedding in disaster management. Therefore, in order to better characterize the geospatial characteristics of landslide attributes, we can construct a landslide management knowledge graph and combine it with knowledge representation learning methods to convert landslide attributes into spatially continuous, low-dimensional feature vectors, so as to explore the implicit correlation between landslide attributes and management solutions, thereby better calculating the similarity between landslide attributes.

[0010] In the field of numerical optimization, swarm intelligence optimization algorithms simulate various group behaviors of social animals and use the information interaction and cooperation between individuals in the group to achieve the purpose of optimization. These algorithms include classic algorithms such as Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Grey Wolf Optimizer (GWO).

[0011] In practical applications, the swarm intelligence optimization algorithm converts engineering optimization problems into function optimization problems, establishes an objective function, and then continuously iterates the algorithm to find the optimal solution to the objective function. It has been widely used in decision-making attribute weight optimization tasks in related fields such as water diversion projects, bridge modeling, and construction period estimation. Therefore, the swarm intelligence algorithm is introduced to optimize the weight of landslide attributes, so as to obtain the optimal weight value of the attribute, thereby further improving the problem-solving quality of case-based reasoning.

[0012] Summarizing the above research, the main shortcomings of the existing technology are:

[0013] (1) Existing studies have ignored the geographical similarities in the spatial distribution of landslide attributes, which results in obvious limitations and low robustness of the model in complex and heterogeneous environments;

[0014] (2) It is difficult to adaptively generate the optimal weight values ​​of landslide attributes, which hinders the reasonable allocation of weights and reduces the accuracy of decision-making results;

[0015] (3) Existing CBR methods directly use the solution of the most similar case as the recommendation result, ignoring the randomness of reasoning caused by false similar cases interfering with case reuse. Summary of the invention

[0016] The present invention proposes a landslide control decision-making assistance method that integrates knowledge graph and case reasoning, which can realize the matching and recommendation of control plans under the guidance of knowledge graph, and provide a refined engineering control decision-making basis for landslide disasters with homogeneous characteristics.

[0017] The present invention adopts the following technical solutions.

[0018] A landslide control auxiliary decision-making method integrating knowledge graph and case-based reasoning. The decision-making method is based on the landslide control auxiliary decision-making model KGCBR (Integrating Knowledge Graph and Case Based Reasoning, KGCBR) that integrates knowledge graph and case-based reasoning. The decision-making model comprehensively considers the geographic spatial characteristics of landslide attributes and the adaptive allocation of weights and introduces a case correction strategy to ensure the reliability and rationality of recommended control measures. The construction process of the decision-making model includes the construction of a landslide control knowledge graph, the determination of a case-based reasoning solution process, and the adaptive weight optimization allocation based on a grey wolf optimization algorithm.

[0019] The construction of the landslide management knowledge graph includes the following contents: First, construct a landslide management knowledge graph, use the TransH knowledge embedding model to capture the geographical similarities of landslide attributes, and integrate them into the case representation. Specifically, build a landslide management knowledge graph based on existing scientific and technological literature, laws and regulations, implementation standards, disaster statistics and existing ontology models. The type information of the landslide management knowledge graph includes the basic attributes of landslide disasters, spatial attributes including longitude and latitude, and landslide management measures.

[0020] When constructing the knowledge graph of landslide management, the disaster statistical data is based on the provincial geological disaster comprehensive prevention and control database as the data source, and the data subject of the data source is structured data; first, based on the data table header name mapping method of the geological disaster comprehensive prevention and control database, the structured data is imported into the Neo4J graph database, and converted into a "head entity-relationship-tail entity" triple structure through the built-in CQL language to complete the construction of the landslide management knowledge graph.

[0021] The case-based reasoning solution process includes case representation, case retrieval and reuse, and case correction; that is, using GWO to achieve adaptive weight optimization allocation to obtain the optimal weight of landslide attributes, and combining local and global similarity reasoning to apply to case retrieval; in order to avoid the randomness of the reasoning results, by adjusting the k value and counting the recommended measures of similar historical cases, the solution of the target case is corrected to achieve matching and recommendation of governance plans under the guidance of knowledge graphs, providing a refined engineering governance decision-making basis for landslide disasters with homogeneous characteristics.

[0022] In the case representation process, the geographical similarity of landslide attributes is taken into account on the basis of the basic landslide management attributes of landslide slope, sliding body length, sliding body width, average sliding body thickness, sliding body volume, sliding body morphology, sliding body lithology, height difference, slope, slope direction, threat object, total number of threatened people, and threatened property, so as to optimize the description of geographical spatial characteristics in landslide management case reasoning;

[0023] When the entity vectors obtained based on vector calculation can reflect the similarity between entities, the knowledge graph embedding (KGE) method is used to perform knowledge representation learning on the entities and relationships in the knowledge graph, and the entity vector representation of text attributes is obtained through vector calculation; for the 1:N, N:1, and N:N entity relationships in the landslide management knowledge graph, the TransH knowledge embedding model is used to perform representation learning on the knowledge triples, and the landslide text attributes are converted into distributed feature vectors. The model can show good performance when processing the above complex relationships. The specific process is shown in the attached figure. Figure 3 ; The TransH model defines a hyperplane W for each relation r in the triple (h, r, t) r , and h⊥ With t ⊥ As the projection of h and t on the hyperplane and satisfying Formula 1, its model scoring function is positioned as Formula 2;

[0024]

[0025]

[0026] In the formula, h represents the head entity, t represents the tail entity, r represents the relationship, and w r is the hyperplane W r The unit normal vector, h ⊥ represents the projection of the head entity on the hyperplane, t ⊥ Represents the projection of the tail entity on the hyperplane, L1 and L2 are the first and second normal forms respectively;

[0027] After learning the knowledge representation of landslide management attributes, the structural form of case representation is determined, that is, the 13 basic attributes of landslide management are combined to form the case representation of landslide management, and its structure is expressed in the formula:

[0028] C=(e1,e2,e3,...e n ) Formula 3;

[0029] In the formula, e n are the basic properties of landslide control, n=1, 2, 3, …, 13.

[0030] In the process of case retrieval and reuse, the K-nearest neighbor (KNN) algorithm is used to select K similar cases from the landslide management case library, namely: first, the local similarity between each attribute of the landslide management target case and the historical case is calculated, and then the global similarity is calculated by combining the weight of each attribute, and then sorted from high to low;

[0031] When calculating local similarity, the landslide management attributes are divided into numerical attributes and text attributes. For numerical attributes, the similarity of the same numerical attribute between the target case and the historical case is expressed by the distance after the two values ​​are normalized, as shown in the formula:

[0032]

[0033] in, represents the value of attribute k of the i-th historical case in the case library, represents the value of attribute k of target case j, and Represent the maximum and minimum values ​​of attribute k respectively, express and For text attributes, the distributed low-dimensional vector of the text attributes obtained after knowledge representation learning is used as the numerical representation of the text attributes, and the cosine similarity is used to calculate the similarity between text attributes, as shown in the formula:

[0034]

[0035] In the formula, and The vectors representing the attribute n of the target case j and the attribute n of the i-th historical case in the case library, Represents the cosine similarity between attributes n;

[0036] When calculating the global similarity between the target case and the historical case, the KNN algorithm is used to obtain k historical cases with higher global similarity as the result of case retrieval; the formula is

[0037]

[0038] In the formula, w m represents the weight of the mth attribute, represents the similarity between the i-th historical case in the case library and the target case attribute m, M represents the number of attributes, represents the global similarity between the i-th historical case in the case library and the target case;

[0039] In the case correction process, k similar cases are obtained from the historical case database, i.e., the geological disaster comprehensive prevention and control database, through case retrieval. Assuming that the historical case with the highest similarity is an erroneous case, the solution of the target case is affected by the biased historical case, and the case is corrected to obtain the solution of the target case. Set the size of the K nearest neighbor value, select the top-k historical cases in descending order of similarity as the screening result, 1≤k≤10, and count the array of engineering control measures. If the total number of a certain control measure is greater than or equal to k / 2, the result is rounded down, it is considered to be a commonly used control measure, and it is added to the recommendation list; in the case correction process, the control measures are divided into 12 types of landslide control measures: interception and drainage, retaining wall, slope protection and greening, anchor rods or anchor cables, slope reduction and load reduction, anchoring, face wall, water interception skeleton, anti-slide pile, protective net, retaining wall, and interception; for the landslide disaster points that have been controlled, 0 and 1 are used to indicate whether the control measures are adopted. For the top-k landslide cases before the calculated similarity ranking, the number of each control measure is counted, and the calculation formula is as follows:

[0040]

[0041] Among them, m represents the similarity ranking, n represents the nth governance measure, S mnIndicates whether the historical case with similarity ranking m adopts the nth governance measure, ω represents the recommendation threshold, which is k / 2 here; the final statistical calculation result sum is rounded down, and each governance measure can only be recommended to the target case as an adoptable governance measure when the statistical calculation result is greater than or equal to 0.

[0042] In the method, the solving performance of the CBR decision-making auxiliary model is improved by optimizing the weight setting, that is, the gray wolf optimization algorithm GWO is used to realize adaptive weight optimization allocation, GWO simulates the hunting behavior of the gray wolf population, and realizes the optimization purpose through the mutual cooperation mechanism within the population. The dynamic optimization process of the wolf position in GWO is used to realize the refined weight allocation of the CBR model; the optimization core of GWO is to obtain the maximum or minimum value of the fitness function; the mean absolute error MAE (Mean Absolute Error, MAE) between the recommended governance measures and the actual governance measures of the KGCBR model is used as the fitness function;

[0043] The steps to implement the algorithm for adaptive weight optimization allocation are as follows:

[0044] Step a, set GWO operation parameters. Determine the maximum number of iterations, cutoff fitness value, solution space range, optimization parameter dimension, number of gray wolves and other related parameters of GWO;

[0045] Step b, initialize the GWO population. Input the weight value of each landslide attribute in the CBR case as an optimization parameter into the GWO;

[0046] Step c, CBR solution process execution and optimization. Assign the GWO initial parameter value as the historical optimal parameter value to the CBR model and execute the case-based reasoning solution process to calculate the fitness function value for each iteration;

[0047] Step d, determine the optimal weight value of the KGCBR model. Adjust the GWO parameters according to the calculated fitness value, iteratively update the optimal solution of the population, and when the number of GWO iterations reaches the maximum number of iterations or the calculated fitness function value reaches the cutoff fitness value, the minimum fitness value of GWO is obtained, which is the current optimal solution of GWO, and then determine the optimal weight value of each landslide attribute.

[0048] In the gray wolf optimization algorithm GWO, the gray wolf population has a strict population system, as shown in the attached Figure 4 As shown in the figure, α wolf is the leader of the wolf pack, representing the optimal solution of the model; β wolf assists α wolf, representing the suboptimal solution of the model; δ wolf obeys the orders of α wolf and β wolf to hunt, representing the general solution of the model; ω wolf follows other high-level wolves to maintain the balance of population relationships, representing the candidate solution of the model

[0049] The hunting behavior of the gray wolf population surrounding prey is defined by the formula

[0050] D=|C·X p (t)-X(t)| Formula 8;

[0051] X(t+1)=X p (t)-A·D i Formula 9;

[0052] A=2a·r1-a Formula 10;

[0053] C = 2·r2 Formula 11;

[0054] Among them, t represents the number of iterations of the current population, D represents the distance between the individual and the prey, and X p (t) and X(t) represent the positions of the prey and the individual, respectively. A and C represent the effective vectors of the synergy coefficients. a represents the convergence factor that decreases from 2 to 0. r1 and r2 are sets of random numbers in [0,1] to avoid falling into the local optimum.

[0055] The prey species of the gray wolf population hunting behavior is defined by the formula

[0056]

[0057] Among them, D α , D β , D δ represents the distance between α, β, δ wolves and other individuals, X α , X β , X γ represents the current position of α, β, γ wolves, X(t) represents the current position of the gray wolf, C1, C2, C3 represent the adjustment coefficient vector; the following formula 13 defines the step size and direction of the ω individual in the wolf pack moving towards α, β, δ, and the following formula 14 defines the final position of the ω wolf;

[0058]

[0059] The algorithm optimization of GWO starts with randomly creating a population of gray wolves (candidate solutions). Specifically, during the iteration process, α, β, and δ wolves estimate the possible potential location of the prey, that is, the optimal solution; the gray wolves update their positions based on the distance between them and the prey. For development and exploration during the search process, the parameter a should be decreased from 2 to 0. If |A|>1, the candidate solution is far away from the prey; if |A|<1, the candidate solution is close to the prey. The flowchart of GWO is as follows Figure 5 shown.

[0060] In the fitness function of the gray wolf optimization algorithm GWO, MAE is used as an evaluation index to measure the deviation between the recommended measures and the actual measures. The smaller the MAE value, the higher the recommendation quality. Let the recommended control measures set be {p1, p2, ..., p3}, and the actual control measures set be {q1, q2, ..., q3}; where the set indicates whether each type of landslide control measure is adopted. If adopted, the score is 1, and if not adopted, the score is 0. Then MAE represents the error rate between the recommended control measures and the actual control measures. The MAE evaluation formula is expressed as:

[0061]

[0062] Among them, n represents the type of landslide control measures, which has a value of 12;

[0063] The GWO fitness function f is expressed as:

[0064]

[0065] Where m is the number of target cases in the validation set, MAE l represents the mean absolute error between the actual governance measures of the lth target case and the recommended governance measures of CBR;

[0066] In the decision-making method, the smaller the fitness function value is, the more reasonable the allocation weight is, which is more conducive to improving the accuracy of CBR solution.

[0067] This paper proposes a landslide management decision-making auxiliary model (Integrating Knowledge Graph and Case Based Reasoning, KGCBR) that integrates knowledge graph and case reasoning. The model comprehensively considers the geographic spatial characteristics of landslide attributes and the adaptive allocation of weights and introduces a case correction strategy to ensure the reliability and rationality of the recommended management measures. First, a landslide management knowledge graph is constructed, and the TransH knowledge embedding model is used to capture the geographical similarity of landslide attributes and integrate it into the case representation. On this basis, GWO is used to achieve adaptive weight optimization allocation to obtain the optimal weight of landslide attributes, and local and global similarity reasoning are combined and applied to case retrieval. In addition, in order to avoid the contingency of the reasoning result, the solution of the target case is corrected by adjusting the k value and counting the recommended measures of similar historical cases, and finally the management plan matching and recommendation under the guidance of the knowledge graph is realized, providing a refined engineering management decision-making basis for landslide disasters with homogeneous characteristics.

[0068] The advantages of the present invention are:

[0069] (1) Construct a knowledge graph for landslide management, use the TransH knowledge embedding model to obtain the geographical similarity of landslide attributes and integrate it into the case representation, so as to enhance the spatial representation ability of the case representation; that is, introduce the TransH knowledge representation learning method, transform the landslide attributes into spatially continuous, low-dimensional feature vectors and integrate them into the case representation, so as to solve the problem of insufficient representation ability of the geographic spatial characteristics of the existing case representation, so as to better calculate the geographical similarity between landslide attributes.

[0070] (2) Based on GWO, an adaptive weight optimization allocation method is designed to obtain the optimal weight of landslide attributes, improve the quality of case retrieval and further improve the accuracy of CBR recommendation results; that is, the adaptive weight optimization allocation of landslide attributes is realized through the GWO algorithm, and the optimal weight of landslide attributes is obtained on the basis of ensuring the rationality of weight allocation.

[0071] (3) By adjusting the k value and counting the recommended measures of similar historical cases, the solution of the target case is corrected, effectively improving the generalization and applicability of the recommendation results; that is, using the case correction strategy to adjust the case reuse results, reduce the interference of false similar cases on the recommendation results, and effectively improve the applicability of the recommended measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0073] Attached Figure 1 It is a flow chart of the landslide control auxiliary decision-making model;

[0074] Attached Figure 2 This is a schematic diagram of an example of a knowledge graph for landslide management;

[0075] Attached Figure 3 It is a schematic diagram of the knowledge representation learning process;

[0076] Attached Figure 4 This is a schematic diagram of the gray wolf population hierarchy;

[0077] Attached Figure 5 It is a schematic diagram of the GWO process;

[0078] Attached Figure 6 This is a schematic diagram of comparative analysis of different recommendation methods. DETAILED DESCRIPTION

[0079] As shown in the figure, a landslide control decision-making assistance method that integrates knowledge graph and case-based reasoning is provided. The decision-making method is based on the landslide control decision-making assistance model KGCBR (Integrating Knowledge Graph and Case Based Reasoning, KGCBR) that integrates knowledge graph and case-based reasoning. The decision-making model comprehensively considers the geographic spatial characteristics of landslide attributes and the adaptive allocation of weights and introduces a case correction strategy to ensure the reliability and rationality of the recommended control measures. The construction process of the decision-making model includes the construction of the landslide control knowledge graph, the determination of the case reasoning solution process, and the adaptive weight optimization allocation based on the grey wolf optimization algorithm.

[0080] The construction of the landslide management knowledge graph includes the following contents: First, construct a landslide management knowledge graph, use the TransH knowledge embedding model to capture the geographical similarities of landslide attributes, and integrate them into the case representation. Specifically, build a landslide management knowledge graph based on existing scientific and technological literature, laws and regulations, implementation standards, disaster statistics and existing ontology models. The type information of the landslide management knowledge graph includes the basic attributes of landslide disasters, spatial attributes including longitude and latitude, and landslide management measures.

[0081] When constructing the knowledge graph of landslide management, the disaster statistical data is based on the provincial geological disaster comprehensive prevention and control database as the data source, and the data subject of the data source is structured data; first, based on the data table header name mapping method of the geological disaster comprehensive prevention and control database, the structured data is imported into the Neo4J graph database, and converted into a "head entity-relationship-tail entity" triple structure through the built-in CQL language to complete the construction of the landslide management knowledge graph.

[0082] The case-based reasoning solution process includes case representation, case retrieval and reuse, and case correction; that is, using GWO to achieve adaptive weight optimization allocation to obtain the optimal weight of landslide attributes, and combining local and global similarity reasoning to apply to case retrieval; in order to avoid the randomness of the reasoning results, by adjusting the k value and counting the recommended measures of similar historical cases, the solution of the target case is corrected to achieve matching and recommendation of governance plans under the guidance of knowledge graphs, providing a refined engineering governance decision-making basis for landslide disasters with homogeneous characteristics.

[0083] In the case representation process, the geographical similarity of landslide attributes is taken into account on the basis of the basic landslide management attributes of landslide slope, sliding body length, sliding body width, average sliding body thickness, sliding body volume, sliding body morphology, sliding body lithology, height difference, slope, slope direction, threat object, total number of threatened people, and threatened property, so as to optimize the description of geographical spatial characteristics in landslide management case reasoning;

[0084] When the entity vectors obtained based on vector calculation can reflect the similarity between entities, the knowledge graph embedding (KGE) method is used to perform knowledge representation learning on the entities and relationships in the knowledge graph, and the entity vector representation of text attributes is obtained through vector calculation; for the 1:N, N:1, and N:N entity relationships in the landslide management knowledge graph, the TransH knowledge embedding model is used to perform representation learning on the knowledge triples, and the landslide text attributes are converted into distributed feature vectors. The model can show good performance when processing the above complex relationships. The specific process is shown in the attached figure. Figure 3 ; The TransH model defines a hyperplane W for each relation r in the triple (h, r, t) r , and h ⊥ With t ⊥ As the projection of h and t on the hyperplane and satisfying Formula 1, its model scoring function is positioned as Formula 2;

[0085]

[0086]

[0087] In the formula, h represents the head entity, t represents the tail entity, r represents the relationship, and w r is the hyperplane W r The unit normal vector, h ⊥ represents the projection of the head entity on the hyperplane, t ⊥ Represents the projection of the tail entity on the hyperplane, L1 and L2 are the first and second normal forms respectively;

[0088] After learning the knowledge representation of landslide management attributes, the structural form of case representation is determined, that is, the 13 basic attributes of landslide management are combined to form the case representation of landslide management, and its structure is expressed in the formula:

[0089] C=(e1,e2,e3,...e n ) Formula 3;

[0090] In the formula, e n are the basic properties of landslide control, n=1, 2, 3, …, 13.

[0091] In the process of case retrieval and reuse, the K-nearest neighbor (KNN) algorithm is used to select K similar cases from the landslide management case library, namely: first, the local similarity between each attribute of the landslide management target case and the historical case is calculated, and then the global similarity is calculated by combining the weight of each attribute, and then sorted from high to low;

[0092] When calculating local similarity, the landslide management attributes are divided into numerical attributes and text attributes. For numerical attributes, the similarity of the same numerical attribute between the target case and the historical case is expressed by the distance after the two values ​​are normalized, as shown in the formula:

[0093]

[0094] in, represents the value of attribute k of the i-th historical case in the case library, represents the value of attribute k of target case j, and Represent the maximum and minimum values ​​of attribute k respectively, express and For text attributes, the distributed low-dimensional vector of the text attributes obtained after knowledge representation learning is used as the numerical representation of the text attributes, and the cosine similarity is used to calculate the similarity between text attributes, as shown in the formula:

[0095]

[0096] In the formula, and The vectors representing the attribute n of the target case j and the attribute n of the i-th historical case in the case library, Represents the cosine similarity between attributes n;

[0097] When calculating the global similarity between the target case and the historical case, the KNN algorithm is used to obtain k historical cases with higher global similarity as the result of case retrieval; the formula is

[0098]

[0099] In the formula, w m represents the weight of the mth attribute, represents the similarity between the i-th historical case in the case library and the target case attribute m, M represents the number of attributes, represents the global similarity between the i-th historical case in the case library and the target case;

[0100] In the case correction process, k similar cases are obtained from the historical case database, i.e., the geological disaster comprehensive prevention and control database, through case retrieval. Assuming that the historical case with the highest similarity is an erroneous case, the solution of the target case is affected by the biased historical case, and the case is corrected to obtain the solution of the target case. The size of the K nearest neighbor value is set, and the top-k historical cases are selected as the screening results in descending order of similarity, 1≤k≤10, and the array of engineering control measures is counted. If the total number of a certain control measure is greater than or equal to k / 2, the result is rounded down, and it is considered to be a commonly used control measure, and it is added to the recommendation list; according to the "Fujian Province Landslide Prevention and Control Design Technical Specifications (Trial)" and the actual situation of the geological disaster comprehensive prevention and control database, in the case correction process, the control measures are divided into 12 types of landslide control measures: interception and drainage, retaining wall, slope protection and greening, anchor rods or anchor cables, slope cutting and load reduction, anchoring, face wall, water interception skeleton, anti-slide pile, protection net, retaining wall, and blocking; for the landslide disaster points that have been controlled, 0 and 1 are used to indicate whether the control measures are adopted.

[0101] For the top-k landslide cases ranked by similarity, the number of each governance measure is counted. The calculation formula is as follows:

[0102]

[0103] Among them, m represents the similarity ranking, n represents the nth governance measure, S mn Indicates whether the historical case with similarity ranking m adopts the nth governance measure, ω represents the recommendation threshold, which is k / 2 here; the final statistical calculation result sum is rounded down, and each governance measure can only be recommended to the target case as an adoptable governance measure when the statistical calculation result is greater than or equal to 0.

[0104] In the method, the solving performance of the CBR decision-making auxiliary model is improved by optimizing the weight setting, that is, the gray wolf optimization algorithm GWO is used to realize adaptive weight optimization allocation, GWO simulates the hunting behavior of the gray wolf population, and realizes the optimization purpose through the mutual cooperation mechanism within the population. The dynamic optimization process of the wolf position in GWO is used to realize the refined weight allocation of the CBR model; the optimization core of GWO is to obtain the maximum or minimum value of the fitness function; the mean absolute error MAE (Mean Absolute Error, MAE) between the recommended governance measures and the actual governance measures of the KGCBR model is used as the fitness function;

[0105] The steps to implement the algorithm for adaptive weight optimization allocation are as follows:

[0106] Step a, set GWO operation parameters. Determine the maximum number of iterations, cutoff fitness value, solution space range, optimization parameter dimension, number of gray wolves and other related parameters of GWO;

[0107] Step b, initialize the GWO population. Input the weight value of each landslide attribute in the CBR case as an optimization parameter into the GWO;

[0108] Step c, CBR solution process execution and optimization. Assign the GWO initial parameter value as the historical optimal parameter value to the CBR model and execute the case-based reasoning solution process to calculate the fitness function value for each iteration;

[0109] Step d, determine the optimal weight value of the KGCBR model. Adjust the GWO parameters according to the calculated fitness value, iteratively update the optimal solution of the population, and when the number of GWO iterations reaches the maximum number of iterations or the calculated fitness function value reaches the cutoff fitness value, the minimum fitness value of GWO is obtained, which is the current optimal solution of GWO, and then determine the optimal weight value of each landslide attribute.

[0110] In the gray wolf optimization algorithm GWO, the gray wolf population has a strict population system, as shown in the attached Figure 4 As shown in the figure, α wolf is the leader of the wolf pack, representing the optimal solution of the model; β wolf assists α wolf, representing the suboptimal solution of the model; δ wolf obeys the orders of α wolf and β wolf to hunt, representing the general solution of the model; ω wolf follows other high-level wolves to maintain the balance of population relationships, representing the candidate solution of the model

[0111] The hunting behavior of the gray wolf population surrounding prey is defined by the formula

[0112] D=|C·X p (t)-X(t)| Formula 8;

[0113] X(t+1)=X p (t)-A·D i Formula 9;

[0114] A=2a·r1-a Formula 10;

[0115] C = 2·r2 Formula 11;

[0116] Among them, t represents the number of iterations of the current population, D represents the distance between the individual and the prey, and X p (t) and X(t) represent the positions of the prey and the individual, respectively. A and C represent the effective vectors of the synergy coefficients. a represents the convergence factor that decreases from 2 to 0. r1 and r2 are sets of random numbers in [0,1] to avoid falling into the local optimum.

[0117] The prey species of the gray wolf population hunting behavior is defined by the formula

[0118]

[0119] Among them, Dα , D β , D δ represents the distance between α, β, δ wolves and other individuals, X α , X β , X γ represents the current position of α, β, γ wolves, X(t) represents the current position of the gray wolf, C1, C2, C3 represent the adjustment coefficient vector; the following formula 13 defines the step size and direction of the ω individual in the wolf pack moving towards α, β, δ, and the following formula 14 defines the final position of the ω wolf;

[0120]

[0121] The algorithm optimization of GWO starts with randomly creating a population of gray wolves (candidate solutions). Specifically, during the iteration process, α, β, and δ wolves estimate the possible potential location of the prey, that is, the optimal solution; the gray wolves update their positions based on the distance between them and the prey. For development and exploration during the search process, the parameter a should be decreased from 2 to 0. If |A|>1, the candidate solution is far away from the prey; if |A|<1, the candidate solution is close to the prey. The flowchart of GWO is as follows Figure 5 shown.

[0122] In the fitness function of the gray wolf optimization algorithm GWO, MAE is used as an evaluation index to measure the deviation between the recommended measures and the actual measures. The smaller the MAE value, the higher the recommendation quality. Let the recommended control measures set be {p1, p2, ..., p3}, and the actual control measures set be {q1, q2, ..., q3}; where the set indicates whether each type of landslide control measure is adopted. If adopted, the score is 1, and if not adopted, the score is 0. Then MAE represents the error rate between the recommended control measures and the actual control measures. The MAE evaluation formula is expressed as:

[0123]

[0124] Among them, n represents the type of landslide control measures, which has a value of 12;

[0125] The GWO fitness function f is expressed as:

[0126]

[0127] Where m is the number of target cases in the validation set, MAE l represents the mean absolute error between the actual governance measures of the lth target case and the recommended governance measures of CBR;

[0128] In the decision-making method, the smaller the fitness function value is, the more reasonable the allocation weight is, which is more conducive to improving the accuracy of CBR solution.

[0129] Example:

[0130] This example uses Fujian Province as the research area. Fujian Province has a complex geological environment, and there are many extreme rainfalls, typhoons and rainstorms. Landslide disasters have occurred frequently over the years. The data used in the present invention comes from the Fujian Provincial Comprehensive Prevention and Control Database of Geological Hazards, including basic information of geological disasters, disaster-bearing bodies, comprehensive evaluation and prevention and other spatial data and attribute data. Based on the integrity and rationality of the data, 307 landslide control case data were selected as experimental data, 200 landslide control case data were selected as typical historical cases and used for knowledge graph construction and model training, 100 case data were selected as target cases and used for model verification, and the remaining 7 case data were used as engineering control cases and used for model testing. In order to verify the effectiveness and superiority of the KGCBR proposed in the present invention, it was compared with the CBR based on the average weight method (denoted as ACBR) and the CBR based on CRITIC (denoted as CCBR), and the experimental results under different k values ​​were compared and analyzed. The results are shown in Table 1 and Appendix. Figure 6 The evaluation index is used to measure the error rate between the actual governance measures and the recommended governance measures. The smaller the value, the better the recommendation effect.

[0131] Table 1 Comparison of MAE values ​​of different recommended methods (unit: %)

[0132]

[0133] From Table 1, Appendix Figure 6 It can be seen that the evaluation index results of the KGCBR proposed in this example under different k values ​​are better than those of the traditional CBR method. It can be seen that KGCBR has good performance in qualitative recommendation of landslide control measures, and can screen out highly similar cases from historical cases with differentiated disaster characteristics to achieve accurate push of highly applicable control measures. When the k value is equal to 9, the MAE value of KGCBR is 16.33%, which has the best performance, so it can be used for landslide control decision-making assistance tasks. In addition, through further analysis, it was found that when the k value is greater than 3, the MAE values ​​of the three methods are relatively lower and the recommendation effect is better. This is because the landslide control in the study area mostly adopts retaining walls, drainage interception, slope cutting and load reduction, slope protection and greening measures. Therefore, with the increase of k value, the model will further recommend more universal control measures, thereby reducing the error rate between actual measures and recommended measures; and when the k value is ≤3, the MAE values ​​of the three methods are relatively high, and the recommendation effect is poor. This is mainly because the number of typical historical cases collected by the present invention is small. In some cases, it is difficult to recommend historical cases with high similarity for the target cases, which leads to large differences between actual measures and recommended measures.

[0134] According to the comparative analysis, it can be found that the KGCBR method proposed in this example has the smallest recommendation error rate when the k value is equal to 9. Therefore, the k value of this model is set to 9 and it is used for the recommendation application task of landslide control measures. The remaining 7 landslide cases that did not participate in training and verification were tested as test samples.

[0135] In this example, the landslide case with landslide ID 40234 is selected for detailed analysis. The basic information of this sample is shown in Table 2.

[0136] By calculating the similarity between the sample and the historical cases, sorting them from high to low according to the similarity, and statistically screening the recommended governance measures of the top 10 historical cases, the top 10 historical cases retrieved are shown in Table 3.

[0137] Table 2 Basic information of the sample

[0138]

[0139] Table 3. Top 10 historical cases with similarity to sample recommendations

[0140]

[0141]

[0142] The k value is set to 9, and the number of each type of control measures in the top 9 historical case recommendation results is counted, as shown in Table 4. When the k value is equal to 9, according to the principle that the statistical control measures are greater than or equal to k / 2, the recommended control measures are "retaining wall + interception and drainage + anchor rods (cables) + slope protection and greening", which is completely consistent with the actual control measures of the sample "retaining wall + interception and drainage + anchor rods (cables) + slope protection and greening", meeting the actual needs of landslide control.

[0143] Table 4 Statistics of governance measures for similar historical cases

[0144]

[0145] The actual measures and recommended measures for the remaining six test cases are shown in Table 5. It can be concluded from the table that the actual measures of the six target cases are basically similar to the recommended measures, but due to the small number of historical cases of control measures such as baffles, protective nets, water interception skeletons, and anti-slide piles, it is difficult to accurately recommend such measures. In future research, it is necessary to further expand the historical case data set to enhance the effect of the recommendation model in landslide control applications.

[0146] Table 5 Test results of other cases

[0147]

Claims

1. A landslide control decision-making support method integrating knowledge graph and case-based reasoning, characterized by: The decision-making method is based on the landslide management auxiliary decision-making model KGCBR that integrates knowledge graph and case reasoning. The decision-making model comprehensively considers the geographic spatial characteristics of landslide attributes and the adaptive allocation of weights and introduces a case correction strategy to ensure the reliability and rationality of the recommended management measures. The construction process of the decision-making model includes the construction of the landslide management knowledge graph, the determination of the case reasoning solution process, and the adaptive weight optimization allocation based on the gray wolf optimization algorithm.

2. According to claim 1, a landslide control decision-making support method integrating knowledge graph and case-based reasoning is characterized by: The construction of the landslide management knowledge graph includes the following contents: First, construct a landslide management knowledge graph, use the TransH knowledge embedding model to capture the geographical similarities of landslide attributes, and integrate them into the case representation. Specifically, build a landslide management knowledge graph based on existing scientific and technological literature, laws and regulations, implementation standards, disaster statistics and existing ontology models. The type information of the landslide management knowledge graph includes the basic attributes of landslide disasters, spatial attributes including longitude and latitude, and landslide management measures.

3. According to claim 2, a landslide control decision-making assistance method integrating knowledge graph and case-based reasoning is characterized by: When constructing the knowledge graph of landslide management, the disaster statistical data is based on the provincial geological disaster comprehensive prevention and control database, and the data subject of the data source is structured data. First, based on the data table header name mapping method of the geological disaster comprehensive prevention and control database, the structured data is imported into the Neo4J graph database, and converted into a "head entity-relationship-tail entity" triple structure through the built-in CQL language to complete the construction of the landslide management knowledge graph.

4. According to claim 3, a landslide control decision-making support method integrating knowledge graph and case-based reasoning is characterized by: The case-based reasoning solution process includes case representation, case retrieval and reuse, and case correction; that is, using GWO to achieve adaptive weight optimization allocation to obtain the optimal weight of landslide attributes, and combining local and global similarity reasoning to apply to case retrieval; in order to avoid the randomness of the reasoning results, by adjusting the k value and counting the recommended measures of similar historical cases, the solution of the target case is corrected to achieve matching and recommendation of governance plans under the guidance of knowledge graphs, providing a refined engineering governance decision-making basis for landslide disasters with homogeneous characteristics.

5. According to claim 4, a landslide control decision-making support method integrating knowledge graph and case-based reasoning is characterized by: In the case representation process, the geographical similarity of landslide attributes is taken into account on the basis of the basic landslide management attributes of landslide slope, sliding body length, sliding body width, average sliding body thickness, sliding body volume, sliding body morphology, sliding body lithology, height difference, slope, slope direction, threat object, total number of threatened people, and threatened property, so as to optimize the description of geographical spatial characteristics in landslide management case reasoning; When the entity vectors obtained based on vector calculation can reflect the similarity between entities, the knowledge graph embedding method KGE is used to learn the knowledge representation of entities and relationships in the knowledge graph, and the entity vector representation of text attributes is obtained through vector calculation; for the 1:N, N:1, and N:N entity relationships in the landslide management knowledge graph, the TransH knowledge embedding model is used to learn the representation of knowledge triples, and the landslide text attributes are converted into distributed feature vectors. The model can show good performance when processing the above complex relationships; the TransH model defines a hyperplane W for each relationship r in the triple (h, r, t) r , and h ⊥ With t ⊥ As the projection of h and t on the hyperplane and satisfying Formula 1, its model scoring function is positioned as Formula 2; In the formula, h represents the head entity, t represents the tail entity, r represents the relationship, and w r is the hyperplane W r The unit normal vector, h ⊥ represents the projection of the head entity on the hyperplane, t ⊥ Represents the projection of the tail entity on the hyperplane, L1 and L2 are the first and second normal forms respectively; After learning the knowledge representation of landslide management attributes, the structural form of case representation is determined, that is, the 13 basic attributes of landslide management are combined to form the case representation of landslide management, and its structure is expressed in the formula: C=(e1,e2,e3,...e n ) Formula 3; In the formula, e n are the basic properties of landslide control, n=1, 2, 3, …, 13.

6. The landslide control decision-making support method integrating knowledge graph and case-based reasoning according to claim 5 is characterized by: In the process of case retrieval and reuse, the K-nearest neighbor algorithm KNN is used to select K similar cases from the landslide management case library, namely: first, the local similarity between each attribute of the landslide management target case and the historical case is calculated, and then the global similarity is calculated by combining the weight of each attribute, and then sorted from high to low; When calculating local similarity, the landslide management attributes are divided into numerical attributes and text attributes. For numerical attributes, the similarity of the same numerical attribute between the target case and the historical case is expressed by the distance after the two values ​​are normalized, as shown in the formula: in, represents the value of attribute k of the i-th historical case in the case library, represents the value of attribute k of target case j, and Represent the maximum and minimum values ​​of attribute k respectively, express and For text attributes, the distributed low-dimensional vector of the text attributes obtained after knowledge representation learning is used as the numerical representation of the text attributes, and the cosine similarity is used to calculate the similarity between text attributes, as shown in the formula: In the formula, and The vectors representing the attribute n of the target case j and the attribute n of the i-th historical case in the case library, Represents the cosine similarity between attributes n; When calculating the global similarity between the target case and the historical case, the KNN algorithm is used to obtain k historical cases with higher global similarity as the result of case retrieval; the formula is In the formula, w m represents the weight of the mth attribute, represents the similarity between the i-th historical case in the case library and the target case attribute m, M represents the number of attributes, Represents the global similarity between the i-th historical case in the case library and the target case.

7. The landslide control decision-making assistance method integrating knowledge graph and case-based reasoning according to claim 6 is characterized by: In the case correction process, k similar cases are obtained from the historical case database, i.e., the geological disaster comprehensive prevention and control database, through case retrieval. Assuming that the historical case with the highest similarity is an erroneous case, the solution of the target case is affected by the biased historical case, and the case is corrected to obtain the solution of the target case. Set the size of the K nearest neighbor value, select the top-k historical cases in descending order of similarity as the screening result, 1≤k≤10, and count the array of engineering control measures. If the total number of statistical measures of a certain control measure is greater than or equal to k / 2, and the result is rounded down, it is considered to be a commonly used control measure and added to the recommendation list; In the case revision process, the control measures are divided into 12 types of landslide control measures: drainage, retaining wall, slope protection and greening, anchor rods or anchor cables, slope reduction and load reduction, anchoring, face protection wall, water interception skeleton, anti-slide piles, protective net, retaining wall, and blocking; for the landslide disaster points that have been controlled, 0 and 1 are used to indicate whether the control measures are adopted. For the top-k landslide cases before the calculated similarity ranking, the number of each control measure is counted, and the calculation formula is as follows: Among them, m represents the similarity ranking, n represents the nth governance measure, S mn Indicates whether the historical case with similarity ranking m adopts the nth governance measure, ω represents the recommendation threshold, which is k / 2 here; the final statistical calculation result sum is rounded down, and each governance measure can only be recommended to the target case as an adoptable governance measure when the statistical calculation result is greater than or equal to 0.

8. The landslide control decision-making support method integrating knowledge graph and case-based reasoning according to claim 7 is characterized by: In the method, the solving performance of the CBR decision-making auxiliary model is improved by optimizing the weight setting, that is, the gray wolf optimization algorithm GWO is used to realize adaptive weight optimization allocation, GWO simulates the hunting behavior of the gray wolf population, and realizes the optimization purpose through the mutual cooperation mechanism within the population. The dynamic optimization process of the wolf position in GWO is used to realize the refined weight allocation of the CBR model; the optimization core of GWO is to obtain the maximum or minimum value of the fitness function; the mean absolute error MAE between the recommended governance measures and the actual governance measures of the KGCBR model is used as the fitness function; The steps to implement the algorithm for adaptive weight optimization allocation are as follows: Step a, set GWO operation parameters. Determine the maximum number of iterations, cutoff fitness value, solution space range, optimization parameter dimension, number of gray wolves and other related parameters of GWO; Step b, initialize the GWO population. Input the weight value of each landslide attribute in the CBR case as an optimization parameter into the GWO; Step c, CBR solution process execution and optimization. Assign the GWO initial parameter value as the historical optimal parameter value to the CBR model and execute the case-based reasoning solution process to calculate the fitness function value for each iteration; Step d, determine the optimal weight value of the KGCBR model. Adjust the GWO parameters according to the calculated fitness value, iteratively update the optimal solution of the population, and when the number of GWO iterations reaches the maximum number of iterations or the calculated fitness function value reaches the cutoff fitness value, the minimum fitness value of GWO is obtained, which is the current optimal solution of GWO, and then determine the optimal weight value of each landslide attribute.

9. The landslide control decision-making support method integrating knowledge graph and case-based reasoning according to claim 8 is characterized by: In the gray wolf optimization algorithm GWO, the gray wolf population has a strict population system, in which α wolf is the leader of the wolf pack, representing the optimal solution of the model; β wolf assists α wolf, representing the suboptimal solution of the model; δ wolf obeys the orders of α wolf and β wolf to hunt, representing the general solution of the model; ω wolf follows other high-level wolves to maintain the balance of population relations, representing the candidate solution of the model The hunting behavior of the gray wolf population surrounding prey is defined by the formula D=|C·X p (t)-X(t)| Formula 8; X(t+1)=X p (t)-A·Di Formula 9; A=2a·r1-a Formula 10; C = 2·r2 Formula 11; Among them, t represents the number of iterations of the current population, D represents the distance between the individual and the prey, and X p (t) and X(t) represent the positions of the prey and the individual, respectively. A and C represent the effective vectors of the synergy coefficients. a represents the convergence factor that decreases from 2 to 0. r1 and r2 are sets of random numbers in [0,1] to avoid falling into the local optimum. The prey prey of the gray wolf population hunting behavior is defined by the formula Among them, D α , D β , D δ represents the distance between α, β, δ wolves and other individuals, X α , X β , X γ represents the current position of α, β, and γ wolves, X(t) represents the current position of the gray wolf, and C1, C2, and C3 represent the adjustment coefficient vector; The following formula 13 defines the step length and direction of the wolf ω individual moving toward α, β, δ, and the following formula 14 defines the final position of the wolf ω; The algorithm optimization of GWO starts with randomly creating a population of gray wolves, that is, creating candidate solutions. Specifically, during the iteration process, α, β, and δ wolves estimate the possible potential location of the prey, that is, the optimal solution; the gray wolves update their positions based on the distance between them and the prey. For development and exploration during the search process, the parameter a should be decreased from 2 to 0. If |A|>1, the candidate solution is far away from the prey; if |A|<1, the candidate solution is close to the prey.

10. The landslide control decision-making assistance method integrating knowledge graph and case-based reasoning according to claim 9 is characterized by: In the fitness function of the gray wolf optimization algorithm GWO, MAE is used as an evaluation index to measure the deviation between the recommended measures and the actual measures. The smaller the MAE value, the higher the recommendation quality. Let the recommended control measures set be {p1, p2, ..., p3}, and the actual control measures set be {q1, q2, ..., q3}; where the set indicates whether each type of landslide control measure is adopted. If adopted, the score is 1, and if not adopted, the score is 0. Then MAE represents the error rate between the recommended control measures and the actual control measures. The MAE evaluation formula is expressed as: Among them, n represents the type of landslide control measures, which has a value of 12; The GWO fitness function f is expressed as: Where m is the number of target cases in the validation set, MAE l represents the mean absolute error between the actual governance measures of the lth target case and the recommended governance measures of CBR; In the decision-making method, the smaller the fitness function value is, the more reasonable the allocation weight is, which is more conducive to improving the accuracy of CBR solution.

Images