Landslide risk dynamic early warning method and system based on Bayesian network

Abstract

The invention provides a landslide risk dynamic early warning method and system based on a Bayesian network, and belongs to the field of geological disaster intelligent prediction, and the method comprises the steps: constructing a disaster-inducing factor data set, and screening key factors based on a Pearson's correlation coefficient and an information gain method; an FP-Growth algorithm is further utilized to extract association rules among high-confidence factors, a Bayesian network structure is guided to be optimized, and a Bayesian network model with a causal relationship is constructed; the model supports an incremental learning mechanism based on a newly added landslide sample, can dynamically update a conditional probability table, and realizes landslide probability prediction and risk grade division in combination with a Bayesian forward reasoning result. The method has the advantages of being high in causal reasoning ability, high in model structure expression ability, excellent in prediction precision and capable of supporting real-time updating and risk partition, and is suitable for an intelligent risk assessment and early warning system for landslide disasters.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction of geological disasters, and specifically relates to a dynamic early warning method and system for landslide risk based on Bayesian networks. Background Technology

[0002] With the increasing intensity of urban development in mountainous areas and the continuous expansion of transportation and energy infrastructure, geological disasters such as landslides pose a serious threat to the safety of engineering facilities. Especially along typical linear engineering projects such as oil pipelines, power transmission lines, mountain roads, railways, and tunnels, landslides are characterized by high incidence, frequent occurrence, and sudden onset. Landslide formation is driven by the nonlinear coupling of multiple factors, including topographic slope, lithological structure, extreme rainfall, vegetation cover, and human engineering activities. Risk assessment requires the integration of geospatial analysis, probabilistic reasoning, and dynamic monitoring technologies. Traditional prediction methods have significant limitations in terms of model interpretability, dynamic adaptability, and engineering relevance.

[0003] Currently, landslide risk assessment methods mainly fall into three categories: physical models, statistical models, and machine learning models. Physical models typically require detailed soil physical parameters, rainfall infiltration characteristics, and slope structure data. They calculate slope safety factors to determine the probability of landslides and can explain local landslide formation mechanisms, but they have extremely high data requirements and are difficult to generalize to large areas. Statistical models, such as information content models, weighted summation methods, and analytic hierarchy process (AHP), are simple to model, but they generally rely on manual weighting, are highly subjective, have poor causal explanation capabilities, and lack dynamic adaptability, making them difficult to cope with changes in the external environment. In recent years, machine learning models have been gradually introduced into the field of landslide prediction, such as support vector machines (SVM), random forests (RF), and XGBoost models, which have shown good performance in handling high-dimensional and nonlinear data. However, most of these models are "black box" models, only providing predictive results and unable to analyze the causal chain of landslide formation, lacking reasoning and explanation capabilities, and making it difficult to achieve post-disaster cause tracing and key factor identification. In addition, existing models are generally static in architecture and lack real-time data update mechanisms, which cannot support the needs of dynamic landslide early warning and limit their practicality in infrastructure scenarios. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a dynamic early warning method for landslide risk based on Bayesian networks, comprising the following steps:

[0005] Step S1: Construct a multidimensional disaster-causing factor dataset and obtain the multidimensional disaster-causing factor set {X1, X... 2, …,X sum}, where sum is the initial total number of disaster-causing factors. The multidimensional disaster-causing factor set undergoes multi-level screening and discretization to obtain a discretized disaster-causing factor set. , where n is the total number of disaster-causing factors after screening, and training and test sets are constructed proportionally;

[0006] Step S2: Use the FP-Growth algorithm to analyze the discretized set of disaster-causing factors. Perform frequent itemset extraction based on the preset minimum support S. min and minimum confidence level C min Generate a form like The set of high-confidence association rules, among which The disaster-causing factors are discrete;

[0007] Step S3: Set the landslide occurrence node Root as the root node, and the discretized set of hazard-causing factors as child nodes; add directed edges between the preceding and succeeding terms according to the association rule set. An initial Bayesian network structure is constructed, and the network topology is optimized using the K2+ hill-climbing algorithm to obtain the final Bayesian network structure.

[0008] Step S4: Based on maximum likelihood estimation and expert prior knowledge, construct a conditional probability table CPT using the training set to quantify the probability distribution of each node under the parent node's state;

[0009] Step S5: Calculate the probability of landslide occurrence under a given combination of environmental factors using forward reasoning. The risk level of landslides is classified by combining geospatial units, and the contribution of each factor is calculated through reverse reasoning. Screening for the main disaster-causing factors.

[0010] Beneficial effects:

[0011] 1. This invention provides a dynamic early warning method for landslide risk based on Bayesian networks. It uses the FP-Growth algorithm to mine high-confidence rules between factors from historical data, guides the addition of edges and topology optimization of the Bayesian network structure, overcomes the oversimplification problem of traditional naive Bayesian structure, improves the model's ability to express non-independent relationships between variables, makes the model more consistent with the real causal dependencies in the landslide disaster process, and improves the accuracy of inference.

[0012] 2. This invention takes landslide occurrence as the target node, constructs a Bayesian network structure that integrates significant correlations between factors, supports forward probabilistic reasoning and reverse causal tracing, and introduces factor contribution to screen the main control factors. It overcomes the limitations of traditional "black box" models that can only output results and cannot identify the causal chain and main control factors, realizes interpretable modeling of landslide occurrence mechanism, and supports post-disaster causal tracing and intelligent identification of key disaster-causing factors.

[0013] 3. This invention designs an incremental parameter update process based on newly added landslide samples, which solves the problem that the existing landslide prediction model architecture is static and difficult to update in real time, and realizes the dynamic adaptation and self-updating capability of the landslide monitoring model. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a dynamic early warning method for landslide risk based on Bayesian networks according to the present invention.

[0015] Figure 2 A comparison of ROC curve results between the basic Bayesian network and the Bayesian network optimized by the association rules of this invention;

[0016] Figure 3 This is a structural block diagram of a landslide risk dynamic early warning system based on Bayesian networks according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0018] Example 1

[0019] like Figure 1 As shown in the figure, an embodiment of the present invention provides a dynamic early warning method for landslide risk based on Bayesian networks, which includes the following steps:

[0020] Step S1: Construct a multidimensional disaster-causing factor dataset and obtain the multidimensional disaster-causing factor set {X1, X... 2, …,X sum}, where sum is the initial total number of disaster-causing factors. The multidimensional disaster-causing factor set undergoes multi-level screening and discretization to obtain a discretized disaster-causing factor set. , where n is the total number of disaster-causing factors after screening, and training and test sets are constructed proportionally;

[0021] Step S2: Use the FP-Growth algorithm to analyze the discretized set of disaster-causing factors. Perform frequent itemset extraction based on the preset minimum support S. min and minimum confidence level C min Generate a form like The set of high-confidence association rules, among which The disaster-causing factors are discrete;

[0022] Step S3: Set the landslide occurrence node Root as the root node, and the discretized set of hazard-causing factors as child nodes; add directed edges between the preceding and succeeding terms according to the association rule set. An initial Bayesian network structure is constructed, and the network topology is optimized using the K2+ hill-climbing algorithm to obtain the final Bayesian network structure.

[0023] Step S4: Based on maximum likelihood estimation and expert prior knowledge, construct a conditional probability table CPT using the training set to quantify the probability distribution of each node under the parent node's state;

[0024] Step S5: Calculate the probability of landslide occurrence under a given combination of environmental factors using forward reasoning. The risk level of landslides is classified by combining geospatial units, and the contribution of each factor is calculated through reverse reasoning. Screening for the main disaster-causing factors.

[0025] In one embodiment, step S1 above: constructing a multi-source heterogeneous geological hazard database, specifically includes:

[0026] Step S11: Based on the field survey and historical landslide database, extract the coordinates of all landslide points within the study area and generate a positive sample set. L i The spatial location of the landslide point; with each landslide point L i A circular buffer zone with a radius of R = 500m is constructed centered on the landslide site. Outside the buffer zone, non-landslide points are randomly selected at a 1:1 ratio to generate a negative sample set. Label positive samples as Y=1 and negative samples as Y=0 to construct a labeled initial dataset. Thus, a multidimensional set of disaster-causing factors {X1, X...} is obtained. 2, …,X sum}, where sum is the initial total number of disaster-causing factors;

[0027] This invention integrates topographic, geological, meteorological, hydrological, vegetation, human activity factors, and historical landslide data of the study area to construct a multidimensional disaster-causing factor dataset that includes elevation, aspect, slope, plane curvature, profile curvature, surface relief, topographic roughness, runoff intensity index (SPI), topographic humidity index (TWI), topographic ruggedness index (TRI), lithology, distance from fault, distance from epicenter, distance from river, land use type, soil type, annual precipitation, normalized difference vegetation index (NDVI), POI kernel density, and distance from road.

[0028] Step S12: Using Pearson correlation coefficient PCC, variance inflation factor VIF, tolerance TOL, and information gain IG, analyze the multidimensional catastrophic factor set {X1, X...} 2, …,Xsum After multi-level screening, an initial set of key disaster-causing factors {X1, X...} is obtained. 2, …,X n}, where n is the total number of disaster-causing factors after screening, specifically including:

[0029] Step S121: Calculate the Pearson correlation coefficient (PCC) among the various hazard-causing factors to measure their linear correlation. The calculation formula is as follows:

[0030] ;

[0031] Among them, A i Let B be the i-th sample of causative factor A; i This is the i-th sample of disaster-causing factor B; and , where are the means of variables A and B respectively; r is the correlation coefficient between disaster-causing factors A and B;

[0032] The larger the absolute value of the PCC value, the stronger the correlation between the two disaster-causing factors. The threshold of |r|>0.7 is set as the basis for judging the redundancy of variables. For a set of variables with strong cross-correlation, only the factors that are most representative of landslide occurrence or have the least correlation with other variables are retained, and the rest are removed.

[0033] Step S122: To avoid strong multicollinearity among factors, calculate the tolerance (TOL) and variance inflation factor (VIF) for each variable, as shown in the following formula:

[0034] ;

[0035] ;

[0036] Among them, R 2 VIF is the variance between variables; TOL is the tolerance; the smaller the TOL value, the better the variable can be explained by other variables, and the more serious the multicollinearity problem is; the larger the VIF value, the greater the influence of the variable on the regression model, which may lead to model instability; VIF>10 or TOL<0.1 are used as the threshold for judging multicollinearity, and factors that do not meet the requirements are eliminated.

[0037] Step S123: Information gain (IG) is introduced as a feature selection index to measure the degree to which each factor reduces the uncertainty of landslide occurrence. Information gain is calculated based on the concept of entropy to determine the information contribution of a factor to the target variable (landslide occurrence) under known conditions. The calculation formula is as follows:

[0038] ;

[0039] ;

[0040] Where X is the disaster-causing factor; Y is the landslide occurrence state, which is a binary variable (1 indicates that a landslide has occurred, and 0 indicates that it has not occurred); y i Let P(y) be the i-th possible value of the landslide state variable Y; i The value of y represents the state of the landslide. i The probability of landslide occurrence; H(Y) represents the entropy of the landslide occurrence variable, which measures the uncertainty of the system; H(Y|X) represents the conditional entropy of the landslide occurrence variable given factor X; IG(Y,X) represents the information gain, which measures the contribution of factor X to reducing the uncertainty of landslide occurrence.

[0041] The higher the IG value, the greater the contribution of the factor to the information of landslide occurrence; set a threshold of IG>0.01 to remove factors that do not meet the requirements;

[0042] Multidimensional disaster-causing factor set {X1, X 2, …,X sum After the above multi-level screening process, the initial set of key disaster-causing factors {X1, X...} is obtained. 2, …,X n}, where n is the total number of disaster-causing factors after screening;

[0043] Step S13: Use Jenks' natural breakpoint method to analyze the initial set of critical disaster-causing factors {X1, X...} 2, …,X n The data is then discretized and hierarchically processed to generate a discretized set of disaster-causing factors. , where n is the total number of discretized disaster-causing factors;

[0044] For the initial set of key disaster-causing factors {X1, X... 2, …,X n For continuous variables in the table, the Jenks natural break method is used to determine the grading threshold by minimizing the within-class variance.

[0045] ;

[0046] Where C is the classification scheme, k is the number of levels, and μ i The mean of the i-th class;

[0047] The obtained discretized set of disaster-causing factors Divided into 5 levels, marked as ,in This indicates that the i-th disaster-causing factor is at the m-th discrete level;

[0048] Step S14: Dataset Partitioning: Divide the discretized set of disaster-causing factors into training set D in a 7:3 ratio. train and test set D test The training set is used for subsequent Bayesian network parameter learning, and the test set is used for model validation.

[0049] In one embodiment, step S2 above: using the FP-Growth algorithm to process the discretized set of disaster-causing factors Perform frequent itemset extraction based on the preset minimum support S. min and minimum confidence level C min Generate a form like The set of high-confidence association rules, among which The discretized hazard factors specifically include:

[0050] Step S21: Construct a transaction database based on training samples, where each transaction consists of discretized disaster-causing factors, serving as the basic data for frequent itemset mining;

[0051] Based on the set of discretized disaster-causing factors Obtain the corresponding discretized disaster-causing factor dataset. Where N is the total number of data points;

[0052] Will Considering it as a transactional database T, for Each transaction in L ij Indicates disaster-causing factors Discrete rank in the j-th sample;

[0053] Complete series A hierarchy of all discretized hazard-causing factors;

[0054] Step S22: Frequent Itemset Mining: Using the FP-Growth frequent pattern growth algorithm, set a minimum support threshold S. min Extract frequent itemsets that meet the support and confidence requirements from the transaction database;

[0055] Statistical analysis of each item Number of occurrences in T Calculate support , retain satisfaction (Minimum support S in this embodiment of the invention) min The frequent terms with support = 0.1 are sorted in descending order; each transaction t is then sorted sequentially. j Frequent items are inserted into a frequent pattern tree (FP-Tree). Each node records the item name and count, and a linked list maintains the connections between nodes with the same item. For each frequent item in the FP-Tree... (Starting from the lowest support), extract its conditional pattern base along the reverse path, reconstruct the conditional FP-Tree based on the conditional pattern base, and recursively mine to generate frequent itemsets; the support calculation formula is:

[0056] ;

[0057] in, For frequent itemsets The number of times it appears in all transactions, where N is the total number of transactions. ,but Frequent itemsets that meet the support threshold;

[0058] Step S23: Association rule generation and filtering: For each factor combination in the frequent itemset, generate a rule of the form... The initial association rules were determined, and candidate association rules were obtained by filtering them based on support and confidence. Then, combining landslide disaster mechanisms and domain knowledge, a second filtering process was performed on the candidate association rules, retaining only the set of association rules that have causal logic with the actual landslide occurrence process. Specifically, this set includes:

[0059] Each factor Having m i Each level, for any two factors and (i≠j), enumerate all its level pairs: There are a total of Candidate rules; for each candidate rule Calculate the confidence level in the transaction database T:

[0060] ;

[0061] Only retain those that meet the requirements. and The rule set R cand For R cand Each rule in the algorithm is reviewed by experts in conjunction with the landslide formation mechanism, and rules that do not conform to geological or hydrological logic are eliminated. After threshold screening and mechanism verification, a high-confidence association rule set R is obtained, which provides causal priors for subsequent Bayesian network structure optimization.

[0062] In one embodiment, step S3 above involves setting the landslide occurrence node Root as the root node and the discretized set of causative factors as child nodes; and adding directed edges between the preceding and succeeding terms based on the association rule set. An initial Bayesian network structure is constructed, and the network topology is optimized using the K2+ hill-climbing algorithm to obtain the final Bayesian network structure, specifically including:

[0063] Step S31: Using the landslide occurrence node Root as the root node of the Bayesian network, the discretized set of disaster-causing factors... As a child node, the initial set of directed edges is Construct the initial Bayesian network structure;

[0064] Traditional Naive Bayes networks assume that all influencing factors are independent and uniformly point to the target variable "whether a landslide has occurred." However, their conditional probability table (CPT) grows rapidly under high-dimensional input, leading to increased parameter learning costs and reduced inference efficiency. Therefore, a divergent Naive Bayes structure is adopted, with the "whether a landslide has occurred" node (Root) as the root node, and the remaining n discretized disaster-causing factors... As child nodes, the initial set of directed edges is: ;

[0065] Step S32: Add the association rule set to the edge set to form the edge set with supplemented rules: ;

[0066] Step S33: Based on the K2+ hill-climbing algorithm, local topology optimization is performed, and the Bayesian Information Criterion (BIC) is introduced to construct the scoring function of the Bayesian network.

[0067] ;

[0068] Where G is the current Bayesian network structure, D is the training dataset, and the total number of samples is N; for the i-th node r i for The number of possible values ​​for q i For its parent node set Pa( The possible configuration number N ij Configure the parent node in the sample with the same number of records as the j-th combination; N ijk For a sample that simultaneously satisfies the condition that the parent node is configured as type j, and The number of records for the k-th value; the first item It is the log-likelihood value of data D under G; the second term It is a penalty term that applies to the number of model parameters with a weight of lnN;

[0069] Step S34: Using E1 as the initial edge set, perform the following iterative search steps on the initial Bayesian network:

[0070] Step S341: In the current edge set E t Using three local operations—adding, deleting, or reversing a single directed edge—can yield candidate structures. ;

[0071] Step S342: For each candidate structure calculate ;

[0072] Step S343: If it exists Then update Otherwise, retain the original structure; terminate when the maximum number of iterations is reached, and output the Bayesian network structure that reflects both data-driven causal priors and is optimal in the overall scoring sense.

[0073] In one embodiment, step S4 above, which involves constructing a conditional probability table CPT based on maximum likelihood estimation and expert prior knowledge using the training set to quantify the probability distribution of each node under the parent node's state, specifically includes:

[0074] Step S41: For each node in the training set and its parent node configuration Pa( Frequency statistics are performed on j, and Dirichlet prior parameters are introduced based on maximum likelihood estimation. compute nodes Given the posterior conditional probability of each value k under the parent node state j, construct the node's conditional probability table CPT.

[0075] For training set D train Each node and its parent node configuration Pa( )=j, Statistics:

[0076] ;

[0077] Dirichlet prior parameters are introduced based on maximum likelihood estimation (MLE). (Set as a uniform constant) Based on the Bayesian update fusion maximum likelihood estimation, the posterior probability is obtained:

[0078] ;

[0079] Where, r i for The number of possible values ​​for q i Pa ( The number of combinations of ) when At that time, it degenerates into pure MLE: ; Put all nodes Under each parent node configuration j, and for each value k... The organization is stored as a conditional probability table (CPT);

[0080] Step S42: When the number of new samples reaches a set threshold, while retaining the original Bayesian network structure and historical parameters, based on each node in the new samples... The frequency statistics of the configuration of the parent node, combined with the unified Dirichlet prior, are used to locally update the original conditional probability table CPT, thereby realizing the continuous correction of the parameters of the Bayesian network structure and the dynamic adaptation of the landslide disaster mode.

[0081] This update process retains the existing Bayesian network structure unchanged, without rebuilding the network topology or clearing or resetting existing parameters; instead, it uses new sample data to perform cumulative statistics and correct the original probability distribution based on the existing conditional probability table (CPT); specifically, it includes the following operations:

[0082] For each node And its corresponding parent node configuration j, count the frequency of newly added samples. and ;

[0083] Update conditional probability terms: ,in, These are Dirichlet priors used for smoothing and to avoid zero-probability issues.

[0084] Update the CPT entries for all nodes to accommodate the incorporation of incremental sample knowledge, ensuring the model's responsiveness to the latest landslide trigger states.

[0085] In one embodiment, step S5 above involves calculating the probability of a landslide occurring under a given combination of environmental factors using forward reasoning. The risk level of landslides is classified by combining geospatial units, and the contribution of each factor is calculated through reverse reasoning. Screening for the main disaster-causing factors, specifically including:

[0086] Step S51: Forward Reasoning: Using the Bayesian network forward reasoning mechanism, input the combination of factors under the current or simulated scenario, and calculate the posterior probability of the landslide node Root. The landslide probability values ​​obtained through reasoning are mapped to spatial units, and risk levels are divided according to a set probability threshold to achieve geospatial zoning of landslide risk.

[0087] Forward reasoning: For any given combination of environmental factors, state variables The posterior probability of a landslide occurring is calculated using the constructed Bayesian network:

[0088] ;

[0089] Where P(Y=1) is the prior probability of a landslide occurring. The CPT constructed in step S4 has a denominator that is a total probability normalization term;

[0090] This allows us to predict the posterior probability of landslides occurring under different environments.

[0091] Project the predicted landslide probability P (Y=1) onto the buffer zone. Within the grid unit, landslide risk levels are divided into five categories: "extremely high, high, medium, low, and extremely low" based on the natural breakpoint method, thus realizing a spatial risk zoning map.

[0092] Step S52: Reverse reasoning: Calculate the marginal probability change of each disaster-causing factor under the premise of landslide occurrence. It is used to identify the main controlling factors that have a significant impact on landslide risk.

[0093] Given that a landslide has occurred (Y=1), evaluate each factor X. i Posterior probabilities for different states:

[0094] ;

[0095] in, for The prior distribution;

[0096] Calculation of the contribution of the main control factors: To measure the degree of influence of each factor on the occurrence of landslides, the contribution increment is defined as follows:

[0097] ;

[0098] By comparison The size of all factors is used to sort them and screen out the dominant disaster-causing factors that have the greatest driving effect on landslide initiation.

[0099] The forward principle of this invention can be used to guide whether to carry out prevention and control measures, while the reverse principle can be used to clarify how to carry out prevention and control measures.

[0100] This invention evaluates the performance advantages of a Bayesian network model in landslide risk prediction by constructing comparative experiments, specifically including the following:

[0101] To verify the effectiveness of association rules in optimizing Bayesian network structures, this invention constructs two model structures respectively:

[0102] Traditional basic model M base : Constructed using a divergent Naive Bayes structure, that is, it only contains the "landslide occurrence" node Root pointing to all disaster-causing factor nodes X. i The initial structure did not introduce any inter-factor association edges;

[0103] The enhanced model M of the present invention enhanced In terms of basic structure, the high-confidence association rules obtained in step S2 are integrated, and the directed edges corresponding to the rules that conform to the geological mechanism are... The network structure was supplemented and optimized using the K2+ hill-climbing algorithm.

[0104] Both models mentioned above are trained on the same training set D. train Parameter learning was completed on the independent test set D. test Performance evaluation was conducted on the above.

[0105] Evaluation indicators include:

[0106] Accuracy: , where TP, TN, FP, and FN represent the matching between the predicted results of landslides / non-landslides and the actual values, respectively;

[0107] Recall: This measures the model's ability to detect actual landslide points.

[0108] AUC: The area under the ROC curve reflects the overall performance of the model in terms of risk discrimination.

[0109] like Figure 2 As shown, M will eventually be base With M enhanced A comparative analysis was conducted on the performance of the three indicators mentioned above. The enhanced model of this invention outperforms the basic model on multiple indicators, demonstrating that introducing significant correlations between factors into the Bayesian network structure helps improve the model's ability to express landslide occurrence mechanisms and its prediction accuracy.

[0110] Example 2

[0111] like Figure 3 As shown, this embodiment of the invention provides a dynamic early warning system for landslide risk based on Bayesian networks, comprising the following modules:

[0112] The module 61 for constructing a discretized disaster-causing factor set is used to construct a multidimensional disaster-causing factor dataset and obtain the multidimensional disaster-causing factor set {X1, X... 2, …,X sum}, where sum is the initial total number of disaster-causing factors. The multidimensional disaster-causing factor set undergoes multi-level screening and discretization to obtain a discretized disaster-causing factor set. , where n is the total number of disaster-causing factors after screening, and training and test sets are constructed proportionally;

[0113] The module 62 for obtaining the association rule set is used to apply the FP-Growth algorithm to the discretized set of disaster-causing factors. Perform frequent itemset extraction based on the preset minimum support S. min and minimum confidence level C min Generate a form like The set of high-confidence association rules, among which The disaster-causing factors are discrete;

[0114] Module 63 for constructing and optimizing the Bayesian network structure is used to set the landslide occurrence node Root as the root node and the discretized set of disaster-causing factors as child nodes; and to add directed edges between the preceding and succeeding terms based on the association rule set. An initial Bayesian network structure is constructed, and the network topology is optimized using the K2+ hill-climbing algorithm to obtain the final Bayesian network structure.

[0115] The Bayesian network parameter learning module 64 is used to construct a conditional probability table CPT based on the training set, using maximum likelihood estimation and expert prior knowledge, to quantify the probability distribution of each node in the state of the parent node.

[0116] The landslide risk bidirectional reasoning and graded early warning module 65 is used to calculate the probability of landslide occurrence under a given combination of environmental factors through forward reasoning. The risk level of landslides is classified by combining geospatial units, and the contribution of each factor is calculated through reverse reasoning. Screening for the main disaster-causing factors.

[0117] A landslide risk dynamic early warning device based on Bayesian networks includes one or more electronic devices, wherein the one or more electronic devices are used to implement a landslide risk dynamic early warning method based on Bayesian networks.

[0118] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement a dynamic early warning method for landslide risk based on Bayesian networks.

[0119] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement a dynamic early warning method for landslide risk based on a Bayesian network.

[0120] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A dynamic early warning method for landslide risk based on Bayesian networks, characterized in that, include: Step S1: Construct a multidimensional disaster-causing factor dataset and obtain the multidimensional disaster-causing factor set {X1, X... 2, …,X sum }, where sum is the initial total number of disaster-causing factors. The multidimensional disaster-causing factor set undergoes multi-level screening and discretization to obtain a discretized disaster-causing factor set. , where n is the total number of disaster-causing factors after screening, and training and test sets are constructed proportionally; Step S2: Use the FP-Growth algorithm to process the discretized set of disaster-causing factors. Perform frequent itemset extraction based on the preset minimum support S. min and minimum confidence level C min Generate a form like The set of high-confidence association rules, among which, The disaster-causing factors are discrete; Step S3: Set the landslide occurrence node Root as the root node, and the discretized set of disaster-causing factors as child nodes; add directed edges between the preceding and succeeding terms according to the association rule set. An initial Bayesian network structure is constructed, and the network topology is optimized using the K2+ hill-climbing algorithm to obtain the final Bayesian network structure. Step S4: Based on maximum likelihood estimation and expert prior knowledge, construct a conditional probability table CPT using the training set to quantify the probability distribution of each node under the parent node's state; Step S5: Calculate the probability of landslide occurrence under a given combination of environmental factors using forward reasoning. The risk level of landslides is classified by combining geospatial units, and the contribution of each factor is calculated through reverse reasoning. Screening for the main disaster-causing factors.

2. The landslide risk dynamic early warning method based on Bayesian networks according to claim 1, characterized in that, Step S1: Construct a multidimensional disaster-causing factor dataset and obtain a multidimensional disaster-causing factor set {X1, X... 2, …,X sum }, where sum is the initial total number of disaster-causing factors. The multidimensional disaster-causing factor set undergoes multi-level screening and discretization to obtain a discretized disaster-causing factor set. , where n is the total number of disaster-causing factors after screening, and training and test sets are constructed proportionally, specifically including: Step S11: Based on the field survey and historical landslide database, extract the coordinates of all landslide points within the study area and generate a positive sample set. , where L i The spatial location of the landslide point; with each landslide point L i A circular buffer zone with a radius of R = 500m is constructed centered on the landslide site. Outside the buffer zone, non-landslide points are randomly selected at a 1:1 ratio to generate a negative sample set. Label positive samples as Y=1 and negative samples as Y=0 to construct a labeled initial dataset. Thus, a multidimensional set of disaster-causing factors {X1, X...} is obtained. 2, …,X sum }, where sum is the initial total number of disaster-causing factors; Step S12: Using Pearson correlation coefficient PCC, variance inflation factor VIF, tolerance TOL, and information gain IG, analyze the multidimensional catastrophic factor set {X1, X...} 2, …,X sum After multi-level screening, an initial set of key disaster-causing factors {X1, X...} is obtained. 2, …,X n }, where n is the total number of disaster-causing factors after screening; Step S13: Use Jenks' natural breakpoint method to analyze the initial set of critical disaster-causing factors {X1, X...} 2, …,X n The data is then discretized and hierarchically processed to generate a discretized set of disaster-causing factors. , where n is the total number of discretized disaster-causing factors; Step S14: Dataset partitioning: Divide the discretized set of disaster-causing factors into a training set D at a ratio of 7:

3. train and test set D test The training set is used for learning Bayesian network parameters, and the test set is used for model validation.

3. The landslide risk dynamic early warning method based on Bayesian networks according to claim 2, characterized in that, Step S2: The FP-Growth algorithm is used to process the discretized set of disaster-causing factors. Perform frequent itemset extraction based on the preset minimum support S. min and minimum confidence level C min Generate a form like The set of high-confidence association rules, among which, The discretized hazard factors specifically include: Step S21: Construct a transaction database based on training samples, where each transaction consists of discretized disaster-causing factors, serving as the basic data for frequent itemset mining; Step S22: Frequent Itemset Mining: Using the FP-Growth frequent pattern growth algorithm, set a minimum support threshold S. min Extract frequent itemsets that meet the support and confidence requirements from the transaction database; Step S23: Association rule generation and filtering: For each factor combination in the frequent itemset, generate a rule of the form... The initial association rules are obtained, and the initial association rules are filtered based on support and confidence to obtain candidate association rules. Combining the landslide disaster mechanism and domain knowledge, the candidate association rules are filtered a second time, and only the set of association rules with causal logic to the actual landslide occurrence process is retained.

4. The landslide risk dynamic early warning method based on Bayesian networks according to claim 3, characterized in that, Step S3: Set the landslide occurrence node Root as the root node, and the discretized set of disaster-causing factors as child nodes; add directed edges between the preceding and succeeding terms according to the association rule set. An initial Bayesian network structure is constructed, and the network topology is optimized using the K2+ hill-climbing algorithm to obtain the final Bayesian network structure, specifically including: Step S31: Using the landslide occurrence node Root as the root node of the Bayesian network, the discretized set of disaster-causing factors... As a child node, the initial set of directed edges is Construct the initial Bayesian network structure; Step S32: Add the association rule set to the edge set to form a rule-supplemented edge set: ; Step S33: Based on the K2+ hill-climbing algorithm, local topology optimization is performed, and the Bayesian Information Criterion (BIC) is introduced to construct the scoring function of the Bayesian network. ； Where G is the current Bayesian network structure, D is the training dataset, and the total number of samples is N; for the i-th node r i for The number of possible values ​​for q i For its parent node set Pa( The possible configuration number N ij Configure the parent node in the sample with the same number of records as the j-th combination; N ijk For a sample that simultaneously satisfies the condition that the parent node is configured as type j, and The number of records for the k-th value; the first item It is the log-likelihood LL of data D under G; the second term It is a penalty term that applies to the number of model parameters with a weight of lnN; Step S34: Using E1 as the initial edge set, perform the following iterative search steps on the initial Bayesian network: Step S341: In the current edge set E t Using three local operations—adding, deleting, or reversing a single directed edge—can yield candidate structures. ; Step S342: For each candidate structure calculate ; Step S343: If it exists Then update Otherwise, retain the original structure; terminate when the maximum number of iterations is reached, and output the Bayesian network structure that reflects both data-driven causal priors and is optimal in the overall scoring sense.

5. The landslide risk dynamic early warning method based on Bayesian networks according to claim 4, characterized in that, Step S4: Based on maximum likelihood estimation and expert prior knowledge, a conditional probability table CPT is constructed using the training set to quantify the probability distribution of each node under the parent node's state. Specifically, this includes: Step S41: For each node in the training set and its parent node configuration Pa( Frequency statistics are performed on j, and Dirichlet prior parameters are introduced based on maximum likelihood estimation. compute nodes Given the posterior conditional probability of each value k under the parent node state j, construct the node's conditional probability table CPT. Step S42: When the number of new samples reaches a set threshold, while retaining the original Bayesian network structure and historical parameters, based on each node in the new samples... The frequency statistics of the configuration of the parent node, combined with the unified Dirichlet prior, are used to locally update the original conditional probability table CPT, thereby realizing the continuous correction of the parameters of the Bayesian network structure and the dynamic adaptation of the landslide disaster mode.

6. The landslide risk dynamic early warning method based on Bayesian networks according to claim 5, characterized in that, Step S5: Calculate the probability of landslide occurrence under a given combination of environmental factors through forward reasoning. The risk level of landslides is classified by combining geospatial units, and the contribution of each factor is calculated through reverse reasoning. Screening for the main disaster-causing factors, specifically including: Step S51: Forward Reasoning: Using the Bayesian network forward reasoning mechanism, input the combination of factors under the current or simulated scenario, and calculate the posterior probability of the landslide node Root. The landslide probability values ​​obtained through reasoning are mapped to spatial units, and risk levels are divided according to a set probability threshold to achieve geospatial zoning of landslide risk. Step S52: Reverse reasoning: Calculate the marginal probability change of each disaster-causing factor under the premise of landslide occurrence. It is used to identify the main controlling factors that have a significant impact on landslide risk.

7. A dynamic early warning system for landslide risk based on Bayesian networks, characterized in that, Includes the following modules: The module for constructing a discretized disaster-causing factor set is used to build a multidimensional disaster-causing factor dataset and obtain the multidimensional disaster-causing factor set {X1, X... 2, …,X sum }, where sum is the initial total number of disaster-causing factors. The multidimensional disaster-causing factor set undergoes multi-level screening and discretization to obtain a discretized disaster-causing factor set. , where n is the total number of disaster-causing factors after screening, and training and test sets are constructed proportionally; The module for obtaining the association rule set is used to apply the FP-Growth algorithm to the discretized set of disaster-causing factors. Perform frequent itemset extraction based on the preset minimum support S. min and minimum confidence level C min Generate a form like The set of high-confidence association rules, among which The disaster-causing factors are discrete; The module for constructing and optimizing the Bayesian network structure is used to set the landslide occurrence node Root as the root node and the discretized set of disaster-causing factors as child nodes; and to add directed edges between the preceding and succeeding terms according to the association rule set. An initial Bayesian network structure is constructed, and the network topology is optimized using the K2+ hill-climbing algorithm to obtain the final Bayesian network structure. The Bayesian network parameter learning module is used to construct a conditional probability table (CPT) based on the training set, using maximum likelihood estimation and expert prior knowledge, to quantify the probability distribution of each node under the state of the parent node. The landslide risk bidirectional reasoning and graded early warning module is used to calculate the probability of landslide occurrence under a given combination of environmental factors through forward reasoning. The risk level of landslides is classified by combining geospatial units, and the contribution of each factor is calculated through reverse reasoning. Screening for the main disaster-causing factors.

8. A dynamic early warning device for landslide risk based on Bayesian networks, characterized in that, It includes one or more electronic devices, wherein the one or more electronic devices are used to implement the method of any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1 to 6.

Images