An event probability prediction method, system and terminal under a tsunami disaster

Abstract

The application discloses a kind of event probability prediction method, system and terminal under tsunami disaster, the method includes: obtaining global tsunami historical data set, from which variable and variable corresponding initial data set are screened out, missing value is filled in the data set after discretization by Markov blanket node state value combination strategy, and complete data set is obtained;Complete data set is used to learn the structure of bayesian network, and the basic bayesian network structure is obtained;Causal edge candidate set is constructed, and the modified network structure is pruned to the basic network structure using causal edge candidate set, and the target network structure is obtained;The variable data under the tsunami disaster to be predicted is obtained, and the variable data is input into the target network structure, to obtain the probability of corresponding event occurrence under the tsunami disaster to be predicted.The application constructs the bayesian network of complete structure, can accurately predict the probability of event occurrence when tsunami disaster comes, so as to make correct emergency decision.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, system, terminal, and computer-readable storage medium for predicting the probability of events during tsunami disasters. Background Technology

[0002] In addressing uncertainty, academia has developed a series of uncertainty theories, including probability theory and evidence theory. The reasoning and decision-making abilities inherent in human cognitive behavior have become the ultimate goal of uncertainty-based artificial intelligence. Among the many uncertainty-related methods, probability theory is frequently used in research, and the emergence of Bayesian networks (BNs) has further increased the attention given to probabilistic methods in the field of artificial intelligence, providing a powerful tool for solving uncertainty-based reasoning problems.

[0003] Unlike computer vision, recommender systems, and natural language processing, which rely on big data, disaster emergency response suffers from randomness and infrequency. This means that the number of samples collected is far less than the amount of data needed to train deep learning models, especially historical samples from tsunami disasters. Secondly, tsunami emergency response requires rapid and accurate decision support. Generating data and training models using deep learning methods is often time-consuming and may not meet real-time requirements. Furthermore, the opaque nature of deep learning models, acting as "black boxes," contradicts the high demands for interpretability in disaster emergency management. Therefore, using Bayesian networks to construct tsunami emergency response models not only adapts to the characteristics of disaster data but also meets the needs of speed, accuracy, and transparency in emergency management.

[0004] However, existing disaster emergency response methods focus on learning Bayesian networks from complete data. When applied to disaster modeling, they face problems such as insufficient data volume, inconsistent data types, and missing data. Especially in the field of tsunami disasters, previous methods have even failed to construct structurally complete Bayesian networks, thus affecting emergency decision-making in tsunami disasters.

[0005] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0006] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for predicting the probability of tsunami events. This invention aims to address the problems faced by existing technologies that focus on learning Bayesian networks from complete data. When applied to tsunami disaster modeling, these technologies suffer from insufficient data volume, inconsistent data types, and missing data, making it impossible to construct a structurally complete Bayesian network. Consequently, they cannot accurately predict the probability of tsunami events, thus affecting emergency decision-making in the field of tsunami disasters.

[0007] To achieve the above objectives, the present invention provides a method for predicting the probability of tsunami events, which includes the following steps:

[0008] Obtain a global historical tsunami dataset, filter out variables and their corresponding initial datasets from the global historical tsunami dataset, and discretize the initial datasets corresponding to the continuous variables to obtain intermediate datasets;

[0009] The missing values ​​in the intermediate dataset are filled using a Markov blanket node state value combination strategy to obtain a complete dataset.

[0010] The complete dataset is used to learn the Bayesian network structure, and the basic Bayesian network structure is obtained.

[0011] Construct a causal edge candidate set, use the causal edge candidate set to modify the basic Bayesian network structure to obtain a modified network structure, and prune the modified network structure to obtain the target network structure;

[0012] Obtain variable data under the tsunami disaster to be predicted, input the variable data into the target network structure for calculation, and obtain the probability of the corresponding event occurring under the tsunami disaster to be predicted.

[0013] Optionally, in the event probability prediction method under the tsunami disaster, the variables include tsunami cause, earthquake magnitude, volcanic eruption index, region code, tsunami validity, number of tsunami impact points, number of casualties, monetary losses, damage to buildings, and maximum wave height.

[0014] The continuous variables among the variables include earthquake magnitude, maximum wave height, tsunami region, and number of tsunami impact points.

[0015] Optionally, the method for predicting the probability of tsunami disasters, wherein obtaining a global historical tsunami dataset, selecting variables and their corresponding initial datasets from the global historical tsunami dataset, and discretizing the initial datasets corresponding to continuous variables to obtain intermediate datasets, specifically includes:

[0016] Obtain a global historical tsunami dataset, and remove duplicate, irrelevant, and severely missing variables from the global historical tsunami dataset to obtain the variables and their corresponding initial datasets;

[0017] Based on domain knowledge and standards, the initial dataset corresponding to the continuous variables in the variables is discretized to obtain a discrete dataset;

[0018] The intermediate dataset is obtained by combining the discrete dataset and the remaining dataset in the initial dataset.

[0019] Optionally, the event probability prediction method under tsunami disasters, wherein the step of imputing missing values ​​in the intermediate dataset using a Markov blanket node state value combination strategy to obtain a complete dataset, specifically includes:

[0020] The intermediate dataset is divided into a set of completely random missing variables, a set of random missing variables, and a set of non-random missing variables.

[0021] The missing values ​​in the set of completely random missing variables are filled using any value other than the existing values, to obtain the first complete dataset;

[0022] The Fast-IAMB algorithm is used to learn the Markov blanket for each missing variable from the set of random missing variables and the set of non-random missing variables. Frequency statistics are performed on the state value combinations of each Markov blanket to obtain the frequency statistics results of the state value combinations of each Markov blanket.

[0023] Based on multiple frequency statistics results, the node with the highest frequency in each statistical result is obtained. The multiple nodes are used to impute missing values ​​for the missing variables corresponding to the random missing variable set and the non-random missing variable set, respectively, to obtain the second complete dataset.

[0024] The complete dataset is obtained by combining the first complete dataset and the second complete dataset.

[0025] Optionally, in the aforementioned method for predicting the probability of events under a tsunami disaster, the step of constructing a causal edge candidate set specifically includes:

[0026] A tsunami knowledge corpus is constructed, which is used to provide expert knowledge in the field of tsunamis and to assist in determining the direction of the edges between the variables;

[0027] Based on the expert knowledge in the tsunami knowledge corpus, several words related to the variable are obtained;

[0028] Web crawlers are used to collect information text from tsunami disaster web pages. The collected information text is preprocessed to obtain target text information. The target text information is then used to match several words to obtain the co-occurrence frequency and order of the words in the target text information.

[0029] An asymmetric matrix is ​​constructed based on the co-occurrence frequency and order of several words in the target text information. The frequency difference of the occurrence order of several words is quantified based on the asymmetric matrix. The causal relationship and direction between the variables are obtained based on the quantified frequency difference. A causal edge candidate set is constructed based on the causal relationship and the direction.

[0030] Optionally, in the aforementioned method for predicting the probability of events during a tsunami disaster, the step of modifying the basic Bayesian network structure using the causal edge candidate set to obtain a modified network structure specifically includes:

[0031] Obtain the missing variables in the basic Bayesian network structure, obtain the mutual information values ​​related to the missing variables, and sort the mutual information values ​​in reverse order to obtain a mutual information value sequence;

[0032] The variables in the mutual information value sequence are compared with the causal edge candidate set. If the comparison is successful, the variables in the mutual information value sequence are added as new edges to the basic Bayesian network structure to obtain the first modified network. The first modified network is checked to see if a loop is formed. If a loop is formed, the operation of adding new edges is canceled.

[0033] Based on the specific characteristics of tsunami disasters, time-sequence constraint rules are established for the variables, and edges in the first correction network that violate the constraint rules are corrected to obtain a second correction network.

[0034] Edges connecting two different subgraphs are selected from the candidate causal edge set as candidate edges, and the candidate edges are sorted in reverse order according to the mutual information value to obtain a candidate edge sequence.

[0035] The edges in the candidate edge sequence are compared with the causal edge candidate set. If the comparison is successful, the edges in the candidate edge sequence are added as new edges to the second correction network to obtain the correction network structure.

[0036] Optionally, in the aforementioned method for predicting the probability of events during a tsunami disaster, the step of pruning the modified network structure to obtain the target network structure specifically includes:

[0037] The modified network structure was pruned using the chi-square test, and the difference between the observed frequency and the expected frequency of each pair of variables in the modified network structure was calculated.

[0038] If the difference is higher than a preset threshold, the edge between the two variables is cut off to obtain the target network structure.

[0039] Furthermore, to achieve the above objectives, the present invention also provides an event probability prediction system under tsunami disasters, wherein the event probability prediction system under tsunami disasters includes:

[0040] The dataset acquisition module is used to acquire global tsunami historical datasets, filter out variables and their corresponding initial datasets from the global tsunami historical datasets, and discretize the initial datasets corresponding to the continuous variables to obtain intermediate datasets;

[0041] The data missing value imputation module is used to impute missing values ​​in the intermediate dataset using a Markov blanket node state value combination strategy to obtain a complete dataset.

[0042] The basic network construction module is used to learn the Bayesian network structure using the complete dataset to obtain the basic Bayesian network structure.

[0043] The network correction module is used to construct a causal edge candidate set, use the causal edge candidate set to correct the basic Bayesian network structure to obtain a corrected network structure, and prune the corrected network structure to obtain the target network structure.

[0044] The event probability prediction module is used to acquire variable data under the tsunami disaster to be predicted, input the variable data into the target network structure for calculation, and obtain the probability of the corresponding event occurring under the tsunami disaster to be predicted.

[0045] In addition, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an event probability prediction program for tsunami disasters stored in the memory and executable on the processor, wherein when the event probability prediction program for tsunami disasters is executed by the processor, the steps of the event probability prediction method for tsunami disasters as described above are implemented.

[0046] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an event probability prediction program under tsunami disaster, and when the event probability prediction program under tsunami disaster is executed by a processor, it implements the steps of the event probability prediction method under tsunami disaster as described above.

[0047] In this invention, a global historical tsunami dataset is acquired, and variables and their corresponding initial datasets are selected. A Markov blanket node state value combination strategy is used to impute missing values ​​in the discretized dataset, resulting in a complete dataset. The complete dataset is then used to learn a Bayesian network structure, yielding a basic Bayesian network structure. A causal edge candidate set is constructed and used to refine the basic network structure. This refined network structure is then pruned to obtain the target network structure. Variable data for the tsunami disaster to be predicted is acquired and input into the target network structure to obtain the probability of the corresponding event occurring under the predicted tsunami disaster. This invention constructs a structurally complete Bayesian network capable of accurately predicting the probability of tsunami events, thereby enabling correct emergency decisions. Attached Figure Description

[0048] Figure 1 This is a flowchart of a preferred embodiment of the event probability prediction method under tsunami disasters of the present invention;

[0049] Figure 2 This is a schematic diagram of the overall framework of the event probability prediction method under tsunami disasters of the present invention;

[0050] Figure 3 This is a flowchart of the modification of the basic Bayesian network structure in the event probability prediction method under tsunami disaster of the present invention;

[0051] Figure 4 This is a schematic diagram of the time sequence constraint rules in the event probability prediction method under tsunami disaster of the present invention;

[0052] Figure 5 This is a schematic diagram of the missing value imputation results of the dataset generated by the Chid model in the event probability prediction method under tsunami disaster of the present invention;

[0053] Figure 6 This is a schematic diagram of the missing value imputation results of the dataset generated by the Sachs model in the event probability prediction method under tsunami disaster of this invention;

[0054] Figure 7 This is a schematic diagram of the missing value imputation results of the dataset generated by the Manual-BN model in the event probability prediction method under tsunami disaster of the present invention;

[0055] Figure 8 This is a schematic diagram illustrating the structural correction of the BN learned by NOTEARS in the event probability prediction method under tsunami disasters of this invention.

[0056] Figure 9 This is a comparison chart before and after the knowledge correction is incorporated into the hill climbing method based on BIC score in the event probability prediction method under tsunami disaster of this invention;

[0057] Figure 10 This is a comparison chart before and after knowledge correction was incorporated into the PC-parallel method in the event probability prediction method under tsunami disasters of this invention;

[0058] Figure 11 This is a comparison chart before and after the DAG-GNN method was incorporated into the knowledge correction in the event probability prediction method under tsunami disaster of this invention;

[0059] Figure 12 This is a leaf node frequency distribution diagram before and after structural correction in the event probability prediction method under tsunami disaster of this invention;

[0060] Figure 13 This is a schematic diagram of a preferred embodiment of the event probability prediction system under tsunami disasters of the present invention;

[0061] Figure 14 This is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0062] This application provides a method and related equipment for predicting the probability of tsunami events. To make the purpose, technical solution, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0063] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0064] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0065] The preferred embodiment of the present invention describes a method for predicting the probability of tsunami disasters, such as... Figure 1 and Figure 2As shown, the event probability prediction method under the tsunami disaster includes the following steps:

[0066] Step S10: Obtain the global tsunami historical dataset, filter out the variables and their corresponding initial datasets from the global tsunami historical dataset, and discretize the initial datasets corresponding to the continuous variables to obtain intermediate datasets.

[0067] In the field of disaster emergency planning, Bayesian network models are commonly used. These models allow relevant personnel to infer the probability of different events, such as casualties and economic losses caused by disasters, through indirect graphical representations and probability distributions, thereby improving emergency planning decisions. The most crucial step in learning a Bayesian network model is structure learning. For historically infrequent disasters like tsunamis, historical data samples often contain significant missing data, and the data types of each variable are inconsistent. Therefore, data preprocessing should be prioritized. Then, to address the issue of limited sample size, knowledge from the tsunami disaster domain is incorporated to compensate for the data deficiencies and refine the Bayesian network structure.

[0068] Specifically, a global tsunami historical dataset is obtained. The global tsunami historical dataset is derived from the publicly available global tsunami dataset provided by the National Oceanic and Atmospheric Administration (NOAA) website. It contains dozens of variables. This invention filters out variables from the global tsunami historical dataset, removing duplicate, irrelevant, and severely missing variables to obtain the variables and their corresponding initial datasets.

[0069] Ten variables were selected, including Tsunami Cause (TCC), Earthquake Magnitude (EM), Volcanic Eruption Index (VEI), Region Code, Tsunami Validity (TEV), Number of Tsunami Runups, Casualty, Loss, Damaged Buildings (HD), and Maximum Wave Height. The initial dataset was obtained by collecting data corresponding to these variables from a global historical tsunami dataset.

[0070] It is understood that the variables include both discontinuous and continuous variables. The continuous variables include earthquake magnitude, maximum wave height, tsunami region, and number of tsunami impact points, while the remaining variables are all discontinuous variables.

[0071] Furthermore, the initial dataset corresponding to the continuous variables in the variables is discretized according to domain knowledge and standards to obtain a discrete dataset. Specifically, earthquake magnitude is divided into five state values ​​based on 4.5, 6, 7, and 8: felt earthquake, moderate earthquake, strong earthquake, major earthquake, and great earthquake. Maximum wave height is divided into six state values ​​based on national standards for tsunami intensity, with boundaries of 0.3, 1, 3, 10, and 20 meters. The tsunami region codes span from 30 to 89, resulting in too many state values ​​as a single variable. The NOAA website indicates that the global distribution of all events in the database is 70% Pacific Ocean, 15% Mediterranean Ocean, 9% Caribbean and Atlantic Ocean, and 6% Indian Ocean. Therefore, the region codes are divided into four regions: Pacific Ocean, Mediterranean Ocean and surrounding Europe, Caribbean Ocean and Atlantic Ocean, and Indian Ocean.

[0072] It should be noted that, for the discretization of the number of tsunami impact points, since there is no domain knowledge available for reference, the quartile discretization method was used, dividing the data into four intervals based on the 25% values ​​of the data sorted from smallest to largest, namely 1, 2, and 6.

[0073] After discretizing the four continuous variables to obtain a discrete dataset, the intermediate dataset is obtained by combining the discrete dataset with the remaining datasets (datasets corresponding to discontinuous variables) in the initial dataset.

[0074] Step S20: Impute missing values ​​in the intermediate dataset using a Markov blanket node state value combination strategy to obtain a complete dataset.

[0075] Specifically, there are generally three situations where variables are missing: completely random missing (MCAR), randomly missing (MAR), and non-random missing (MNAR). In this invention, the intermediate dataset is divided into a set of completely random missing variables, a set of randomly missing variables, and a set of non-random missing variables.

[0076] The missing values ​​in the set of completely random missing variables are filled using any value other than the existing values, to obtain the first complete dataset.

[0077] Understandably, completely random missing values ​​mean that the missing values ​​are not related to any other variables. In other words, the occurrence of completely random missing values ​​is random, but they may be correlated with other non-missing variables. When the variable is MCAR, a non-existent state value can be used to fill the missing value. Under this data, the missing values ​​of earthquake magnitude and volcanic eruption index are assumed to have not occurred, and a value other than the existing state values ​​is used to fill the missing value.

[0078] Furthermore, the Fast-IAMB algorithm is used to learn the Markov blanket for each missing variable from the set of random missing variables and the set of non-random missing variables. Frequency statistics are performed on the state value combinations of each Markov blanket to obtain the frequency statistics results of the state value combinations of each Markov blanket.

[0079] Specifically, for the remaining sets of randomly missing variables and non-randomly missing variables, this invention employs a strategy based on the combination of Markov blanket node state values. Under the fidelity assumption, all variables (nodes) belonging to the specified Markov blanket (MB) are strongly correlated with that specified variable, while non-MB variable nodes are considered redundant features of the variable and should be deleted to optimize the missing value imputation process. Specifically, the Fast-IAMB algorithm is first used to learn the MB of the target variable from the remaining complete data, and frequency statistics are performed on each unique combination of MB state values ​​in the dataset. In cases where Fast-IAMB fails to learn the MB, the node with the largest mutual information value with the target node is used as the unique node in the MB, as shown in the following formula:

[0080]

[0081]

[0082] Here, X and Y represent two different variables. MI(X,Y) obtains the mutual information between variables X and Y. x represents a state value of variable X, y represents a state value of variable Y, P(x,y) represents the joint probability distribution of variables x and y, P(x) represents the marginal probability distribution of variable x, P(y) represents the marginal probability distribution of variable y, V represents all variables including X and Y, and mb is the Markov blanket variable that uses the variable Y with the largest mutual information value with variable X when the Fast-IAMB algorithm cannot learn MB.

[0083] Furthermore, based on multiple frequency statistics, the node with the highest frequency in each statistical result is obtained, and the missing values ​​of the missing variables corresponding to the random missing variable set and the non-random missing variable set are filled using multiple nodes respectively to obtain the second complete dataset; the first complete dataset and the second complete dataset are combined to obtain the complete dataset.

[0084] Understandably, after obtaining the frequency statistics, for each missing target node value, based on its corresponding MB state value combination, the non-missing target node value with the highest occurrence frequency is selected as the estimated value. In the case of a perfect match (i.e., when TMB and MMB are equal), the formula is as follows:

[0085]

[0086]

[0087] Where TMB is the state value combination of the Markov blanket when the target variable X is not missing, MMB is the state value combination of the Markov blanket when the target variable X is missing, and S X It is the set of all possible state values ​​of X, and Count() is a counting function representing the total number of samples that satisfy the conditions in parentheses. i S represents X The i-th state value in the set, where i represents the ordinal number of the state value, f(x) i |MMB) represents the condition where the dependent variable is x, given the MMB. i The function estimate is obtained. The above formula aims to find the optimal state value x∈S given MMB. X This ensures that x appears most frequently when TMB and MMB are completely consistent; this strategy implicitly applies the idea of ​​maximum likelihood estimation, thereby ensuring the accuracy and robustness of the filling results.

[0088] Furthermore, in the absence of a perfect match (i.e., when TMB and MMB are not equal), the most similar state value combination is selected, as shown in the following formula, where ADS represents the Absolute Deviation Score:

[0089]

[0090]

[0091] Where n represents the number of Markov blanket variables of the target variable, and ADS represents the absolute deviation score. For example, if the Markov blanket variables of X are [Y, Z, W], then n = 3, TMB = [1, 3, 2], MMB = [2, 3, 1], then ADS = 2.

[0092] Step S30: Use the complete dataset to learn the Bayesian network structure and obtain the basic Bayesian network structure.

[0093] Specifically, using the existing complete dataset to learn the Bayesian network algorithm, the Bayesian network structure is learned on the complete dataset. It is found that the basic Bayesian network structure has a number of problems, such as missing some nodes, some edges with uncertain directions, multiple subgraphs being disconnected, and too many edges leading to structural complexity.

[0094] Step S40: Construct a causal edge candidate set, use the causal edge candidate set to modify the basic Bayesian network structure to obtain a modified network structure, and prune the modified network structure to obtain the target network structure.

[0095] It is understandable that, due to the various problems mentioned above in the basic Bayesian network structure, such as missing nodes, undetermined directions of some edges, disconnected subgraphs, and excessive edges leading to structural complexity, this invention constructs a causal edge candidate set and uses the causal edge candidate set to correct the basic Bayesian network structure.

[0096] Specifically, the process of constructing a candidate set of causal edges involves first constructing a tsunami knowledge corpus. This tsunami knowledge corpus provides expert knowledge in the tsunami domain and assists in determining the direction of edges between the variables. Based on the expert knowledge in the tsunami knowledge corpus, several words related to the variables are obtained.

[0097] Understandably, based on expert knowledge in the field of tsunami research, words related to the meanings of the 10 variables in the dataset are identified. Some of these variables may correspond to more than one word, and for ease of data collection, some words are represented by prefix letters, thus encompassing several word forms or synonyms.

[0098] Furthermore, a web crawler is used to collect information text from tsunami disaster web pages. The collected information text is preprocessed to obtain target text information. The target text information is then used to match several words to obtain the co-occurrence frequency and order of the words in the target text information.

[0099] An asymmetric matrix is ​​constructed based on the co-occurrence frequency and order of several words in the target text information. The frequency difference of the occurrence order of several words is quantified based on the asymmetric matrix. The causal relationship and direction between the variables are obtained based on the quantified frequency difference. A causal edge candidate set is constructed based on the causal relationship and the direction.

[0100] Understandably, the process of constructing a causal edge candidate set based on word co-occurrence frequency and their order in the text rests on a linguistic assumption: in narrative structures, causes are typically mentioned before effects. In the preprocessed text, the co-occurrence of all variable-related words is recorded and stored in an asymmetric matrix. The x and y axes of this matrix correspond to the ten variables in the tsunami dataset, and each element [i, j] records the number of instances where the related word for variable i appears before the related word for variable j. By comparing elements [i, j] and [j, i], the frequency differences in the order of related word occurrences can be quantified, thereby inferring the causal relationship and its directionality between variables.

[0101] As shown in the formula below: Where CM[j,i] is the co-occurrence matrix of variable j appearing before variable i, CM[i,j] is the co-occurrence matrix of variable i appearing before variable j, ω is the preset threshold, and V i →V j Let i and j represent variables, and this formula represents V. i and V j The number of times these two variables co-occur must be greater than a certain threshold to represent V. i and V j The nodes corresponding to the variables i and j are related. This method yields a candidate set of causal edges for subsequent structural correction tasks.

[0102] Furthermore, such as Figure 3 As shown, the step of using the causal edge candidate set to modify the basic Bayesian network structure to obtain a modified network structure specifically includes:

[0103] Step S401: Obtain the missing variables in the basic Bayesian network structure, obtain the mutual information values ​​related to the missing variables, and sort the mutual information values ​​in reverse order to obtain a mutual information value sequence.

[0104] Step S402: Compare the variables in the mutual information value sequence with the causal edge candidate set. If the comparison is successful, add the variables in the mutual information value sequence as new edges to the basic Bayesian network structure to obtain the first modified network. Check whether the first modified network forms a loop. If a loop is formed, cancel the operation of adding new edges.

[0105] Understandably, to address the issue of missing nodes, this invention obtains a sequence of mutual information values. The variables in this sequence are then checked against a causal edge candidate set to determine if an edge exists between two variables, and this edge is added to the original BN structure. Considering the possibility of multiple missing nodes, newly added edges might connect these nodes, potentially leading to loops. Therefore, after each new edge is added, the network structure is checked to see if a loop is formed. If a loop is formed, the edge addition operation is revoked, thus completing the missing node replacement. The specific algorithm for completing the missing node replacement is shown in Table 1 below.

[0106] Table 1. Algorithm for Completing Missing Nodes

[0107]

[0108] Step S403: Based on the special characteristics of the tsunami disaster field, time sequence constraint rules are formulated for the variables, and the edges in the first correction network that violate the constraint rules are corrected to obtain the second correction network.

[0109] Specifically, based on the temporal constraint that later events cannot lead to earlier events, this invention, taking into account the particularities of the tsunami disaster field, formulates the following temporal constraint rules to more accurately reflect the causal relationships between events: Rule 1: Tsunami Cause (TCC) precedes Tsunami Effectiveness (TEV); Rule 2: Detected wave height precedes the number of tsunami runups; Rule 3: The three variables representing disaster losses—casualty, loss, and housing damage—follow the events of the other variables.

[0110] Based on the above rules, a hierarchical structure of constraints is constructed, such as... Figure 4 As shown, rules 1 and 2 are used to identify and correct paths that violate constraints in the graph, typically achieved by reversing each edge on the path. For rule 3, it first checks if the node representing the loss is a leaf node; if not, only the other two loss nodes are allowed as its children; otherwise, the child and parent nodes are reversed. These rules not only ensure the temporal logical consistency of the Bayesian network structure but also enhance its ability to explain causal relationships specific to the tsunami disaster domain. In subsequent improvements to the BN connectivity method, temporal constraints also provide a certain direction-determining function.

[0111] Step S404: Select the edges connecting two different subgraphs from the causal edge candidate set as candidate edges, and sort the candidate edges in reverse order according to the mutual information value to obtain the candidate edge sequence.

[0112] Understandably, to address the issue of multiple disconnected subgraphs, following the approach of completing missing nodes mentioned above, edges are added based on mutual information values ​​and a candidate set of causal edges to achieve network connectivity. Specifically, firstly, edges connecting two different subgraphs are identified from the candidate set of causal edges as candidate edges. Then, these candidate edges are sorted in reverse order based on their mutual information values ​​to obtain a candidate edge sequence.

[0113] Step S405: Compare the edges in the candidate edge sequence with the causal edge candidate set. If the comparison is successful, add the edges in the candidate edge sequence as new edges to the second correction network to obtain the correction network structure.

[0114] Furthermore, if no suitable candidate edge is found in the causal edge candidate set, time series structure constraints and BIC scoring are introduced to help determine the direction of the edge. This method comprehensively utilizes quantitative analysis and domain knowledge to ensure that the final BN structure is not only technically connected but also logically sound, more closely reflecting the complexity of actual tsunami disasters.

[0115] Furthermore, the modified network structure is pruned to obtain the target network structure, specifically including:

[0116] The modified network structure is pruned using the chi-square test. The difference between the observed frequency and the expected frequency of each pair of variables in the modified network structure is calculated. If the difference is higher than a preset threshold, the edge between the two variables is pruned to obtain the target network structure.

[0117] To optimize complex BN structures and improve inference efficiency, a chi-square test was used for pruning. If the chi-square test value between two variables... 2 If the correlation is below a given threshold, it is considered statistically significant; otherwise, they are considered independent, and the edges between them are removed.

[0118]

[0119]

[0120] Where, χ 2 It is the chi-square statistic, used to quantify the difference between observed frequencies and expected frequencies; O ab This refers to the observation frequencies in row a and column b of a contingency table (a contingency table is a table obtained by cross-counting the different state values ​​of each pair of variables), i.e., the actual number of times a specific category combination occurs in the data; E ab This is the expected frequency of the contingency table, calculated based on the assumption that the two variables are independent, n a It is the marginal frequency of row a in the contingency table, n b is the marginal frequency of column b, and n is the total sample size, which is the sum of the frequencies of all cells in the contingency table.

[0121] Step S50: Obtain variable data under the tsunami disaster to be predicted, input the variable data into the target network structure for calculation, and obtain the probability of the corresponding event occurring under the tsunami disaster to be predicted.

[0122] After modifying the basic Bayesian network structure, the resulting target network structure can accurately predict the probability of the corresponding event occurring under the predicted tsunami disaster. Specifically, the variable data under the predicted tsunami disaster (data corresponding to ten variables under this predicted tsunami disaster) is obtained, and the variable data is input into the target network structure for calculation to obtain the probability of the corresponding event occurring under the predicted tsunami disaster. Using the probability as a basis, better emergency planning decisions can be made.

[0123] Furthermore, the present invention provides corresponding calculation results to verify the beneficial effects of the present invention.

[0124] First, to verify the missing value imputation effect, this invention introduces Alarm, Child, Sachs, and Asia from the BN standard library, as well as a manually constructed Manual-BN Bayesian network model to generate corresponding samples for experimental verification of the discrete missing value imputation method. The datasets are shown in the table below. To simulate data imputation with small sample sizes as closely as possible, data with sample sizes of 400, 600, 800, 1000, 1200, and 1400 were generated from these five models. 35% of the data was randomly extracted from each node in turn. The prediction effect of each round of sample size was obtained by averaging the prediction metrics of each variable in the model. The comparison methods included K-Nearest Neighbor Imputation (KNN), Naive Bayes Classification (NBC), and Gradient Boosting Classifier (GBDT) from the commonly used machine learning library Scikit-learn, as well as the Expectation-Maximization (EM) algorithm commonly used in BN, and the increasingly popular Chain Equation Multiple Imputation (MICE) method. The following shows a line graph comparison of Child, Sachs, and Manual-BN, representing the Child model (e.g., ...). Figure 5 As shown), Sachs model (such as...) Figure 6 As shown), Manual-B model (as shown) Figure 7 The image shows the missing value imputation results for the generated dataset, referencing four metrics: Accuracy, F1 score, Recall, and Time. The magenta line (MBI) represents the missing value imputation method. From left to right, the first three metrics are better the higher they are, and the last one, Time, is better the lower it is.

[0125] Combination Figure 5 , Figure 6 and Figure 7 It is evident that the MBI algorithm consistently outperforms other methods in accuracy and maintains an advantage in recall and F1 score across most sample specifications, while being slightly less efficient in terms of time compared to the two less accurate methods.

[0126] Secondly, the effect of modifying the basic Bayesian network structure (incorporating a knowledge-modified network) was verified. As shown in Table 2, this invention presents experimental results using six evaluation metrics to evaluate the BN structure. Each metric is divided into two columns, representing the BN metric values ​​before and after knowledge incorporation. All parameters of the basic algorithm used default parameters or the optimal parameters provided in the paper. During chi-square pruning, χ²... 2The value is set to 0.05. The BIC score in the experiment was calculated based on 257 samples with no missing data in the original dataset. Furthermore, the BIC score calculation formula shows that a reduction in the number of data samples often leads to higher scores for simpler BN structures. Therefore, in tsunami disaster research with limited sample data, the BIC score should not be used as the sole evaluation criterion. Multiple indicators should be considered comprehensively to evaluate and select the optimal structure of the Bayesian network. The total number of edges in a BN intuitively reflects its complexity and is crucial for understanding the overall complexity of the network structure; the average in-degree indicates the average number of parent nodes per node, helping to understand how many other variables influence each variable on average; depth refers to the length of the longest causal chain in the network, reflecting the maximum number of steps required to go from cause to effect in a BN; leaf nodes typically represent observed results, such as the specific impact of the tsunami, and their number is important for understanding the network's output layer; the Markov blanket size reflects the local complexity of each node in the network and directly affects the efficiency of inference.

[0127] Table 1 Comparison before and after knowledge integration and revision

[0128]

[0129] The table reveals potential shortcomings of BN structures before integrating prior knowledge. When the average in-degree is not equal to the number of edges divided by the number of variables, it suggests the possible loss of variable nodes in the BN structure. Furthermore, the case where the number of leaf nodes is zero indicates the presence of cycles in the BN structure, violating the principle that BNs must be DAGs. These findings demonstrate the necessity and importance of considering domain knowledge during model construction.

[0130] By comparing the metric values ​​before and after knowledge integration, significant changes were observed after structural optimization. Even after considering measures such as supplementing missing nodes and connecting disconnected subgraphs, the median number of edges in the network decreased from 12 to 11, demonstrating the effectiveness of structural simplification. Simultaneously, the average Markov blanket size also decreased, reflecting the refinement of dependencies between variables. Notably, the longest causal chain length stabilized at around 5 after optimization, indicating moderate control of network complexity. Furthermore, the average number of leaf nodes increased from 2.7 to 3, with a limited increase and stable distribution, without observing extreme fluctuations, further confirming the robustness of the network structure after optimization. These results collectively point to the effectiveness of the adopted optimization strategy in improving the rationality and interpretability of the model structure, providing guidance for the construction of future tsunami BN models.

[0131] like Figure 8 As shown, Figure 8The example demonstrates the specific changes at each step of the structural correction process, including filling in missing nodes, time sequence constraints, and connecting multiple subgraphs, using NOTEARS as an example. The changes to specific nodes and edges are marked in green.

[0132] like Figure 9 , Figure 10 and Figure 11 As shown below, using several basic algorithms' learned BN as examples, we demonstrate the BN structure after incorporating knowledge and making corrections. The red markers indicate leaf nodes, and the yellow-marked paths are one of the longest causal chains.

[0133] Before knowledge-based correction, Region, Height, Casualty, and TCC appeared as leaf nodes at least four times in the Bayesian network. This reflects that BN structures obtained solely from data often treat these four variables as the final outcome variables. However, it is clearly inappropriate to use TCC (cause of the tsunami) as the outcome of the tsunami emergency BN. Therefore, after knowledge-based correction, the frequency of TCC dropped to the end, and more BN structures considered Region, TEV, Casualty, and Loss to be the final node variables for the tsunami emergency BN. Furthermore, the former showed one instance of EM being treated as a leaf node but no instance of HD, while the latter never showed EM but had HD as a leaf node twice, further demonstrating the effectiveness of the structural correction. Specific data is as follows... Figure 12 (a) and Figure 12 As shown in (b) of the diagram.

[0134] As can be seen, this invention provides a method for predicting the probability of tsunami events. It involves acquiring a global historical tsunami dataset, selecting variables and their corresponding initial datasets, imputing missing values ​​in the discretized dataset using a Markov blanket node state value combination strategy to obtain a complete dataset, learning a Bayesian network structure using the complete dataset to obtain a basic Bayesian network structure, constructing a causal edge candidate set, using this set to modify the basic network structure, pruning the modified network structure to obtain the target network structure, acquiring variable data for the tsunami event to be predicted, inputting the variable data into the target network structure to obtain the probability of the corresponding event occurring under the predicted tsunami event. The proposed Bayesian network construction framework for tsunami disasters effectively solves the problem of missing disaster data and integrates domain knowledge to correct problems such as missing nodes, structural loops, and disconnections in the Bayesian network structure. This allows the Bayesian network for tsunami disasters to be better applied to downstream emergency planning and decision-making tasks, and also provides a new solution for emergency planning modeling of other disasters.

[0135] Furthermore, such as Figure 13As shown, based on the above-mentioned event probability prediction method under tsunami disasters, the present invention also provides an event probability prediction system under tsunami disasters, wherein the event probability prediction system under tsunami disasters includes:

[0136] The dataset acquisition module 51 is used to acquire a global tsunami historical dataset, filter out variables and their corresponding initial datasets from the global tsunami historical dataset, and discretize the initial datasets corresponding to the continuous variables to obtain an intermediate dataset;

[0137] The data missing value filling module 52 is used to fill in the missing values ​​of the intermediate dataset using a Markov blanket node state value combination strategy to obtain a complete dataset.

[0138] The basic network construction module 53 is used to learn the Bayesian network structure using the complete dataset to obtain the basic Bayesian network structure.

[0139] The network correction module 54 is used to construct a causal edge candidate set, use the causal edge candidate set to correct the basic Bayesian network structure to obtain a corrected network structure, and prune the corrected network structure to obtain the target network structure.

[0140] The event probability prediction module 55 is used to acquire variable data under the tsunami disaster to be predicted, input the variable data into the target network structure for calculation, and obtain the probability of the corresponding event occurring under the tsunami disaster to be predicted.

[0141] Furthermore, such as Figure 14 As shown, based on the above-mentioned event probability prediction method and system under tsunami disaster, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 14 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0142] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a tsunami event probability prediction program 40, which can be executed by the processor 10 to implement the tsunami event probability prediction method of this application.

[0143] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the event probability prediction method under the tsunami disaster.

[0144] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0145] In one embodiment, when processor 10 executes event probability prediction program 40 under tsunami disaster in memory 20, the following steps are performed:

[0146] Obtain a global historical tsunami dataset, filter out variables and their corresponding initial datasets from the global historical tsunami dataset, and discretize the initial datasets corresponding to the continuous variables to obtain intermediate datasets;

[0147] The missing values ​​in the intermediate dataset are filled using a Markov blanket node state value combination strategy to obtain a complete dataset.

[0148] The complete dataset is used to learn the Bayesian network structure, and the basic Bayesian network structure is obtained.

[0149] Construct a causal edge candidate set, use the causal edge candidate set to modify the basic Bayesian network structure to obtain a modified network structure, and prune the modified network structure to obtain the target network structure;

[0150] Obtain variable data under the tsunami disaster to be predicted, input the variable data into the target network structure for calculation, and obtain the probability of the corresponding event occurring under the tsunami disaster to be predicted.

[0151] The variables include the cause of the tsunami, earthquake magnitude, volcanic eruption index, region code, tsunami validity, number of tsunami impact points, number of casualties, monetary losses, damage to buildings, and maximum wave height.

[0152] The continuous variables among the variables include earthquake magnitude, maximum wave height, tsunami region, and number of tsunami impact points.

[0153] The step of obtaining a global tsunami historical dataset, selecting variables and their corresponding initial datasets from the global tsunami historical dataset, and discretizing the initial datasets corresponding to the continuous variables to obtain intermediate datasets, specifically includes:

[0154] Obtain a global historical tsunami dataset, and remove duplicate, irrelevant, and severely missing variables from the global historical tsunami dataset to obtain the variables and their corresponding initial datasets;

[0155] Based on domain knowledge and standards, the initial dataset corresponding to the continuous variables in the variables is discretized to obtain a discrete dataset;

[0156] The intermediate dataset is obtained by combining the discrete dataset and the remaining dataset in the initial dataset.

[0157] Specifically, the step of imputing missing values ​​in the intermediate dataset using a Markov blanket node state value combination strategy to obtain a complete dataset includes:

[0158] The intermediate dataset is divided into a set of completely random missing variables, a set of random missing variables, and a set of non-random missing variables.

[0159] The missing values ​​in the set of completely random missing variables are filled using any value other than the existing values, to obtain the first complete dataset;

[0160] The Fast-IAMB algorithm is used to learn the Markov blanket for each missing variable from the set of random missing variables and the set of non-random missing variables. Frequency statistics are performed on the state value combinations of each Markov blanket to obtain the frequency statistics results of the state value combinations of each Markov blanket.

[0161] Based on multiple frequency statistics results, the node with the highest frequency in each statistical result is obtained. The multiple nodes are used to impute missing values ​​for the missing variables corresponding to the random missing variable set and the non-random missing variable set, respectively, to obtain the second complete dataset.

[0162] The complete dataset is obtained by combining the first complete dataset and the second complete dataset.

[0163] The construction of the causal edge candidate set specifically includes:

[0164] A tsunami knowledge corpus is constructed, which is used to provide expert knowledge in the field of tsunamis and to assist in determining the direction of the edges between the variables;

[0165] Based on the expert knowledge in the tsunami knowledge corpus, several words related to the variable are obtained;

[0166] Web crawlers are used to collect information text from tsunami disaster web pages. The collected information text is preprocessed to obtain target text information. The target text information is then used to match several words to obtain the co-occurrence frequency and order of the words in the target text information.

[0167] An asymmetric matrix is ​​constructed based on the co-occurrence frequency and order of several words in the target text information. The frequency difference of the occurrence order of several words is quantified based on the asymmetric matrix. The causal relationship and direction between the variables are obtained based on the quantified frequency difference. A causal edge candidate set is constructed based on the causal relationship and the direction.

[0168] Specifically, the step of modifying the basic Bayesian network structure using the causal edge candidate set to obtain a modified network structure includes:

[0169] Obtain the missing variables in the basic Bayesian network structure, obtain the mutual information values ​​related to the missing variables, and sort the mutual information values ​​in reverse order to obtain a mutual information value sequence;

[0170] The variables in the mutual information value sequence are compared with the causal edge candidate set. If the comparison is successful, the variables in the mutual information value sequence are added as new edges to the basic Bayesian network structure to obtain the first modified network. The first modified network is checked to see if a loop is formed. If a loop is formed, the operation of adding new edges is canceled.

[0171] Based on the specific characteristics of tsunami disasters, time-sequence constraint rules are established for the variables, and edges in the first correction network that violate the constraint rules are corrected to obtain a second correction network.

[0172] Edges connecting two different subgraphs are selected from the candidate causal edge set as candidate edges, and the candidate edges are sorted in reverse order according to the mutual information value to obtain a candidate edge sequence.

[0173] The edges in the candidate edge sequence are compared with the causal edge candidate set. If the comparison is successful, the edges in the candidate edge sequence are added as new edges to the second correction network to obtain the correction network structure.

[0174] Specifically, pruning the modified network structure to obtain the target network structure includes:

[0175] The modified network structure was pruned using the chi-square test, and the difference between the observed frequency and the expected frequency of each pair of variables in the modified network structure was calculated.

[0176] If the difference is higher than a preset threshold, the edge between the two variables is cut off to obtain the target network structure.

[0177] In summary, this invention provides a method and related equipment for predicting the probability of tsunami events. The method includes: acquiring a global historical tsunami dataset; selecting variables and their corresponding initial datasets from the dataset; imputing missing values ​​in the discretized dataset using a Markov blanket node state value combination strategy to obtain a complete dataset; learning a Bayesian network structure using the complete dataset to obtain a basic Bayesian network structure; constructing a causal edge candidate set; using the causal edge candidate set to modify the basic network structure; pruning the modified network structure to obtain a target network structure; acquiring variable data for the tsunami disaster to be predicted; inputting the variable data into the target network structure to obtain the probability of the corresponding event occurring under the tsunami disaster to be predicted. This invention constructs a structurally complete Bayesian network that can accurately predict the probability of tsunami events, thereby enabling correct emergency decisions.

[0178] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0179] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0180] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for predicting the probability of tsunami events, characterized in that, The event probability prediction method under the tsunami disaster includes: Obtain a global historical tsunami dataset, filter out variables and their corresponding initial datasets from the global historical tsunami dataset, and discretize the initial datasets corresponding to the continuous variables to obtain intermediate datasets; The missing values ​​in the intermediate dataset are filled using a Markov blanket node state value combination strategy to obtain a complete dataset. The complete dataset is used to learn the Bayesian network structure, and the basic Bayesian network structure is obtained. Construct a causal edge candidate set, use the causal edge candidate set to modify the basic Bayesian network structure to obtain a modified network structure, and prune the modified network structure to obtain the target network structure; Obtain variable data under the tsunami disaster to be predicted, input the variable data into the target network structure for calculation, and obtain the probability of the corresponding event occurring under the tsunami disaster to be predicted; The process of imputing missing values ​​in the intermediate dataset using a Markov blanket node state value combination strategy to obtain a complete dataset specifically includes: The intermediate dataset is divided into a set of completely random missing variables, a set of random missing variables, and a set of non-random missing variables. The missing values ​​in the set of completely random missing variables are filled using any value other than the existing values, to obtain the first complete dataset; The Fast-IAMB algorithm is used to learn the Markov blanket for each missing variable from the set of random missing variables and the set of non-random missing variables. Frequency statistics are performed on the state value combinations of each Markov blanket to obtain the frequency statistics results of the state value combinations of each Markov blanket. Based on multiple frequency statistics results, the node with the highest frequency in each statistical result is obtained. The multiple nodes are used to impute missing values ​​for the missing variables corresponding to the random missing variable set and the non-random missing variable set, respectively, to obtain the second complete dataset. The complete dataset is obtained by combining the first complete dataset and the second complete dataset.

2. The event probability prediction method under tsunami disaster according to claim 1, characterized in that, The variables include the cause of the tsunami, earthquake magnitude, volcanic eruption index, region code, tsunami validity, number of tsunami impact points, number of casualties, monetary losses, damage to buildings, and maximum wave height. The continuous variables among the variables include earthquake magnitude, maximum wave height, tsunami region, and number of tsunami impact points.

3. The event probability prediction method under tsunami disaster according to claim 2, characterized in that, The process of obtaining a global tsunami historical dataset, selecting variables and their corresponding initial datasets from the global tsunami historical dataset, and discretizing the initial datasets corresponding to the continuous variables to obtain intermediate datasets, specifically includes: Obtain a global historical tsunami dataset, and remove duplicate, irrelevant, and severely missing variables from the global historical tsunami dataset to obtain the variables and their corresponding initial datasets; Based on domain knowledge and standards, the initial dataset corresponding to the continuous variables in the variables is discretized to obtain a discrete dataset; The intermediate dataset is obtained by combining the discrete dataset and the remaining dataset in the initial dataset.

4. The event probability prediction method under tsunami disaster according to claim 1, characterized in that, The construction of the causal edge candidate set specifically includes: A tsunami knowledge corpus is constructed, which is used to provide expert knowledge in the field of tsunamis and to assist in determining the direction of the edges between the variables; Based on the expert knowledge in the tsunami knowledge corpus, several words related to the variable are obtained; Web crawlers are used to collect information text from tsunami disaster web pages. The collected information text is preprocessed to obtain target text information. The target text information is then used to match several words to obtain the co-occurrence frequency and order of the words in the target text information. An asymmetric matrix is ​​constructed based on the co-occurrence frequency and order of several words in the target text information. The frequency difference of the occurrence order of several words is quantified based on the asymmetric matrix. The causal relationship and direction between the variables are obtained based on the quantified frequency difference. A causal edge candidate set is constructed based on the causal relationship and the direction.

5. The event probability prediction method under tsunami disaster according to claim 4, characterized in that, The step of modifying the basic Bayesian network structure using the causal edge candidate set to obtain a modified network structure specifically includes: Obtain the missing variables in the basic Bayesian network structure, obtain the mutual information values ​​related to the missing variables, and sort the mutual information values ​​in reverse order to obtain a mutual information value sequence; The variables in the mutual information value sequence are compared with the causal edge candidate set. If the comparison is successful, the variables in the mutual information value sequence are added as new edges to the basic Bayesian network structure to obtain the first modified network. The first modified network is checked to see if a loop is formed. If a loop is formed, the operation of adding new edges is canceled. Based on the specific characteristics of tsunami disasters, time-sequence constraint rules are established for the variables, and edges in the first correction network that violate the constraint rules are corrected to obtain a second correction network. Edges connecting two different subgraphs are selected from the candidate causal edge set as candidate edges, and the candidate edges are sorted in reverse order according to the mutual information value to obtain a candidate edge sequence. The edges in the candidate edge sequence are compared with the causal edge candidate set. If the comparison is successful, the edges in the candidate edge sequence are added as new edges to the second correction network to obtain the correction network structure.

6. The event probability prediction method under tsunami disaster according to claim 5, characterized in that, The step of pruning the modified network structure to obtain the target network structure specifically includes: The modified network structure was pruned using the chi-square test, and the difference between the observed frequency and the expected frequency of each pair of variables in the modified network structure was calculated. If the difference is higher than a preset threshold, the edge between the two variables is cut off to obtain the target network structure.

7. A probability prediction system for tsunami disasters, characterized in that, The event probability prediction system under tsunami disasters is used to implement the event probability prediction method under tsunami disasters according to any one of claims 1-6, and the event probability prediction system under tsunami disasters includes: The dataset acquisition module is used to acquire global tsunami historical datasets, filter out variables and their corresponding initial datasets from the global tsunami historical datasets, and discretize the initial datasets corresponding to the continuous variables to obtain intermediate datasets; The data missing value imputation module is used to impute missing values ​​in the intermediate dataset using a Markov blanket node state value combination strategy to obtain a complete dataset. The basic network construction module is used to learn the Bayesian network structure using the complete dataset to obtain the basic Bayesian network structure. The network correction module is used to construct a causal edge candidate set, use the causal edge candidate set to correct the basic Bayesian network structure to obtain a corrected network structure, and prune the corrected network structure to obtain the target network structure. The event probability prediction module is used to acquire variable data under the tsunami disaster to be predicted, input the variable data into the target network structure for calculation, and obtain the probability of the corresponding event occurring under the tsunami disaster to be predicted.

8. A terminal, characterized in that, The terminal includes: a memory, a processor, and an event probability prediction program for tsunami disasters stored in the memory and executable on the processor. When the event probability prediction program for tsunami disasters is executed by the processor, it implements the steps of the event probability prediction method for tsunami disasters as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an event probability prediction program under tsunami disasters, which, when executed by a processor, implements the steps of the event probability prediction method under tsunami disasters as described in any one of claims 1-6.

Images