Dangerous chemical accident dynamic affair map construction method based on multi-source data fusion

Abstract

The invention discloses a multi-source data fusion-based dangerous chemical accident dynamic event map construction method, relates to the technical field of event map construction, and applies static mapping of a dangerous chemical event map to a time dimension to describe dynamic, sudden and timely characteristics of dangerous chemical accident occurrence. Based on a large number of accident investigation report texts, a Word2vec training word vector technology is applied, and through information extraction, knowledge fusion, ontology construction and dynamic reasoning of event knowledge, a dangerous chemical accident affair graph is systematically constructed. And a ResNet method is adopted in the knowledge fusion process, so that the equalization problem of a sample data set is solved. And general event generalization is carried out by combining causal relationship feature extraction of the hazardous chemical accidents, and a relatively complete hazardous chemical accident affair map is further constructed.

Description

Technical Field

[0001] The present invention relates to the technical field of event graph construction, and in particular to a method for constructing a dynamic event graph of hazardous chemical accidents by fusing multi-source data. Background Art

[0002] Dangerous chemical accidents often have dynamic, sudden and timely evolution characteristics, which are closely related to time attributes. Therefore, time information is crucial for management decisions. Although researchers in the financial field have constructed financial event maps, there are still few studies on hazardous chemical incidents.

[0003] In the prior art, most of the research on hazardous chemical incidents focuses on a specific element, such as Chinese named entity recognition, entity linking, relationship extraction and other technical processes. However, an important feature of accident event knowledge is high dynamics. The construction of its event knowledge graph needs to include the timeliness of accident risk knowledge. The method for constructing a dynamic event graph of accidents provided by the present invention provides a time dimension control for accident risk cognition, and utilizes the characteristics of information acquisition and knowledge fusion, ontology construction and knowledge dynamic reasoning to construct a top-down dynamic event graph of hazardous chemical accidents. Then, through the developed intelligent question-answering system, visualized risk dynamic control of hazardous chemical accidents is achieved. Summary of the invention

[0004] The purpose of the present invention is to solve the problems in the prior art and to propose a method for constructing a dynamic event map of dangerous chemical accidents based on multi-source data fusion.

[0005] A method for constructing a dynamic causal map of hazardous chemical accidents based on multi-source data fusion adopts a bottom-up approach to construct the causal map, which includes a data layer and a model layer. The data layer includes:

[0006] The information extraction module aims to combine the contextual information of the hazardous chemical accident investigation report text to more accurately identify the trigger words contained in the event and determine the event type;

[0007] The knowledge fusion module aims to integrate knowledge to improve its quality and resolve redundant and conflicting information. It mainly includes entity linking, which links the hazardous chemical incident entity objects extracted from the text with the corresponding named entities in the hazardous chemical accident knowledge base. The entity points to the same fact in the real world. Through entity linking, the diversity and ambiguity of entity references are resolved.

[0008] The model layer includes:

[0009] The ontology building module aims to build the ontology and establish the conceptual template of the model layer;

[0010] A dynamic causal graph module for hazardous chemical accidents is constructed, and an open source graph database is used to store the causal graph. The database uses elements such as color, arrows, and text to clearly and intuitively display the relationship between entities in the form of a graph. According to the above method, the extraction of causal event pairs and fuzzy causal event pairs is completed to construct a hazardous chemical safety causal graph. The cause events and result events of hazardous chemical accidents are nodes in the graph, and the corresponding causal relationships are edges in the graph. The degree of correlation between events is judged according to the weights, so as to obtain the causal relationship evolution path of hazardous chemical accidents. The edges representing causal relationships in the causal graph are all one-way arrows, which can only point from the cause event to the result event. However, there are multiple arrows pointing to the same result event or one event pointing to multiple result events. Such events may be the path that affects the occurrence of the accident.

[0011] In the above-mentioned method for constructing a dynamic event graph of hazardous chemical accidents based on multi-source data fusion, the information extraction module adopts the CNN (neural convolutional network)-BiLSTM (bidirectional long short-term memory network)-Attention (attention mechanism) model to identify the trigger words of hazardous chemical accidents and complete the classification of hazardous chemical accidents, which specifically includes the following parts:

[0012] Document Embedding: The document embedding part embeds the input document based on Word2Vec to obtain the vector representation of the document.

[0013] The feature extraction part consists of three parts: word-level feature extraction based on CNN, sentence-level feature extraction based on BiLSTM, and important feature extraction based on the Attention mechanism;

[0014] Output layer: The output layer uses the softmax classifier to identify the category to which the trigger word belongs and completes event classification.

[0015] In the above-mentioned method for constructing a dynamic event graph of hazardous chemical accidents based on multi-source data fusion, the entity linking adopts a supervised algorithm and combines CNN to connect hazardous chemical accident entities. The document vector of the hazardous chemical accident investigation report obtained by Word2vec is trained by CNN to obtain local features between word vectors, so that the word vector contains more local semantic information and better represents the implicit features contained in the text data. At the same time, local connection and weight sharing are used in CNN training to effectively reduce the number of parameters of the CNN network, reduce complexity, and improve the model training speed.

[0016] In the event type recognition task, CNN is used for word-level feature extraction. The convolutional layer in CNN uses multiple convolution kernels to slide on the input text sequence to obtain words and their context features, and presents them in the form of multiple feature maps, which are then pooled and concatenated to form a single feature map.

[0017] In the above-mentioned method for constructing a dynamic event graph of hazardous chemical accidents based on multi-source data fusion, the ontology construction module adopts data-driven semi-automatic ontology construction, sets four hazardous chemical accident attributes: basic attributes of accident events, event causal relationship attributes, direct cause attributes and indirect cause attributes, and includes the following processes:

[0018] S1: Calculation of conceptual parallelism;

[0019] S2: Extract the subordinate and superior relationships of entities, mainly extracting the causal relationship of entities;

[0020] S3: Generate ontology, use hierarchical clustering method to cluster dynamic concepts at different levels, assign common super words, and build the ontology structure.

[0021] In the above-mentioned method for constructing a dynamic event graph of hazardous chemical accidents based on multi-source data fusion, for event graphs with complex time characteristics, dynamic event graph reasoning based on multi-relationship cyclic events is used to supplement them. The main body of the model includes a multi-relationship proximity aggregator and a time series event encoder. First, the hazardous chemical accident entities and relationships are embedded to construct a four-tuple risk event. The knowledge unit in the four-tuple can be represented as (h; r; s; t), where h and t are the head entity and the tail entity, respectively, representing the entity relationship under the timestamp. The multi-relationship proximity aggregator aggregates the proximity information of each timestamp entity and extracts the proximity features of the entity. The process of the multi-relationship proximity aggregator is divided into the following steps:

[0022] S1: Get the domain information corresponding to all relations of the quadruple;

[0023] S2: Use an aggregator to filter and retain the domain information of the target entity to be predicted;

[0024] S3: Calculate the neighboring information of the target entity as a cumulative average vector and obtain the neighboring features represented by the vector;

[0025] The time series event encoder establishes a joint probability model of the dynamic knowledge graph by obtaining a vector representation of the fusion of domain information at different timestamps to describe the dependencies of dynamic risk events of hazardous chemical accidents on multiple relationships and multiple time dimensions. The first layer of the time series event encoder is a long short memory network LSTM layer, which captures the longer time information of hazardous chemical accident risk events, and regards entity prediction and relationship prediction as a multi-class problem, and uses a multi-class cross entropy loss function for prediction.

[0026] In the above-mentioned method for constructing a dynamic event graph of hazardous chemical accidents based on multi-source data fusion, the maximum pooling operation is used in the construction of the ontology model, and then all the feature values ​​obtained by pooling are spliced ​​together to form a single feature graph.

[0027] Compared with the existing technology, the present invention is based on a large number of accident investigation report texts, uses Word2vec training word vector technology, and systematically constructs a dangerous chemical incident story map through information extraction, knowledge fusion, ontology construction and dynamic reasoning of event knowledge. The ResNet method is used in the knowledge fusion process to achieve the balance problem of the sample data set, combined with the causal relationship feature extraction of hazardous chemical accidents, generalization of overall events, and further constructs a relatively complete hazardous chemical incident story map. The visualization technology is used to show the evolution process and law of the causal relationship of events in hazardous chemical accidents, and an intelligent question-and-answer system for hazardous chemical accidents is developed to realize intelligent analysis of hazardous chemical accidents. The focus is on correctly understanding Chinese questions and accurately distinguishing the behavioral events therein, matching events from the event knowledge base, and obtaining relevant information, returning the best answer to the user, and completing one question and answer. Experiments show that the present invention can effectively improve the accuracy of entity and relationship prediction. The research results have good application prospects and have positive guiding significance for the prevention of hazardous chemical accidents and safety risk control. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a structural schematic diagram of a transport unit in a method for constructing a dynamic event graph of hazardous chemical accidents based on multi-source data fusion proposed by the present invention.

[0029] Figure 2 A schematic diagram of the structure of logistics transportation integration in a method for constructing a dynamic event graph of dangerous chemical accidents based on multi-source data fusion proposed by the present invention DETAILED DESCRIPTION

[0030] Reference Figure 1-2The present invention discloses a method for constructing a dynamic causal map of hazardous chemical accidents based on multi-source data fusion. A bottom-up method is used to construct the causal map. The bottom-up construction method first summarizes and organizes the entities to form basic concepts, and then gradually extracts entities upward to form high-level concepts. In other words, entities are extracted from public linked data, and entities with higher credibility are selected to be added to the knowledge base. Then, a conceptual model of the top-level ontology is constructed. The technical structure of the causal map mainly includes four steps: information extraction, knowledge fusion, ontology construction, and construction of a dynamic causal map of hazardous chemical accidents.

[0031] Information extraction and knowledge fusion are the main links in constructing the event graph, which solves the problem of how to automatically extract candidate knowledge units from multi-source heterogeneous information. Text information extraction generally includes two tasks: named entity recognition and relationship extraction. The experimental data comes from a large number of accident investigation report texts. In Word2vec, the training efficiency of the training model CBOW is relatively high. At the same time, for the document-level event extraction of hazardous chemical accident incidents in the present invention, it is necessary to better understand the semantic information of the entire investigation report text. Therefore, the present invention adopts the CBOW model for training to obtain the document vector representation of hazardous chemical accident incidents.

[0032] Through the analysis of the text description of the hazardous chemical accident investigation report, it can be seen that the text extracted from the hazardous chemical accident investigation report usually contains one or more event trigger words, so it is necessary to combine the contextual information of the entire document to more accurately identify the trigger words contained in the event. Trigger words are important representations that indicate the occurrence of an event and determine the type of event. By identifying event trigger words, the type of event can be determined. Therefore, the purpose of event trigger word identification is to identify the event contained in the sentence and determine its type. Therefore, event type identification is modeled as a text classification task. Considering the need to obtain better text semantic information, the present invention uses the CNN-BiLSTM-Attention model to identify the trigger words of hazardous chemical accident events and complete the classification of hazardous chemical accident events. The model structure is as follows Figure 1 As shown in the figure, it mainly consists of three parts: (1) document embedding part; (2) feature extraction part; (3) output layer. The document embedding part performs WordEmbedding on the input document based on Word2Vec to obtain the vector representation of the document; the feature extraction part consists of three parts, namely word-level feature extraction based on CNN, sentence-level feature extraction based on BiLSTM, and important feature extraction based on the Attention mechanism. The output layer uses the features extracted from the second part of the model through a softmax classifier to identify the category to which the trigger word belongs, thus completing event classification.

[0033] After information extraction, the entities of hazardous chemical incidents, the relationships between entities, attributes and some time information can be obtained. Although the optimized deep learning model has advantages in entity recognition and relationship extraction, these results also contain a considerable amount of redundant and conflicting information. At the same time, in order to improve the logic and level of knowledge, it is necessary to integrate knowledge to improve the quality of knowledge. In other words, it is necessary to delete incorrect and redundant knowledge and eliminate contradictory and conflicting knowledge. Entity linking is the main task of knowledge fusion.

[0034] Entity linking refers to linking the hazardous chemical incident entity objects extracted from the text with the corresponding named entities in the hazardous chemical accident knowledge base, which point to the same fact in the real world. The entity names in the field of hazardous chemical incidents are complex and diverse. Through entity linking, the diversity and ambiguity of entity references can be effectively resolved. The current entity linking algorithms mainly include unsupervised algorithms and supervised algorithms. Unsupervised algorithms are generally based on dictionary matching or similarity calculations, and have high requirements for data standardization. The entity data in the field of hazardous chemical accidents is complex and irregular, so a supervised algorithm is used in combination with a convolutional neural network (CNN) to connect hazardous chemical accident entities.

[0035] The CNN model can capture the local correlation of data, thereby obtaining more potential implicit features. Therefore, by training the document vector of the hazardous chemical accident investigation report obtained through Word2vec through CNN, the local features between word vectors can be obtained, so that the word vector contains more local semantic information and better represents the implicit features contained in the text data. At the same time, the use of local connections and weight sharing in CNN training can effectively reduce the number of CNN network parameters, reduce complexity, and increase the speed of model training.

[0036] Therefore, in the event type recognition task, CNN is used for word-level feature extraction. The convolution layer in CNN uses multiple convolution kernels to slide on the input text sequence to obtain the words and their context features, and present them in the form of multiple feature maps (feature maps), which are then pooled and concatenated to form a single feature map.

[0037] Figure 2 The figure shows the process of CNN acquiring word-level features of text data (the feature maps obtained by sliding different convolution kernels are represented by different colors in the figure), where the vector representation of the sentence obtained by Word2vec is Where si is the i-th sentence in the document, w i,j is the vector corresponding to the j words in the i-th sentence, with a dimension of dw.

[0038] In the convolution operation, multiple convolution kernels are used to obtain the features of a word and its surrounding words. Therefore, the convolution kernel needs to be set first. The convolution kernel is usually set as a matrix In the form of, then use the set convolution kernel w i,j Convolution operation is performed to obtain the feature map. The feature map of a single word can be expressed as follows:

[0039] c j =f(u·w i,j:i,j+k-1 +β)

[0040] Among them, w i,j:i,j+k-1 Represents the word vector of word j and its surrounding k words, is the bias, f is the nonlinear activation function, and in order to obtain the feature map of the entire sentence, it is necessary to use the set convolution kernel to perform a convolution operation on each word in the sentence, and finally obtain the feature map of the entire sentence as follows:

[0041] g=[c1,c2,...,c Nw-k+1 ]

[0042] in, N w Indicates the maximum length of a word sequence.

[0043] In order to obtain different features of words, multiple different convolution kernels are usually used in CNN. Assuming that m1 convolution kernels are used during the convolution operation, there will be m1 feature maps, such as Figure 2 As shown, there are 3 feature maps.

[0044] In CNN, after convolution, the extracted features are input to the pooling layer, in which the input features can be deduplicated and de-noised through various pooling operations to reduce irrelevant interference items, thereby reducing the feature dimension and obtaining the most important features while reducing the amount of data operations. Among the pooling operations, the most effective one is the maximum pooling operation, so the present invention adopts the maximum pooling operation in model construction, and then splices all the feature values ​​obtained by pooling together to form a single feature map.

[0045] The hazardous chemical accident investigation report text was used for pre-training, and the training was performed based on 5429 hazardous chemical field annotation data processed from the report. The parameter training results of the previous CNN-BiLSTM-Attention model were transferred to the embedding layer and CRF layer of the new model to form three comparative experiments with different parameters. The experimental results are shown in Table 1. From the experimental results, it is proved that the use of CNN-BiLSTM-Attention can effectively improve the entity recognition effect of the model.

[0046] Table 1 Comparison of model training results

[0047]

[0048] The fact map can be logically divided into the data layer and the model layer. After knowledge fusion, a series of basic facts in the hazardous field are correctly expressed, but facts are not equivalent to knowledge. In order to obtain a knowledge system with a logical structure, it is also necessary to build an ontology and establish a conceptual template for the model layer. With the help of the ontology library, the relationships, relationships, attributes and entity types between hazardous chemical accident entities can be standardized. At present, commonly used ontology construction tools mainly include visual manual construction and semi-automatic construction. The Java-based Jena reasoning engine provides a method for semi-automatically building an ontology by calling an algorithm. It can store large-scale RDF data and provide an API for processing OWL ontology language and RDFS ontology, which greatly improves the efficiency of ontology construction. The data-driven semi-automatic ontology construction process can be divided into three stages:

[0049] 1. Calculation of parallel relationships between concepts;

[0050] 2. Extract the subordinate and superior relationships of entities, mainly extracting the cause-effect relationship of entities;

[0051] 3. Generate ontology. Generally, hierarchical clustering method is used to cluster dynamic concepts at different levels, assign common super words, and construct the ontology structure.

[0052] According to the construction of bottom-up knowledge and technology processes, the model was dynamically established based on data-driven and supplemented by manual construction, combining the knowledge characteristics of the field of hazardous chemical accidents. Based on the Jena reasoning engine, the hazardous chemical accident domain ontology was semi-automatically constructed, and four hazardous chemical accident attributes were set: basic attributes of accident events, event causal relationship attributes, direct cause attributes, and indirect cause attributes. For the event graph reasoning with complex time characteristics, dynamic event graph reasoning (Multi-Net) based on multi-relationship cyclic events was used to supplement the knowledge graph. The main body of the Multi-Net model includes a multi-relationship neighbor aggregator and a time encoder. First, the hazardous chemical accident entities and relations are embedded to construct a four-tuple risk event. The knowledge unit in the four-tuple can be represented as (h; r; s; t), where h and t are the head entity and the tail entity, respectively, representing the entity relationship under the timestamp. The function of the multi-relationship proximity aggregator is to aggregate the proximity information of each timestamp entity and extract the proximity features of the entity. The process of the improved relationship aggregator can be divided into the following steps. (1) Obtain the domain information corresponding to all relations of the quadruple; (2) Use an aggregator to filter and retain the domain information of the target entity to be predicted; (3) Calculate the neighboring information of the target entity as a cumulative average vector and obtain the neighboring features represented by the vector.

[0053] After obtaining the vector representation of the fusion of domain information at different timestamps, the time series event encoder establishes a joint probability model of the dynamic knowledge graph to describe the dependencies of dynamic risk events of hazardous chemical accidents in multiple relationships and multiple time dimensions. The characteristic of this model is that the first layer of the time series event encoder is a long short memory network LSTM layer. The LSTM time series modeling system can capture the longer time information of hazardous chemical accident risk events. Entity prediction and relationship prediction are regarded as a multi-class problem, and a multi-class cross entropy loss function can be used for prediction. In the hazardous chemical accident risk knowledge base, there is an imbalance in the distribution of entity and relationship data, which greatly affects the fitting and generalization capabilities of the model.

[0054] Constructing a dynamic event map of hazardous chemical accidents, the present invention uses the open source graph database Neo4j to store the event map. Neo4j is a high-performance NoSQL graph database implemented by Java language, with more simplified data modeling, mining knowledge systems from point to surface, object-oriented thinking, simple multi-dimensional data representation and other obvious advantages. It provides a complete graph query language, which can query and analyze related data better and faster. Neo4j has a dedicated query language-Cypher statement. The present invention imports the constructed hazardous chemical safety accident event related data into the Neo4j database through Cypher OAD CSV statement, CypherCREATE statement, Cypher MATCH statement, Cypher MERGE statement and neo4j-import tool. The database uses elements such as color, arrows, and text to clearly and intuitively display the relationship between entities in a graphical manner. According to the above method, the extraction of causal event pairs and fuzzy causal event pairs is completed to construct a hazardous chemical safety event map. The cause events and result events of hazardous chemical accidents are nodes in the graph, and the corresponding causal relationships are edges in the graph. After the cause-and-effect graph is constructed, the correlation between the events can be displayed, and the degree of correlation between the events can be judged according to the weights, so as to obtain the evolution path of the causal relationship of hazardous chemical accidents. The edges representing the causal relationship in the cause-and-effect graph are all one-way arrows, which can only point from the cause event to the result event. However, there are multiple arrows pointing to the same result event or one event pointing to multiple result events. Such events may be the path that affects the occurrence of the accident.

[0055] It is known from common technical knowledge that the present invention can be implemented by other embodiments that do not deviate from its spirit or essential features. Therefore, the above disclosed embodiments are only illustrative in all respects and are not exclusive. All changes within the scope of the present invention or within the scope equivalent to the present invention are included in the present invention.

Claims

1. A method for constructing a dynamic event graph of dangerous chemical accidents based on multi-source data fusion, characterized in that: A bottom-up approach is adopted to construct a truth map, which includes a data layer and a model layer. The data layer includes: The information extraction module aims to combine the contextual information of the hazardous chemical accident investigation report text to more accurately identify the trigger words contained in the event and determine the event type; The knowledge fusion module aims to integrate knowledge to improve its quality and resolve redundant and conflicting information. It mainly includes entity linking, which links the hazardous chemical incident entity objects extracted from the text with the corresponding named entities in the hazardous chemical accident knowledge base. The entity points to the same fact in the real world. Through entity linking, the diversity and ambiguity of entity references are resolved. The model layer includes: The ontology building module aims to build the ontology and establish the conceptual template of the model layer; A dynamic causal graph module for hazardous chemical accidents is constructed, and an open source graph database is used to store the causal graph. The database uses elements such as color, arrows, and text to clearly and intuitively display the relationship between entities in the form of a graph. According to the above method, the extraction of causal event pairs and fuzzy causal event pairs is completed to construct a hazardous chemical safety causal graph. The cause events and result events of hazardous chemical accidents are nodes in the graph, and the corresponding causal relationships are edges in the graph. The degree of correlation between events is judged according to the weights, so as to obtain the causal relationship evolution path of hazardous chemical accidents. The edges representing causal relationships in the causal graph are all one-way arrows, which can only point from the cause event to the result event. However, there are multiple arrows pointing to the same result event or one event pointing to multiple result events. Such events may be the path that affects the occurrence of the accident.

2. According to claim 1, a method for constructing a dynamic event map of dangerous chemical accidents based on multi-source data fusion is characterized by: The information extraction module uses the CNN (neural convolutional network)-BiLSTM (bidirectional long short-term memory network)-Attention (attention mechanism) model to identify the trigger words of hazardous chemical accidents and complete the classification of hazardous chemical accidents, which specifically includes the following parts: Document Embedding: The document embedding part embeds the input document based on Word2Vec to obtain the vector representation of the document. The feature extraction part consists of three parts: word-level feature extraction based on CNN, sentence-level feature extraction based on BiLSTM, and important feature extraction based on the Attention mechanism; Output layer: The output layer uses the softmax classifier to identify the category to which the trigger word belongs and completes event classification.

3. The method for constructing a dynamic event graph of dangerous chemical accidents based on multi-source data fusion according to claim 1 is characterized by: The entity linking adopts a supervised algorithm and combines CNN to connect hazardous chemical accident entities. The document vector of the hazardous chemical accident investigation report obtained by Word2vec is trained by CNN to obtain local features between word vectors, so that the word vector contains more local semantic information and better represents the implicit features contained in the text data. At the same time, local connection and weight sharing are adopted in CNN training to effectively reduce the number of parameters of the CNN network, reduce complexity, and improve the model training speed. In the event type recognition task, CNN is used for word-level feature extraction. The convolutional layer in CNN uses multiple convolution kernels to slide on the input text sequence to obtain words and their context features, and presents them in the form of multiple feature maps, which are then pooled and concatenated to form a single feature map.

4. The method for constructing a dynamic event graph of dangerous chemical accidents based on multi-source data fusion according to claim 4 is characterized by: The ontology construction module adopts data-driven semi-automatic ontology construction, sets four hazardous chemical accident attributes: basic attributes of accident events, event causal relationship attributes, direct cause attributes and indirect cause attributes, and includes the following processes: S1: Calculation of conceptual parallelism; S2: Extract the subordinate and superior relationships of entities, mainly extracting the causal relationship of entities; S3: Generate ontology, use hierarchical clustering method to cluster dynamic concepts at different levels, assign common super words, and build the ontology structure.

5. The method for constructing a dynamic event graph of dangerous chemical accidents based on multi-source data fusion according to claim 1 is characterized by: For the event graph with complex time characteristics, dynamic event graph reasoning based on multi-relationship cyclic events is used to supplement it. The main body of the model includes a multi-relationship proximity aggregator and a time series event encoder. First, the dangerous chemical accident entities and relations are embedded to construct a four-tuple risk event. The knowledge unit in the four-tuple can be represented as (h; r; s; t), where h and t are the head entity and tail entity, respectively, indicating the entity relationship under the timestamp. The multi-relationship proximity aggregator aggregates the proximity information of each timestamp entity and extracts the proximity features of the entity. The process of the multi-relationship proximity aggregator is divided into the following steps: S1: Get the domain information corresponding to all relations of the quadruple; S2: Use an aggregator to filter and retain the domain information of the target entity to be predicted; S3: Calculate the neighboring information of the target entity as a cumulative average vector and obtain the neighboring features represented by the vector; The time series event encoder establishes a joint probability model of the dynamic knowledge graph by obtaining a vector representation of the fusion of domain information at different timestamps to describe the dependencies of dynamic risk events of hazardous chemical accidents on multiple relationships and multiple time dimensions. The first layer of the time series event encoder is a long short memory network LSTM layer, which captures the longer time information of hazardous chemical accident risk events, regards entity prediction and relationship prediction as a multi-class problem, and uses a multi-class cross entropy loss function for prediction.

6. The method for constructing a dynamic event graph of dangerous chemical accidents based on multi-source data fusion according to claim 4 is characterized by: The maximum pooling operation is used in the construction of the ontology model, and then all the feature values ​​obtained by pooling are spliced ​​together to form a single feature map.

Images