Chemical engineering field event prediction method and system based on affair atlas

Through the event prediction method based on the event map, BiLSTM-CRF and graph neural network are used to build the event map in the chemical industry, solving the problems of low accuracy of event prediction and insufficient early warning in the chemical industry, achieving more accurate risk identification and early warning, and improving the The safety of the chemical industry.

CN118917681APending Publication Date: 2024-11-08LIAONING UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410927255.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The chemical industry has low accuracy in event prediction and insufficient early warning, making it difficult for the existing technology to effectively identify potential risks and causal relationships in chemical accidents.

Method used

Using event prediction method based on event map, the event element information is automatically identified from chemical accident investigation reports through BiLSTM-CRF, the event map in the chemical industry is constructed, and the graph neural network is used to capture the high-order relationship between events and calculate candidate subsequent events and events. The similarity of context is predicted.

Benefits of technology

It improves the accuracy and comprehensiveness of event prediction, can effectively identify potential risks and causal relationships, provides more accurate early warnings and risk assessments, and enhances the safety of the chemical industry.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917681A_ABST
    Figure CN118917681A_ABST
Patent Text Reader

Abstract

The invention discloses a chemical engineering field event prediction method and system based on a affair graph, and the method comprises the following steps: (1), defining an event trigger word and an event element, and carrying out the event extraction from an accident cause part in a chemical engineering accident investigation report; (2) constructing a chemical engineering field affair map; (3) carrying out initial embedding on the chemical engineering event, inputting the initial embedding into the BiLSTM network, and obtaining the representation of an event sequence; (4) stacking a plurality of embedded propagation layers to perform high-order vector representation on event nodes by using a graph neural network according to the event graph; and (5) calculating the vector similarity between each candidate subsequent event and the event context, and selecting the candidate event with the maximum similarity as a prediction result. In order to solve the problem that the current event prediction accuracy is low, the event prediction model based on the event graph is provided; for the problems of event missing and insufficient early warning in the chemical industry field, the method is applied to the chemical industry field, and a chemical industry field event prediction system based on the affair atlas is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of big data and artificial intelligence technology, and in particular to a method and system for predicting events in the chemical industry based on a fact map. Background Art

[0002] The chemical industry occupies an important position in the national economy. It is a basic industry and a pillar industry. Once an accident occurs, it will have great destructive power and social impact. With the continuous development of science and technology, especially the application of technologies such as big data and artificial intelligence, the ability to predict chemical incidents has been significantly improved. By mining and analyzing historical data, potential safety hazards and risk points can be identified, and possible chemical incidents can be predicted.

[0003] Event prediction is the process of using historical data and models to predict possible future events or trends. Event prediction for the chemical industry can help identify potential safety hazards and risks in advance, and take preventive measures to reduce the probability of accidents and improve production safety.

[0004] Event graphs are becoming an important technology in the field of event prediction. Event graphs can more comprehensively reflect the interaction and evolution process between events. Through event graphs, we can clearly see the causal relationship and sequential relationship between various events in chemical events, which helps to reveal the internal logic of the development of events and provide a basis for predicting possible future events. Using chemical industry accident investigation reports as a corpus, constructing event graphs, and developing a chemical industry event prediction system based on event graphs can provide early warnings for chemical accidents and effectively respond to industry risks. Summary of the invention

[0005] In order to solve the problem of low accuracy in current event prediction, an event prediction method based on causal graph is proposed; for the problems of missing events and insufficient early warning in the chemical industry, the event prediction method is applied to the chemical industry, and an event prediction system in the chemical industry based on causal graph is studied and implemented.

[0006] The present invention provides the following technical solution: a method for predicting chemical industry events based on event graph, comprising the following steps:

[0007] (1) Construct a graph of events, treating events as nodes in the graph and the impact relationships between events as edges.

[0008] Taking chemical events in chemical accident investigation reports as the original corpus, the BiLSTM-CRF method is used to automatically identify and extract structured event element information from unstructured text, including operators, time, location, chemical process flow, emergency response, accident causes, and production equipment; events and their element information are used as nodes, and the semantic relationships between events are used as edges to construct a cause-and-effect graph.

[0009] The specific implementation process of extracting structured event element information is as follows: using the accident investigation report written in accordance with the specifications as the original corpus, defining event trigger words and event elements according to the characteristics of events in the chemical industry, extracting the relationship between events from the text based on the trigger words and event elements using a deep learning-based event extraction method, and identifying and parsing the semantic associations between events in the text.

[0010] The specific implementation process of constructing the event graph is: based on the event trigger words and event elements obtained by event extraction; extracting the event chain from them, and then constructing the chemical industry event graph on this basis, in which the nodes represent events and the edges represent the relationship between events.

[0011] (2) Given the word vectors of the four components of an event, concatenate the four word vectors as the representation vector of the entire event.

[0012] Encode the context of the event, including the operator, time, location, chemical process flow, emergency response, cause of the accident, and production equipment information; encode the candidate subsequent events to describe the subsequent events that are likely to occur; through these codes, the model can obtain the initial representation of the event context and candidate events for the derivation and prediction of subsequent events; a single event structure contains four elements v, a 0 、a 1 、a 2 , where v is the predicate of the event, a 0 is the subject of the event, a 1 is the object of the event, a 2 is the indirect object of the event, and the word vectors of the four elements are concatenated as the representation vector of the event.

[0013] (3) The event context and candidate events are input into the BiLSTM network respectively to obtain the sequence representation of the event; the BiLSTM is used to capture the dependency relationship between the previous and next contexts in the event sequence, and the model updates the representation vector of each event node so that the updated event node contains richer sequence information.

[0014] (4) By stacking multiple embedding propagation layers, similar signals in high-order connectivity are captured and event representation is enriched. According to the event graph, a graph neural network is used to represent event nodes with high-order vectors. Through the iterative aggregation of the graph neural network, the representation of each event node is more finely encoded.

[0015] (5) Calculate the vector similarity between each candidate subsequent event and the event context, and select the candidate event with the greatest similarity as the prediction result.

[0016] A prediction system used in the chemical industry event prediction method of the event graph includes:

[0017] (1) Registration and login module: used to manage users’ personal accounts and information and ensure the security and integrity of the system. It mainly includes login, registration and personal information modification functions;

[0018] (2) Schematic Diagram Construction Module: It is used for users to understand the process of constructing a schema, improve the schema, and fill in the corresponding event nodes. The schema construction module provides users with text input, event extraction, and schema representation functions.

[0019] (3) Chemical event prediction module: The user takes the event context and candidate events as input to the system. The event prediction module uses the event prediction algorithm to calculate the similarity between the event context and multiple candidate events to obtain the prediction result. The prediction result is the probability of each candidate event occurring. The system visualizes the event prediction result and displays the event context and prediction result to the user in a graphical manner. The user can perform event graph query based on the event element. The system will return the node containing the event element and other nodes connected to the node.

[0020] (4) System management module: mainly used by administrators to manage server resources and ensure normal use of users, including user query, system monitoring and user management functions.

[0021] The beneficial effects created by the present invention are as follows: first, an event prediction model EPEG based on the event graph is proposed. The model is based on the event graph, uses BiLSTM to learn the contextual representation of events, and uses multiple embedded propagation layers in the graph neural network to fully capture the relationship between events, obtain the high-order representation vector of each event, and achieve comprehensive and accurate event prediction based on the similarity between candidate subsequent events and event contexts. Secondly, an event prediction system for the chemical industry is developed based on this model. The system can predict events in a given chemical event context, and judge possible chemical events and potential risks. The design of specific modules includes event graph construction, chemical event prediction, registration and login, and system management. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a functional module diagram of the chemical industry event prediction system based on the event graph;

[0023] Figure 2 It is a diagram of the event structured representation of the chemical industry event prediction system based on the event graph;

[0024] Figure 3 It is a schematic diagram of event prediction of a method for predicting events in the chemical industry based on event graphs;

[0025] Figure 4 This is a detailed structural diagram of an event prediction method based on event graph;

[0026] Figure 5 This is a schematic diagram of embedding propagation of an event prediction method based on event graph;

[0027] Figure 6 This is the home page of a chemical industry event prediction system based on the event graph;

[0028] Figure 7 The accuracy of different propagation layer hop counts provided by the embodiment of the present invention Figure 8 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0030] The present invention provides a chemical industry event prediction method and system based on a causal graph, and the design concept is as follows: (1) The construction of a causal graph is the basis for chemical industry event prediction. In the process of construction, chemical industry events in chemical accident investigation reports are used as the original corpus to predict chemical industry events. In the process of using the system, users can also improve the graph according to the chemical industry texts they have mastered. Users can upload texts. After data verification, the system extracts events and finally stores them in a structured form in a graph database. (2) To predict chemical industry events, users must first enter event contexts and candidate events. Through event entry, the system obtains event contexts and candidate events. Next, embedding learning is used to obtain event vector representations, and the similarity between context events and candidate events is calculated. Commonly used similarity calculation methods include cosine similarity, Euclidean distance, etc. By comparing the similarity, the degree of correlation between candidate events and context events can be determined. Finally, the system presents the prediction results and event context to the user in the form of a graph. Users can intuitively understand the event context, make decisions based on the prediction results, and take corresponding measures to deal with potential risks and problems.

[0031] The functional modules of the chemical industry event prediction system based on the event graph designed based on this method are as follows Figure 1 As shown. It includes four functional modules: registration and login, event graph construction, chemical event prediction and system management.

[0032] The representation structure of an event refers to the way to describe and express an event. Event trigger words and event elements can be used to specifically represent a chemical event. A chemical field event contains an event trigger word and multiple event elements. Event trigger words refer to the key words that trigger an event. They are usually represented by predicates and verb-object phrases. They are the core elements of an event and the main features for identifying an event. For example, when describing a leakage incident, the event trigger word may be "trigger"; when describing a fire incident, the event trigger word may be "ignite"; when describing an explosion incident, the event trigger word may be "detonate". Event elements refer to specific descriptions related to the event. Usually, a complete event description will have one or more event elements, including the occurrence of the event, location, equipment, and personnel. The structured representation of an event is as follows: Figure 2 shown.

[0033] Given an event context and a list of candidate subsequent events, the most reasonable subsequent event is selected from the list of candidate events. This is the definition of the event prediction task. The model first vectorizes the event context and all the events in the candidate events, then calculates the similarity between each candidate subsequent event and the event in the context as the confidence of the candidate event as the prediction result, and finally selects the event with the largest confidence (formation, suffocation environment) as the predicted subsequent event of the context. The event prediction diagram is as follows Figure 3 shown.

[0034] This paper proposes an event prediction method based on event graph, which aims to improve the accuracy of event prediction by capturing the relationship and evolution trend between events. Figure 4 shown.

[0035] The event structure is defined as four elements consisting of v, a 0 、a 1 、a 2 , where v is the predicate of the event, a 0 is the subject of the event, a 1 is the object of the event, a 2 is the indirect object of the event. The information contained in the four components is important for describing the entire event. The initial event representation is learned by composing pre-trained word embeddings of verbs and independent variables. For arguments consisting of multiple words, the present invention follows the specification and only uses words recognized by the grammatical analyzer. Words outside the vocabulary and non-existent independent variables are represented by zero vectors. First, it is necessary to build a vocabulary containing all unique words in the dataset. This vocabulary includes every word in the text, and only the most common words are retained. Next, a unique integer index is assigned to each word in the vocabulary. This index will be used to identify each word in the vectorization process. Finally, given an event and the word vectors of the four components of the event, the four word vectors are spliced ​​as the representation vector of the entire event, the mapping of the event to the representation vector is completed, and the representation vector of the event is used as the initial input of the model.

[0036] BiLSTM can capture contextual information in the input sequence because it can consider both past and future information. This ability enables BiLSTM to better understand the contextual meaning when processing sequence data, thereby generating more accurate vector representations. At the same time, BiLSTM can capture long-term dependencies and feature interactions in the input sequence, thereby generating vector representations that contain more feature interaction information. This makes it helpful to improve the prediction accuracy and performance of the model. The BiLSTM structure is introduced to capture the information of the preceding events and the subsequent events of each event. The event representation vector E is input into a BiLSTM with a hidden layer state size of h, and the forward hidden state and reverse hidden state are concatenated as a new and more abstract representation vector of the event, so that it contains information about the entire context.

[0037] The graph structure provides a logical chain between event nodes. On this basis, embedding propagation is performed between connected event nodes. The process is formulated through two main operations: message construction and message aggregation, as shown in formula m u←i =f(e i ,e u ,p ui ), where e i and e u is the representation vector of event u and event i, and mu→i is the similarity signal of event u and event i. The message is constructed based on the graph convolutional network, and pui is set to the graph Laplacian norm. Among them, Nu and Ni represent the first-hop neighbors of event u and event i. From the perspective of message passing, considering that the message being propagated should decay with the path length, pui can be interpreted as a discount factor, and the calculation formula is as follows:

[0038] m u←i =f(e i ,e u ,p ui )

[0039]

[0040] Among them, w1 and w2 belong to R′×d, which are trainable weight matrices for extracting useful information for propagation, and d′ is the transformation size. i Different from the traditional graph convolutional network, e i and e u The interaction between i ⊙e uThe messages transmitted are summed up to ⊙, where ⊙ represents the element-wise product. This makes the messages dependent on the affinity between ei and eu, so more messages can be transmitted. This improves the representation ability of the model. In the message aggregation phase, the messages propagated from the neighborhood of u are aggregated to improve the representation of u. Specifically, the aggregation function is defined as the following formula:

[0041]

[0042] in, Denotes the representation of event u obtained after the first embedding propagation layer. In addition to the messages propagated from neighbors Nu, this paper also considers the self-connection of u: mu←u, which preserves the information of the original features. The advantage of the embedding propagation layer is that it explicitly utilizes the first-order connectivity information to associate event representations. The embedding propagation diagram is as follows Figure 5 shown.

[0043] High-order propagation is performed on the basis of first-order propagation. By enhancing the representation through first-order connectivity modeling, more embedding propagation layers can be stacked to explore high-order connectivity information. This high-order connectivity is crucial for encoding interaction signals to estimate the correlation scores between events. The propagated message is shown in the following formula:

[0044]

[0045] For the prediction subsequent event layer in the method, after propagation through L layers, multiple representations of event u are obtained, namely Since the representations obtained in different layers emphasize messages passed through different connections, they have different contributions in event prediction. They are concatenated to form the final embedding of the event, as shown in formula Here, ‖ is a concatenation operation. This not only enriches the embedding with the embedding propagation layer, but also allows the propagation to be controlled by adjusting L. After further abstracting each event node through the high-order event representation, the final event high-order representation is obtained, where n is the number of events in the event context, c is the candidate subsequent event, and the given context event is calculated. With each subsequent candidate event The similarity between them is Simj, and then the candidate subsequent event with the highest score is selected as the prediction result of the context. The similarity calculation formula is:

[0046] and

[0047] For the method proposed in this paper, in order to verify the effectiveness of the model design and test its performance, this paper compared it with seven other representative event prediction methods on three datasets: SinaNews corpus, CMCNC, and MCNC, and conducted relevant experiments to obtain the accuracy of different models, as shown in Table 1. In the comparative experiments on the three datasets, the results achieved by the neural network-based models PairLSTM and SGNN are significantly better than the count-based PMI and Bigram models. The main reason is that the low-dimensional dense embedding of learning events is more effective for script event prediction than sparse feature representation. Compared with "DeepWalk" (pair-based), "PairLSTM" (chain-based) and "SGNN, EPEG" (graph-based), the graph-based model achieved better performance than the pair-based and chain-based models. This confirms that the event graph structure is more effective than event pairs and event chains, and can provide richer event interaction information for event prediction.

[0048] Table 1: Accuracy of different models

[0049]

[0050]

[0051] In order to verify the role of each module in the model, this paper conducted ablation experiments on the MCNC and SinaNews datasets to explore the impact of BiLSTM, GCN and the proposed model EPEG on event representation, as well as the impact on the final event prediction performance, as shown in Table 2.

[0052] Table 2: Accuracy table of ablation experiments

[0053]

[0054] This paper compares several commonly used similarity scoring metrics to study their impact on performance. As shown in Table 3, it is found that different scoring metrics do have different impacts on performance, and the Euclidean similarity function obtains the best results.

[0055] Table 3: Accuracy table of different similarity functions

[0056]

[0057] In event prediction tasks, the number of hops is usually used to represent the number of rounds of node information transmission. As the number of hops increases, the neighbor information obtained by each node will become richer and more comprehensive, so that the relationship between nodes and the correlation between events can be better captured. However, if the hop count K is set too large, it will bring noise to the event node representation, reduce the difference between events, and thus lead to poor experimental results. The accuracy of different propagation layer hops is as follows: Figure 7 As shown, it can be seen that with the increase of K value, the accuracy of the model first gradually increases, and then decreases significantly. The experimental results are best when the number of hops is 2.

Claims

1. A method for predicting chemical industry events based on event graphs. It is characterized in that It includes the following steps: (1) Construct an event graph, treating events as nodes in the graph and the impact relationships between events as edges; (2) Given the word vectors of the four components of an event, concatenate the four word vectors as the representation vector of the entire event; (3) Input the event context and candidate events into the BiLSTM network respectively to obtain the sequence representation of the event; (4) Capture similar signals in high-order connectivity and enrich event representation by stacking multiple embedding propagation layers; (5) Calculate the vector similarity between each candidate subsequent event and the event context, and select the candidate event with the greatest similarity as the prediction result.

2. A method for predicting chemical industry events based on event graph according to claim 1, It is characterized in that The specific implementation process of step (1) is as follows: using the chemical incidents in the chemical accident investigation report as the original corpus, using the BiLSTM-CRF method to automatically identify and extract structured event element information from the unstructured text, including operating personnel, time, location, chemical process flow, emergency treatment, accident cause, and production equipment; Events and their element information are used as nodes, and the semantic relationships between events are used as edges to construct a matter graph.

3. A method for predicting chemical industry events based on event graph according to claim 1, It is characterized in that The specific implementation process of step (2) is as follows: encoding the context of the event, including the operator, time, location, chemical process flow, emergency treatment, cause of the accident, and production equipment information; encoding the candidate subsequent events to describe the subsequent events that are likely to occur; through these codes, the model can obtain the initial representation of the event context and the candidate events for the derivation and prediction of subsequent events; a single event structure contains four elements v, a 0 、a 1 、a 2 , where v is the predicate of the event, a 0 is the subject of the event, a 1 is the object of the event, a 2 is the indirect object of the event, and the word vectors of the four elements are concatenated as the representation vector of the event.

4. A method for predicting chemical industry events based on event graph according to claim 1, It is characterized in that The specific implementation process of step (3) is: using BiLSTM to capture the dependency relationship between the previous and next contexts in the event sequence, and the model updates the representation vector of each event node so that the updated event node contains richer sequence information.

5. A method for predicting chemical industry events based on event graph according to claim 1, It is characterized in that The specific implementation process of step (4) is as follows: based on the event graph, a graph neural network is used to represent the event nodes with high-order vectors, and through iterative aggregation of the graph neural network, the representation of each event node is more finely encoded.

6. A method for predicting chemical industry events based on event graph according to claim 2, It is characterized in that The specific implementation process of extracting structured event element information is as follows: using the accident investigation report written in accordance with the specifications as the original corpus, defining event trigger words and event elements according to the characteristics of events in the chemical industry, extracting the relationship between events from the text based on the trigger words and event elements using a deep learning-based event extraction method, and identifying and parsing the semantic associations between events in the text.

7. A method for predicting chemical industry events based on event graph according to claim 2, It is characterized in that The specific implementation process of constructing the event graph is: based on the event trigger words and event elements obtained by event extraction; extracting the event chain from them, and then constructing the chemical industry event graph on this basis, in which the nodes represent events and the edges represent the relationship between events.

8. A prediction system used in a method for predicting chemical industry events using a fact map according to any one of claims 1 to 8, It is characterized in that include: (1) Registration and login module: used to manage users’ personal accounts and information and ensure the security and integrity of the system. It mainly includes login, registration and personal information modification functions; (2) Schematic Diagram Construction Module: It is used for users to understand the process of constructing a schema, improve the schema, and fill in the corresponding event nodes. The schema construction module provides users with text input, event extraction, and schema representation functions. (3) Chemical event prediction module: The user takes the event context and candidate events as input to the system. The event prediction module uses the event prediction algorithm to calculate the similarity between the event context and multiple candidate events to obtain the prediction result. The prediction result is the probability of each candidate event occurring. The system visualizes the event prediction result and displays the event context and prediction result to the user in a graphical manner. Users can query the event graph based on event elements, and the system will return the nodes containing the event elements and other nodes connected to the nodes; (4) System management module: mainly used by administrators to manage server resources and ensure normal use of users, including user query, system monitoring and user management functions.