Similarity calculation and extraction method of petrochemical accident investigation report based on fault tree and event diagram

Abstract

The present invention provides a similarity calculation and extraction method for petrochemical accident investigation reports based on fault trees and event graphs, including constructing a fault tree for chemical accident investigation reports; grammatical analysis of petrochemical accident investigation reports; constructing an event graph model; a user inputs search information, and a server extracts petrochemical accident investigation reports similar to the search content. The present invention performs text analysis on petrochemical accident investigation reports and fault tree structures, constructs an event graph based on a BiGRU-Attention model, and extracts similar accident investigation reports through similarity calculations between search information and event graphs. The present invention uses event graphs to display the dynamic evolution process of complex accidents in an intuitive form, and uses the system extraction of event relations of the BiGRU-Attention model to improve the drawn event graphs, improve the matching degree of similar retrieval calculations, and provide support for intelligent push of accident analysis and emergency response plans.

Description

Technical Field

[0001] The present invention relates to the technical field of similarity calculation and extraction of accident investigation reports in the petrochemical industry, and in particular to a similarity calculation and extraction method of petrochemical accident investigation reports based on a fault tree and a causal graph. Background Art

[0002] In recent years, serious and serious accidents involving hazardous chemicals have occurred frequently in the chemical industry. Chemical accidents have brought huge economic losses to companies and individuals, caused irreparable damage to the environment, and even caused heavy casualties. Therefore, accurate and efficient retrieval of similar petrochemical accident investigation reports is of great significance for realizing the study of accident evolution laws, accident prevention, and intelligent recommendation of emergency response plans.

[0003] The existing technology mostly uses the matching of keywords or keyword groups to retrieve similar accident investigation reports. However, if exact matching is used, there will be fewer similar accident reports. If fuzzy matching is used, there will be more similar accident reports, and the accuracy will decrease. Patent CN106502996A proposes a semantic matching judicial document retrieval method based on term frequency-inverse word frequency (TF-IDF). The accuracy of retrieval is improved by adjusting the weights of words in documents and the entire text set. It is relatively efficient, but when applied to the petrochemical industry, it is not effective for long text terms such as accident investigation reports, and the evolution process and logical laws of events cannot be reflected. Patent CN105930473A proposes a similar document retrieval method based on random forest technology. By constructing a feature tree and training it with random forest technology, a feature weight tree is obtained, and a similarity matrix is ​​generated according to the query conditions. However, the evolution process of petrochemical accidents is complex, the nodes of the tree are unclear, and the use effect is not good. CN108038091A proposes a similarity calculation and retrieval method and system for judicial document cases based on a causal graph. Similar cases are found by matching retrieval information with the causal graph. This method is accurate and fast and is based on a complete causal graph. However, the petrochemical industry does not have such an application yet, nor does it have a comprehensive causal graph for petrochemical accidents. Moreover, petrochemical accidents are complex and varied, with a variety of causes. It is difficult to quickly and accurately find highly similar accident investigation reports.

[0004] Therefore, in order to quickly and efficiently extract similar petrochemical accident cases, implement targeted intelligent push, and provide guidance for emergency response to accidents, it is necessary to develop a similarity calculation and extraction method for accident investigation reports suitable for the petrochemical industry. Summary of the invention

[0005] In order to quickly and efficiently extract similar petrochemical accident cases, implement targeted intelligent push, and provide guidance for emergency handling of accidents, the present invention provides a similarity calculation and extraction method for petrochemical accident investigation reports based on fault trees and causal graphs.

[0006] The present invention adopts the following technical solutions:

[0007] A similarity calculation and extraction method for petrochemical accident investigation reports based on fault tree and event graph is proposed. Fault tree and event graph are constructed, and similar accident investigation reports are extracted through similarity calculation between retrieval information and event graph.

[0008] Preferably, the method comprises:

[0009] Step 1: Construct a fault tree for chemical accident investigation report;

[0010] Step 2: Syntax analysis of petrochemical accident investigation report;

[0011] Step 3: Construct a story-telling model;

[0012] Step 4: The user inputs search information, and the server extracts petrochemical accident investigation reports similar to the search content.

[0013] Preferably, step 1 specifically includes:

[0014] Step 1.1: Extract major chemical accidents at home and abroad in recent years from the existing accident investigation platform data, and select the accident investigation analysis report text;

[0015] Step 1.2: Perform failure analysis on the accident investigation and analysis report from top to bottom, use Boolean logic to combine low-level events, trace the necessary leading events that lead to the established consequences, establish the fault logic path from the top-level unexpected events to all possible underlying basic causes, and construct the accident fault tree.

[0016] Preferably, the accident investigation and analysis report texts selected in step 1.1 are complete.

[0017] Preferably, step 2 specifically includes:

[0018] Based on the fault tree and using big data technology, the TextRank algorithm is used to perform grammatical analysis on a large amount of text in the petrochemical accident investigation report, construct the structural features of the events that caused the accident, classify them into different element types and labels, and generate descriptive attributes of the events.

[0019] Preferably, the process of performing grammatical analysis using the TextRank algorithm is:

[0020] First, each sentence in the accident investigation report is segmented and POS tagged. Then, other words except the specified POS are filtered out, such as accident attribute words such as fire, explosion, leakage, etc., and words that appear in the stop word list and words with a length of less than 2 are filtered out.

[0021] Secondly, select a word from the remaining words in a loop, and combine it with the four words behind it into four edges, and the weight of each edge is 1. When this edge appears again later, the weight is increased by 1, and the TextRank algorithm is applied to iterate the weight of each node until convergence;

[0022] Finally, the Rank values ​​in the results are sorted to analyze and annotate the syntax of the accident investigation report.

[0023] Preferably, the descriptive attributes of an event include four elements: participants, trigger words, time, and location of the event.

[0024] Preferably, step 3 specifically includes:

[0025] The trigger words are classified into text sentiment based on the BiGRU-Attention model, and corresponding weights are assigned to the extracted deep information of the trigger word text. The text sentiment polarity of the trigger word text features with different weights is classified to realize the weight relationship between events and trigger words. Finally, the incident story map is extracted and generated to study the inherent attributes and evolution characteristics of accidents and construct an incident story map model.

[0026] Preferably, the process of text sentiment classification of trigger words based on the BiGRU-Attention model is:

[0027] Firstly, the Bidirectional Gated Recurrent (BiGRU) neural network layer is used to extract the deep information of the trigger word text;

[0028] Secondly, the attention layer is used to assign corresponding weights to the extracted deep information of the trigger word text, and the trigger word text features with different weights are put into the softmax function layer for text sentiment polarity classification to realize the weight relationship between events and trigger words.

[0029] Preferably, step 4 specifically includes:

[0030] The user inputs the search information into the server. The server extracts keywords based on the search information input by the user, and performs matching calculations with the causal graph in the database. The server returns petrochemical accident investigation reports similar to the search information according to the degree of matching, and extracts petrochemical accident investigation reports similar to the search content.

[0031] Preferably, the search information is one or more words, a sentence or a paragraph.

[0032] Preferably, the user can re-add and modify the search information, and the server will recalculate and return the revised search results until the best match is achieved.

[0033] The present invention has the following beneficial effects:

[0034] The present invention performs text analysis on petrochemical accident investigation reports and fault tree structures, constructs a causal graph based on the BiGRU-Attention model, and extracts similar accident investigation reports by calculating the similarity between retrieval information and the causal graph. The present invention uses the causal graph to display the dynamic evolution process of complex accidents in an intuitive form, and uses the systematic extraction of event relations of the BiGRU-Attention model to improve the drawn causal graph, improve the matching degree of similar retrieval calculation, and provide support for the intelligent push of accident analysis and emergency response plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention.

[0036] Figure 2 This is an example of the accident fault tree of Example 2.

[0037] Figure 3 This is an example of a cause-and-effect diagram of the tank area fire and explosion accident in Example 2.

[0038] Figure 4 This is an example of the accident fault tree of Example 3.

[0039] Figure 5 This is an example of the story diagram of the tank area poisoning incident in Example 3. DETAILED DESCRIPTION

[0040] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

[0041] Example 1

[0042] A similarity calculation and extraction method for petrochemical accident investigation reports based on fault tree and event graph is proposed. Fault tree and event graph are constructed, and similar accident investigation reports are extracted through similarity calculation between retrieval information and event graph.

[0043] Specifically include:

[0044] Step 1: Construct a fault tree for chemical accident investigation report;

[0045] Step 1.1: Extract major chemical accidents at home and abroad in recent years from the existing accident investigation platform data, and select accident investigation and analysis report texts; among them, the selected accident investigation and analysis report texts are all complete.

[0046] Step 1.2: Perform failure analysis on the accident investigation and analysis report from top to bottom, use Boolean logic to combine low-level events, trace the necessary leading events that lead to the established consequences, establish the fault logic path from the top-level unexpected events to all possible underlying basic causes, and construct the accident fault tree.

[0047] Step 2: Syntax analysis of petrochemical accident investigation report;

[0048] Based on the fault tree and using big data technology, the TextRank algorithm is used to perform grammatical analysis on a large amount of text in the petrochemical accident investigation report, construct the structural characteristics of the events that caused the accident, classify them into different element types and labels, and generate four elements of event participants, trigger words, time, and location as the descriptive attributes of the event.

[0049] The process of grammatical analysis using the TextRank algorithm is as follows:

[0050] Each sentence in the accident investigation report is segmented and POS tagged, and other words except the specified POS are filtered out, such as accident attribute words such as fire, explosion, and leakage, and words that appear in the stop word list and words with a length of less than 2 are filtered out. Then, a word is selected from the remaining words in a loop, and it is combined with the following 4 words into 4 edges, and the weight of each edge is 1. When this edge appears again later, the weight is added by 1 on the basis, and the TextRank algorithm is applied to iterate the weight of each node until convergence. Finally, the Rank values ​​in the results are sorted to achieve grammatical analysis and annotation of the accident investigation report.

[0051] Step 3: Construct a story-telling model;

[0052] Based on the BiGRU-Attention model, the trigger words are classified by text sentiment, the extracted deep information of the trigger words is assigned corresponding weights, the trigger words with different weights are classified by text sentiment polarity, the weight relationship between events and trigger words is realized, the inherent attributes and evolution characteristics of accidents are studied, and the incident story map model is constructed. It provides the necessary data foundation for the retrieval of similar accident reports and can effectively improve the efficiency of accident analysis.

[0053] The process of text sentiment classification of trigger words based on the BiGRU-Attention model is as follows:

[0054] Firstly, the Bidirectional Gated Recurrent (BiGRU) neural network layer is used to extract the deep information of the trigger word text. Secondly, the attention layer is used to assign corresponding weights to the extracted deep information of the trigger word text. The trigger word text features with different weights are put into the softmax function layer for text sentiment polarity classification, realizing the weight relationship between events and trigger words.

[0055] Step 4: The user inputs search information, and the server extracts petrochemical accident investigation reports similar to the search content.

[0056] The user inputs the search information into the server. The server extracts keywords based on the search information input by the user, and performs matching calculations with the causal graph in the database. The server returns petrochemical accident investigation reports similar to the search information according to the degree of matching, and extracts petrochemical accident investigation reports similar to the search content.

[0057] The search information is one or more words, a sentence or a paragraph, etc.

[0058] Users can add and modify search information, and the server will recalculate and return the revised search results until the best match is achieved.

[0059] Example 2

[0060] The fire and explosion accident in the tank area is used as an example to illustrate.

[0061] This paper collects fire and explosion accidents in tank areas at home and abroad in recent years, and conducts deductive failure analysis on the accident investigation and analysis reports from top to bottom according to the accident investigation reports. It uses Boolean logic to combine low-order events, and by tracing the necessary leading events that lead to fire and explosion accidents in tank areas, it establishes a fault logic path from the top-level unexpected events to all possible underlying basic causes, and constructs a fault tree for tank area fire and explosion accidents.

[0062] During implementation, the explosion and fire of the No. 5 oil tank as a result of the accident was taken as the top event. The leading event leading to the explosion and fire was that the oil and gas on the top of the No. 5 tank reached the explosion limit and encountered an ignition source to cause an explosion and fire. The oil and gas that reached the explosion limit were formed by the mixture of oil and gas and air accumulated on the tank top. The accumulation of oil and gas on the tank top was due to cracks in the prefabricated arch plate on the tank top and pressure relief in the vents, which led to the accumulation of oil and gas leakage. The ignition source that caused the accident was lightning sparks. The thunderstorm weather and the unreliable electrical connection of the tank body steel bars and metal parts caused sparks. The unreliable electrical connection of the tank body steel bars and metal parts mainly included the inability to weld the electrical contact of the oil tank anti-induction lightning shielding net, and the leakage of steel bars of the concrete oil tank body structure and tank top facilities. The specific accident fault tree is as follows Figure 2 shown.

[0063] The TextRank algorithm is used to perform grammatical analysis on the text of the collected tank area fire and explosion accident reports, and the four elements of participants, trigger words, time, and location of the generated events are extracted. The BiGRU-Attention model is used to perform sentiment classification on simple event relationships in the accident, and a story map of the tank area fire and explosion incident is generated.

[0064] In the implementation, the collected tank area fire and explosion accident report text is divided into individual sentences and a vector representation is found for each sentence. The similarity between the sentence vectors is calculated and stored in a matrix. Then the similarity matrix is ​​used for TextRank calculation to obtain phrases such as explosive mixed gas, ignition source, casualties, economic losses, environmental damage, social impact, failure of anti-static measures, undesigned anti-static facilities, and excessive feed flow rate. These extracted phrases represented by vectors are applied to the BiGRU layer in the BiGRU-Attention model, and the output vector of the BiGRU layer is placed in the Attention layer. The corresponding weights are assigned to the extracted deep-level information of the text to further highlight the key information of the text and generate a story map of the tank area fire and explosion, such as Figure 3 shown.

[0065] The user enters search information on the client, such as "storage tank fire and explosion accident caused by lightning strike", and the search information is submitted to the server. The server extracts keywords based on the search information entered by the user: lightning strike, storage tank, fire and explosion, and matches them with the elements of the existing event graph. According to the degree of matching, the petrochemical accident investigation report similar to the search information is returned.

[0066] Example 3

[0067] The following is an example of a poisoning accident in a tank area.

[0068] Tank poisoning accidents at home and abroad in recent years are collected. According to the accident investigation report, a deductive failure analysis is performed on the accident investigation and analysis report from top to bottom. The Boolean logic is used to combine low-order events. By tracing the necessary leading events that lead to tank poisoning accidents, a fault logic path from the top-level unexpected events to all possible underlying basic causes is established, and a tank poisoning accident fault tree is constructed.

[0069] In the implementation, oil and gas poisoning is the top event. The leading event leading to oil and gas poisoning is that the oil and gas concentration exceeds the standard and the operators are in the area where the oil and gas concentration exceeds the standard. The oil evaporates in the high-temperature pump room, and the doors and windows of the pump room are closed and the ventilation is poor, which leads to the oil and gas concentration exceeding the standard. The possible source of the oil is the oil in the oil receiving basin or the oil leakage of the pump; and the operators are in this environment because of their low personal protection awareness, violation of operating regulations or lack of basic safety knowledge. The specific accident fault tree is as follows: Figure 4 shown.

[0070] The TextRank algorithm is used to perform grammatical analysis on the text of the collected tank area poisoning accident reports, and the four elements of participants, trigger words, time, and location of the generated events are extracted. The BiGRU-Attention model is used to perform sentiment classification on simple event relationships in the accident, and a story map of the tank area poisoning incident is generated.

[0071] In the implementation, the collected tank area poisoning accident report text is divided into individual sentences and a vector representation is found for each sentence. The similarity between the sentence vectors is calculated and stored in a matrix. Then the similarity matrix is ​​used for TextRank calculation to obtain phrases such as oil and gas accumulation, poor ventilation, failure to take safety measures, problems with safety measures, low individual protection awareness, violation of operating regulations, failure to conduct anti-poisoning education before operation, oil leakage at flange connection, oil outflow from connecting pipe, oil leakage from pump, oil storage in the oil receiving basin, lax closure of the air intake branch valve of the oil pipeline, and oil and gas accumulated in the pipeline leaking into the oil tank. These extracted phrases represented by vectors are applied to the BiGRU layer in the BiGRU-Attention model, and the output vector of the BiGRU layer is placed in the Attention layer. The corresponding weights are assigned to the extracted deep information of the text to further highlight the key information of the text and generate a story map of the tank area poisoning incident, such as Figure 5 shown.

[0072] The user enters search information on the client, such as "oil and gas poisoning accident caused by pump leakage", and the search information is submitted to the server. The server extracts keywords based on the search information entered by the user: pump leakage, oil and gas, poisoning, and matches them with the elements of the existing event graph. According to the degree of matching, the petrochemical accident investigation report similar to the search information is returned.

[0073] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A similarity calculation and extraction method for petrochemical accident investigation reports based on fault tree and event diagram, characterized in that: Construct fault trees and event graphs, and extract similar accident investigation reports by calculating the similarity between retrieval information and event graphs; The method includes: Step 1: Construct a fault tree for chemical accident investigation report; Step 2: Syntax analysis of petrochemical accident investigation report; Step 3: Construct a story-telling model; Step 4: The user inputs search information, and the server extracts petrochemical accident investigation reports similar to the search content. Step 2 specifically includes: Based on the fault tree and using big data technology, the TextRank algorithm is used to perform grammatical analysis on a large amount of text in the petrochemical accident investigation report, construct the structural features of the events that caused the accident, classify them into different element types and labels, and generate descriptive attributes of the events; The process of grammatical analysis using the TextRank algorithm is as follows: First, each sentence in the accident investigation report is segmented and POS tagged. Then, other words except the specified POS are filtered out, and words that appear in the stop word list and words with a length of less than 2 are filtered out. Secondly, select a word from the remaining words in a loop, and combine it with the four words behind it into four edges, and the weight of each edge is 1. When this edge appears again later, the weight is increased by 1, and the TextRank algorithm is applied to iterate the weight of each node until convergence; Finally, the Rank values ​​in the results are sorted to analyze and annotate the syntax of the accident investigation report.

2. The method for similarity calculation and extraction of petrochemical accident investigation reports based on fault tree and event diagram according to claim 1 is characterized in that: Step 1 specifically includes: Step 1.1: Extract major chemical accidents at home and abroad in recent years from the existing accident investigation platform data, and select the accident investigation analysis report text; Step 1.2: Perform failure analysis on the accident investigation and analysis report from top to bottom, use Boolean logic to combine low-level events, trace the necessary leading events that lead to the established consequences, establish the fault logic path from the top-level unexpected events to all possible underlying basic causes, and construct the accident fault tree.

3. The similarity calculation and extraction method of petrochemical accident investigation report based on fault tree and event diagram according to claim 2 is characterized in that: The accident investigation and analysis report texts selected in step 1.1 are all complete.

4. The method for similarity calculation and extraction of petrochemical accident investigation reports based on fault tree and event diagram according to claim 1 is characterized in that: The descriptive attributes of an event include four elements: participants, trigger words, time, and location.

5. The method for similarity calculation and extraction of petrochemical accident investigation reports based on fault tree and event diagram according to claim 1 is characterized in that: Step 3 specifically includes: The trigger words are classified into text sentiment based on the BiGRU-Attention model, and corresponding weights are assigned to the extracted deep information of the trigger word text. The text sentiment polarity of the trigger word text features with different weights is classified to realize the weight relationship between events and trigger words. Finally, the incident story map is extracted and generated to study the inherent attributes and evolution characteristics of accidents and construct an incident story map model.

6. The method for similarity calculation and extraction of petrochemical accident investigation reports based on fault tree and event diagram according to claim 5 is characterized in that: The process of text sentiment classification of trigger words based on the BiGRU-Attention model is as follows: Firstly, a bidirectional gated recurrent neural network layer is used to extract features from the deep information of the trigger word text; Secondly, the attention layer is used to assign corresponding weights to the extracted deep information of the trigger word text, and the trigger word text features with different weights are put into the softmax function layer for text sentiment polarity classification to realize the weight relationship between events and trigger words.

7. The method for similarity calculation and extraction of petrochemical accident investigation reports based on fault tree and event diagram according to claim 1 is characterized in that: Step 4 specifically includes: The user inputs the search information into the server. The server extracts keywords based on the search information input by the user, and performs matching calculations with the causal graph in the database. The server returns petrochemical accident investigation reports similar to the search information according to the degree of matching, and extracts petrochemical accident investigation reports similar to the search content.

8. The method for similarity calculation and extraction of petrochemical accident investigation reports based on fault tree and event diagram according to claim 1 is characterized in that: The search information is one or more words, a sentence or a paragraph.

9. The method for similarity calculation and extraction of petrochemical accident investigation reports based on fault tree and event diagram according to claim 1 is characterized in that: Users can add and modify search information, and the server will recalculate and return the revised search results until the best match is achieved.

Images