Method and system for processing and analyzing survey report of water traffic accident

Abstract

The invention belongs to the field of water traffic, and discloses a water traffic accident investigation report processing and analyzing method and system, and the method comprises the steps: carrying out the pre-training of a language model through the text data of a water traffic accident, and obtaining a keyword extraction model; labeling the output of the keyword extraction model by using a conditional random field; based on the keyword extraction model, extracting accident causes and risk points of the to-be-analyzed water traffic accident report; calculating the correlation degree between the accident causes and between the risk points, and constructing a coupling relation matrix; screening out accident cause strong coupling relation pairs and risk point strong coupling relation pairs; performing word frequency analysis on the accident causes and the risk points to obtain importance degrees of the accident causes and the risk points; performing scene classification on a to-be-analyzed water traffic accident report, and constructing an accident scene set; the analysis efficiency is greatly improved, the accuracy is remarkably enhanced, the quantitative analysis is more accurate, and the result reliability is high.

Description

Technical Field

[0001] The present invention relates to the field of water transportation, and in particular to a method and system for processing and analyzing water transportation accident investigation reports. Background Art

[0002] In the field of water transportation, the analysis of accident investigation reports used to rely mainly on manual methods, with expert groups interpreting and evaluating them. This traditional method has many disadvantages, such as consuming a lot of manpower and time costs, and the analysis results are easily affected by the subjective factors of experts, lacking objectivity and consistency. With the development of information technology, some studies have begun to try to use some simple data statistics and analysis methods, but for unstructured accident report text data, these methods are difficult to deeply explore the key information and potential connections.

[0003] In recent years, artificial intelligence technology has been gradually applied in various fields, but it is still in the exploratory stage in the processing of water traffic accident investigation reports. Although some related studies have tried to use machine learning algorithms, there are deficiencies in the accuracy and adaptability of the models, and they fail to fully meet actual needs. For example, some early text analysis models often fail to accurately extract the causes and risk points of accidents when processing accident reports with numerous professional terms and complex text structures, and it is also difficult to effectively classify and analyze accident scenes.

[0004] The main defects of existing technologies are: manual analysis is inefficient and cannot quickly process a large number of accident reports; qualitative analysis is mostly used, lacking precise quantitative analysis, making it difficult to accurately measure the degree of risk; the ability to mine complex semantics and implicit information in accident reports is limited, resulting in incomplete identification of accident causes and risk points; the results obtained by different analysis methods and personnel vary greatly, lacking reliability and stability.

[0005] Therefore, how to provide a method and system for processing and analyzing water traffic accident investigation reports is a problem that needs to be solved urgently. Summary of the invention

[0006] The embodiment of the present invention provides a method and system for processing and analyzing water traffic accident investigation reports to solve the problems in the prior art of low efficiency of manual analysis, lack of accurate quantitative analysis, incomplete identification of accident causes and risk points, and lack of reliability and stability.

[0007] In order to have a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not a general review, nor is it intended to identify key / important components or to describe the scope of protection of these embodiments. Its only purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0008] According to a first aspect of an embodiment of the present invention, a method for processing and analyzing a water traffic accident investigation report is provided.

[0009] In one embodiment, a method for processing and analyzing a water traffic accident investigation report includes:

[0010] The language model is pre-trained using the water traffic accident text data, so that the language model can obtain the semantic features and language patterns in the water traffic accident text data, and a keyword extraction model is obtained; the output of the keyword extraction model is annotated using the conditional random field;

[0011] Based on the keyword extraction model, the causes of accidents and risk points are extracted from the water traffic accident reports to be analyzed; the similarity algorithm is used to calculate the degree of correlation between the causes of accidents and the risk points, and a coupling relationship matrix is ​​constructed; the coupling relationship matrix and the coupling degree threshold are used to screen out the strong coupling relationship pairs of accident causes and risk points;

[0012] According to the word frequency statistical analysis technology, the word frequency analysis of the accident causes and risk points is carried out to obtain the importance of the accident causes and risk points;

[0013] Using the topic model and combining the causes of accidents, the water traffic accident reports to be analyzed are classified into scenarios and an accident scenario set is constructed.

[0014] In one embodiment, before pre-training the language model using the water traffic accident text data, the method further includes:

[0015] Collect water traffic accident investigation reports and convert them into text data in a unified format;

[0016] Clean the text data in the water traffic accident investigation report and remove noise data, irrelevant symbols and repeated information;

[0017] The text data in the water traffic accident investigation report is subjected to lexical and syntactic analysis, and the professional terms are standardized to obtain the water traffic accident text data.

[0018] In one embodiment, labeling the output of the keyword extraction model using the conditional random field includes:

[0019] Construct the characteristic function and transfer function of the conditional random field; use the conditional random field and combine it with expert experience to annotate the output of the keyword extraction model.

[0020] In one embodiment, a similarity algorithm is used to calculate the correlation between the causes of accidents and the risk points, and a coupling relationship matrix is ​​constructed, including:

[0021] The cosine similarity calculation formula is used to calculate the coupling degree between the causes of accidents and between the risk points, and the coupling degree is saved in the form of a matrix.

[0022] In one embodiment, using the coupling relationship matrix and the coupling degree threshold, the accident cause strong coupling relationship pairs and the risk point strong coupling relationship pairs are screened out, including:

[0023] A coupling degree threshold is set in advance, and a coupling connection is established between two keywords that exceed the coupling degree threshold. The coupling connection status is displayed using a chord diagram, and the number of coupling connections for each keyword is statistically obtained. The keywords include accident causes and risk points.

[0024] According to the coupling connections established between the keywords, the strongly coupled relationship pairs of accident causes and the strongly coupled relationship pairs of risk points are obtained.

[0025] In one embodiment, the word frequency analysis is performed on the causes of accidents and risk points according to the word frequency statistical analysis technology, and the importance of the causes of accidents and risk points is obtained, including:

[0026] Through the accident cause list and risk point list, the word frequency statistics of accident cause and risk point are carried out to obtain the word frequency parameters of each accident cause and risk point;

[0027] Arrange in descending order according to the word frequency parameter and display it in a bar graph; determine the coordinates of the corresponding keyword in the scatter plot based on the statistical number of coupling connections for each keyword, and determine the importance of the corresponding accident causes and risk points based on the coordinates of each keyword in the scatter plot.

[0028] In one embodiment, the topic model is used to classify the water traffic accident reports to be analyzed in combination with the causes of the accident, and an accident scene set is constructed including:

[0029] Determine the training parameters of the topic model, including the frequency amplification coefficient, the number of training topics, and the number of topic word displays; amplify the frequency of the cause of the accident and remove stop words;

[0030] Through the perplexity experiment, the required number of scenes and the magnification factor are selected and observed; the topic model is trained using the corpus to obtain a trained topic model;

[0031] Combining the number of scenes, the magnification factor and the trained topic model, the accident scene set is obtained.

[0032] In one embodiment, when the frequency of the word causing the accident is amplified, the frequency amplification formula is: ;

[0033] In the formula, Indicates the updated word The word frequency, represents its original word frequency;

[0034] Indicates the frequency increase coefficient of a certain word, represents the set of accident causes, and k represents the words in the accident report.

[0035] In one embodiment, constructing the accident scenario set includes:

[0036] Classify and name each scene in the accident scene set; calculate the coupling degree between the accident cause and the risk point in a scene through the correlation analysis of scenes, accident causes and risk points, establish the coupling connection through the preset threshold, and display it through the Sankey diagram.

[0037] According to a second aspect of an embodiment of the present invention, a water traffic accident investigation report processing and analysis system is provided.

[0038] In one embodiment, the water traffic accident investigation report processing and analysis system includes:

[0039] The keyword extraction model training module is used to pre-train the language model using the water traffic accident text data, so that the language model can obtain the semantic features and language patterns in the water traffic accident text data, and obtain the keyword extraction model; the output of the keyword extraction model is annotated using the conditional random field;

[0040] The coupling analysis module is used to extract the causes of accidents and risk points from the water traffic accident reports to be analyzed based on the keyword extraction model; the similarity algorithm is used to calculate the degree of correlation between the causes of accidents and the risk points, and a coupling relationship matrix is ​​constructed; the coupling relationship matrix and the coupling degree threshold are used to screen out the strong coupling relationship pairs of accident causes and risk points;

[0041] Importance analysis module, used to perform word frequency analysis on accident causes and risk points based on word frequency statistical analysis technology to obtain the importance of accident causes and risk points;

[0042] The scene classification module is used to utilize the topic model and combine the causes of the accident to classify the water traffic accident reports to be analyzed and construct an accident scene set.

[0043] According to a third aspect of an embodiment of the present invention, a computer device is provided.

[0044] In some embodiments, the computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0045] According to a fourth aspect of embodiments of the present invention, a computer-readable storage medium is provided.

[0046] In one embodiment, the computer-readable storage medium stores a computer program, and the computer program implements the steps of the above method when executed by a processor.

[0047] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0048] (1) Greatly improved analysis efficiency: The equipment of the present invention can quickly process a large number of accident reports through automated data collection and intelligent analysis algorithms. Compared with manual analysis, it greatly shortens the analysis cycle and improves the timeliness and response speed of water traffic safety management.

[0049] (2) Significantly enhanced accuracy: With the help of advanced large language models and optimized algorithms, the identification of accident causes and risk points is more accurate and comprehensive, reducing omissions and misjudgments, and can more accurately mine key information in accident reports, providing a reliable basis for risk assessment. For example, in the analysis of complex accident scenarios, it can accurately identify the coupling effects of multiple factors, while existing technologies may only be able to discover some obvious factors.

[0050] (3) Quantitative analysis is more accurate: Through quantitative analysis methods such as cosine similarity, not only can the correlation between accident causes and risk points be determined, but also their coupling degree and risk importance can be accurately measured, providing an effective means for the quantitative assessment of water traffic risks and making up for the shortcomings of existing technologies in quantitative analysis.

[0051] (4) High reliability of results: The present invention is based on a large amount of accident data and a scientific algorithm model. The analysis results have good stability and consistency and are not affected by human subjective factors. They can provide reliable decision-making support for water traffic safety management departments and shipping companies. Compared with traditional methods and some simple data analysis techniques, the reliability has made a qualitative leap.

[0052] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0054] Figure 1 is a flow chart of a method for processing and analyzing a water traffic accident investigation report according to an exemplary embodiment;

[0055] Figure 2 is a principle block diagram of a water traffic accident investigation report processing and analysis system according to an exemplary embodiment;

[0056] Figure 3 is a structural schematic diagram of a computer device according to an exemplary embodiment;

[0057] Figure 4 is a diagram showing the cause of an accident according to an exemplary embodiment;

[0058] Figure 5 is a risk factor display diagram according to an exemplary embodiment;

[0059] Figure 6 is an importance display diagram according to an exemplary embodiment;

[0060] Figure 7 is a diagram showing a perplexity calculation result according to an exemplary embodiment. DETAILED DESCRIPTION

[0061] The following description and accompanying drawings fully illustrate the specific embodiments of this article so that those skilled in the art can practice them. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. The scope of the embodiments of this article includes the entire scope of the claims, as well as all available equivalents of the claims. Herein, the terms "first", "second", etc. are only used to distinguish one element from another, without requiring or implying any actual relationship or order between these elements. In fact, the first element can also be called the second element, and vice versa. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that the structure, device or equipment including a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also include elements inherent to such structure, device or equipment. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the structure, device or equipment including the elements. Each embodiment is described in a progressive manner herein, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the embodiments can be referred to each other.

[0062] The terms "longitudinal", "lateral", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc. in this document indicate the orientation or position relationship based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing this document and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In the description of this document, unless otherwise specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a mechanical connection or an electrical connection, it can also be the internal communication of two elements, it can be a direct connection, or it can be an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0063] As used herein, the term "plurality" means two or more than two, unless otherwise specified.

[0064] In this document, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, U / V means: U or V.

[0065] In this article, the term "and / or" is a description of the association relationship of objects, indicating that three relationships may exist. For example, U and / or V means: U or V, or, U and V.

[0066] It should be understood that, although the various steps in the flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the figure may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0067] Each module in the device or system of the present application can be implemented in whole or in part by software, hardware, or a combination thereof. The above modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to the above modules.

[0068] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0069] Figure 1An embodiment of a method for processing and analyzing a water traffic accident investigation report of the present invention is shown.

[0070] In this optional embodiment, the method for processing and analyzing a water traffic accident investigation report includes:

[0071] S101. Pre-training a language model using water traffic accident text data, so that the language model acquires semantic features and language patterns in the water traffic accident text data, and obtains a keyword extraction model; and annotating the output of the keyword extraction model using a conditional random field.

[0072] S102. Based on the keyword extraction model, the causes of accidents and risk points are extracted from the water traffic accident reports to be analyzed; the similarity algorithm is used to calculate the degree of correlation between the causes of accidents and between the risk points, and a coupling relationship matrix is ​​constructed; the coupling relationship matrix and the coupling degree threshold are used to screen out the strong coupling relationship pairs of accident causes and the strong coupling relationship pairs of risk points.

[0073] S103. Perform word frequency analysis on the causes of the accident and the risk points according to word frequency statistical analysis technology to obtain the importance of the causes of the accident and the risk points.

[0074] S104. Using the topic model and combining the causes of the accident, the water traffic accident reports to be analyzed are classified into scenarios and an accident scenario set is constructed.

[0075] In this optional embodiment, before pre-training the language model using the water traffic accident text data, the method further includes:

[0076] Collect water traffic accident investigation reports and convert them into text data in a unified format.

[0077] The text data in the water traffic accident investigation report is cleaned and the noise data, irrelevant symbols and repeated information are removed.

[0078] The text data in the water traffic accident investigation report is subjected to lexical and syntactic analysis, and the professional terms are standardized to obtain the water traffic accident text data.

[0079] In this optional embodiment, labeling the output of the keyword extraction model using the conditional random field includes:

[0080] Construct the characteristic function and transfer function of the conditional random field; use the conditional random field and combine it with expert experience to annotate the output of the keyword extraction model.

[0081] In this optional embodiment, a similarity algorithm is used to calculate the degree of association between the causes of accidents and between the risk points, and a coupling relationship matrix is ​​constructed, including:

[0082] The cosine similarity calculation formula is used to calculate the coupling degree between the causes of accidents and between the risk points, and the coupling degree is saved in the form of a matrix.

[0083] In this optional embodiment, using the coupling relationship matrix and the coupling degree threshold, the accident cause strong coupling relationship pairs and the risk point strong coupling relationship pairs are screened out, including:

[0084] A coupling degree threshold is set in advance, and a coupling connection is established between two keywords that exceed the coupling degree threshold. The coupling connection status is displayed using a chord diagram, and the number of coupling connections for each keyword is statistically obtained. The keywords include accident causes and risk points.

[0085] According to the coupling connections established between the keywords, the strongly coupled relationship pairs of accident causes and the strongly coupled relationship pairs of risk points are obtained.

[0086] In this optional embodiment, the word frequency analysis is performed on the cause of the accident and the risk point according to the word frequency statistical analysis technology, and the importance of the cause of the accident and the risk point is obtained, including:

[0087] Through the accident cause list and risk point list, the word frequency statistics of accident causes and risk points are performed to obtain the word frequency parameters of each accident cause and risk point.

[0088] Arrange in descending order according to the word frequency parameter and display it in a bar graph; determine the coordinates of the corresponding keyword in the scatter plot based on the statistical number of coupling connections for each keyword, and determine the importance of the corresponding accident causes and risk points based on the coordinates of each keyword in the scatter plot.

[0089] In this optional embodiment, the topic model is used to classify the water traffic accident report to be analyzed in combination with the cause of the accident, and the accident scene set is constructed including:

[0090] Determine the training parameters of the topic model, including the word frequency amplification coefficient, the number of training topics and the number of topic word displays; amplify the word frequency of the cause of the accident and remove stop words.

[0091] Through the perplexity experiment, the required number of scenes and the amplification factor are selected and observed; the topic model is trained using the corpus to obtain a trained topic model.

[0092] Combining the number of scenes, the magnification factor and the trained topic model, the accident scene set is obtained.

[0093] In this optional embodiment, when the frequency of the word causing the accident is amplified, the frequency amplification formula is: ;

[0094] In the formula, Indicates the updated word The word frequency, represents its original word frequency; Indicates the frequency increase coefficient of a certain word, represents the set of accident causes, and k represents the words in the accident report.

[0095] In this optional embodiment, constructing the accident scenario set includes:

[0096] Classify and name each scene in the accident scene set; calculate the coupling degree between the accident cause and the risk point in a scene through the correlation analysis of scenes, accident causes and risk points, establish the coupling connection through the preset threshold, and display it through the Sankey diagram.

[0097] Figure 2 An embodiment of a water traffic accident investigation report processing and analysis system of the present invention is shown.

[0098] In this optional embodiment, the water traffic accident investigation report processing and analysis system includes:

[0099] The keyword extraction model training module 201 is used to pre-train the language model using the water traffic accident text data so that the language model can obtain the semantic features and language patterns in the water traffic accident text data and obtain the keyword extraction model; the output of the keyword extraction model is annotated using the conditional random field.

[0100] The coupling analysis module 202 is used to extract the causes of accidents and risk points from the water traffic accident report to be analyzed based on the keyword extraction model; use the similarity algorithm to calculate the degree of correlation between the causes of accidents and between the risk points, and construct a coupling relationship matrix; use the coupling relationship matrix and the coupling degree threshold to screen out the strong coupling relationship pairs of accident causes and the strong coupling relationship pairs of risk points.

[0101] The importance analysis module 203 is used to perform word frequency analysis on the accident causes and risk points based on word frequency statistical analysis technology to obtain the importance of the accident causes and risk points.

[0102] The scene classification module 204 is used to utilize the topic model and combine the causes of the accident to classify the water traffic accident reports to be analyzed and construct an accident scene set.

[0103] In order to facilitate understanding of the above technical solutions of the present invention, the above technical solutions of the present invention are further described below from the perspective of architecture and principle, as follows:

[0104] Before constructing a marine accident analysis framework, it is necessary to define some basic concepts, including the maritime transport system, maritime accidents, accident causes, accident scenarios, hazards, etc. The Maritime Transport System (MTS) is a complex social-technical environment (STE) system, which mainly includes three elements: human-ship environment. If the shore-side operation management agency and safety regulatory agency are included in the system, the main elements of the maritime transport system can be divided into four subsystems: human, ship, environment, and management. The human subsystem is composed of the crew on the ship, which is generally divided into the deck department and the engine department, including the captain, the first / second / third mate, the pilot, the duty driver / helmsman, the sailor, the chief engineer, the mechanic, etc.; the ship subsystem is mainly composed of the hull, the main engine, the steering gear, the electrical system, the air conditioning system, the fire protection system, etc.; the environmental subsystem includes natural environmental elements such as wind, waves, currents, and visibility, as well as navigation environment elements such as waterways, navigation marks, lighthouses, bridges, docks, and other ships; the management subsystem includes regulatory agencies, shipping companies, shipowners and other operating agencies, as well as shipyards and other ship design and production and construction agencies.

[0105] Maritime Accident refers to traffic accidents that occur to ships and floating facilities in the ocean, coastal waters and inland navigable waters, such as collision, stranding, water ingress, sinking, capsizing, hull damage, fire, explosion, main engine damage, cargo damage, crew casualties, marine pollution, etc. An accident is a special event formed by a series of conditions and events, which is generally distinguished by Accident or Incident. This invention mainly analyzes data that have entered statistics and formed accident investigation reports, so Accident is more often used to represent it.

[0106] Accident Cause (AC) refers to a series of events and conditions that lead to an accident. According to Heinrich's accident causation theory, the occurrence of an accident is a chain reaction of a series of events. The causes of marine accidents include unsafe human behavior, unsafe ship conditions, unsafe environmental factors and lack of management.

[0107] Accident Scenario (AS) refers to the entire life cycle of an accident, including before, during and after the accident. In this article, the accident scenario is defined as the situation that leads to the accident, that is, the set of accident causes: .

[0108] The term hazard has many definitions in the field of safety science, usually referring to the source of danger to people, property, and the environment. From the perspective of system theory, this paper defines system hazard (SH) as a collection of dangerous points (hazard, h) that may cause system collapse (accident): .

[0109] System hazards of the marine transport system include collision hazard, self-sinking hazard, grounding hazard, fire hazard, etc. according to different types of marine accidents. Danger points are the system entity components and potential factors that constitute system hazards, which can also be divided into four categories: people, ships, environment, and management. For example, in the case of collision hazards, human danger points include entities such as ship drivers, captains, sailors on duty, and potential factors such as knowledge reserves, safety awareness, mental state, and physiological state; ship danger points include entities such as engines, steering gear, navigation aids, and hulls; environmental danger points include entities such as wind, waves, currents, and navigation obstacles; management danger points include entities such as VTS centers, shipping companies, pilot stations, various management regulations, and potential factors such as safety culture and training.

[0110] Danger points are closely related to the causes of accidents, and the causes of accidents must include dangerous points. For example, neglecting lookout is an unsafe behavior that leads to maritime accidents, and the crew members responsible for lookout during the voyage of the ship are dangerous points. Danger points are interconnected and interact with each other, and they are constantly changing in a complex and dynamic way in the maritime transportation system. A certain dangerous point will trigger a risk event under specific time and space conditions, and may even become the cause of an accident.

[0111] The present invention relates to a technical device for processing and analyzing water traffic accident investigation reports using a large language model, and is mainly used in water traffic safety management, shipping enterprise risk prevention and control, etc.

[0112] The device of the present invention is mainly composed of a data acquisition and preprocessing module, a model training and optimization module, a risk coupling analysis module, a scene classification and recognition module and a result display and output module.

[0113] 1) Data collection and preprocessing module: responsible for collecting water traffic accident investigation reports from multiple channels such as maritime-related websites and shipping company databases, and converting them into text data in a unified format. Cleaning operations are performed on the text content in the report to remove noise data, irrelevant symbols and repeated information. At the same time, lexical and syntactic analysis is performed to standardize professional terms to lay the foundation for subsequent analysis. For example, the names of ship equipment that are expressed differently in different reports are unified into standard terms to facilitate system recognition and processing.

[0114] 2) Model training and optimization module: BERT (Bidirectional Encoder Representation from Transformers) and other advanced large language models are used as core algorithms. First, a large amount of water traffic accident text data is used to pre-train the BERT model so that it can learn the semantic features and language patterns in the accident report. On this basis, combined with expert experience and annotated data, the model is fine-tuned through methods such as Conditional Random Field (CRF) to further optimize the model's performance in accident cause and risk point extraction tasks. For example, for the description of specific accident causes such as "negligence in lookout" and "equipment failure", the model parameters are adjusted through expert annotation and feedback to improve its recognition accuracy.

[0115] 3) Risk coupling analysis module: Based on the trained model, the extracted accident causes and risk points are subjected to coupling degree analysis. The cosine similarity algorithm is used to calculate the degree of correlation between the factors, construct a coupling relationship matrix, and screen out strong coupling relationship pairs according to the set threshold. At the same time, the importance of risk points is evaluated in combination with the contextual information and statistical data in the accident report, and the key risk points are determined by comprehensively considering factors such as word frequency, occurrence location, and number of related factors. For example, in a certain accident report, if "radar failure" and "ship collision" frequently appear in the text at the same time and are closely related, they are identified as high-coupling risk points, and their importance is evaluated based on their word frequency and impact in multiple reports.

[0116] 4) Scene classification and identification module: The improved LDA (Latent Dirichlet Allocation) topic model is used to classify the accident reports. In the word segmentation process, a priori accident cause list is introduced, and the frequency of related words is amplified to enhance their weight in model training, so as to more accurately identify different accident scenes. Through training, multiple topic models are obtained, each topic represents an accident scene, and the topic words are the key accident causes and risk points in the scene. Further analyze the distribution patterns of accident causes and risk points in different scenes, as well as the conversion relationship and possibility between scenes. For example, based on the analysis of a large number of accident reports, "collision scenes caused by bad weather" and "accident scenes caused by ship equipment failures" are identified, and the action mechanism and relationship of various risk factors in different scenes are analyzed.

[0117] 5) Result display and output module: The results of risk coupling analysis and scenario classification and identification are displayed in intuitive charts, such as bar charts, pie charts, chord charts, scatter plots, and Sankey diagrams. Bar charts are used to compare the word frequencies of different accident causes and risk points, pie charts show the proportion of various factors in the whole, chord charts present the coupling relationship network between risk points, scatter plots combine word frequency and coupling connection number to reflect the importance distribution of risk points, and Sankey diagrams show the complex relationship between scenarios, accident causes, and risk points. At the same time, a detailed analysis report is generated, including accident overview, key causes, risk assessment results, and prevention recommendations, to provide users with a comprehensive decision-making reference.

[0118] The connection relationship and interaction process between each module are as follows: the data acquisition and preprocessing module transmits the processed text data to the model training and optimization module. The trained model analyzes and processes the input data, and sends the extracted accident causes and risk point information to the risk coupling analysis module and the scene classification and recognition module respectively. After completing their respective analysis tasks, these two modules summarize the results to the result display and output module, and finally present them to the user. The working process of the entire system is: first, the data acquisition and preprocessing module obtains and processes the accident report data, then the model training and optimization module performs model training and optimization, and then uses the trained model to analyze the data. The risk coupling analysis module and the scene classification and recognition module work in parallel, and finally the result display and output module displays the analysis results in an intuitive form and outputs them in the form of a report.

[0119] The specific analysis is as follows:

[0120] The extraction of accident causes and danger points is a keyword extraction task. The extraction process mainly includes three steps: 1. Select the Chinese BERT pre-training model; 2. Sequence labeling of the training set based on expert experience and CRF, and on this basis, fine-tune the BERT pre-training model to obtain a keyword extraction model.

[0121] 1. BERT-Embedding, an embedding tool based on the BERT model

[0122] The core of the BERT pre-training model is to complete BERT-Embedding. For ease of understanding, a description of the cause of the accident in the corpus is used to explain the BERT-Embedding process: the input layer is the original text information, where [CLS] is a classification identifier, [SEP] is a discontinuity identifier, and the sentence is randomly divided into different word blocks; the embedding layer includes token embedding, segment embedding, and position embedding. Through Embedding, the original description "The captain and the sailor were on duty at the bridge, but failed to maintain an uninterrupted regular lookout" is transformed into a word tensor sequence with three layers of embedding. Since the true semantic information of each word is related to its context, word order, and the semantic set it possesses, a word tensor with high accuracy can be obtained after word embedding and then training with a large amount of text data. This part of the training task has been pre-selected, thus saving the high training cost. Fine-tuning training on specific corpus can be directly performed based on this pre-trained model to complete downstream tasks.

[0123] The present invention is oriented to accident investigation report data, which is highly professional and small in size, so the pre-trained Chinese BERT model (with 12 layers, 768 hidden layers, 12 heads, and 110M parameters) is selected as the pre-trained model. The pre-trained word embedding can more accurately describe the semantic relationship between characters, and on this basis complete the downstream task of keyword extraction.

[0124] 2. Data labeling based on CRF

[0125] CRF is a classic method for sequence labeling. Its core idea is to treat each character in the sequence as a whole when labeling the sequence, considering the relationship between the previous and the next, rather than each independent point. The labeling results of each point have a certain dependency, and training is performed in units of paths. Therefore, through training, the model can understand the regularity of the output sequence in addition to understanding the text, and can better complete the task of extracting key information.

[0126] When using CRF for keyword extraction, you need to define the feature function: ;

[0127] in, Represents the input sequence, which is a character i in the sequence. If the prediction is correct, the output value is 1, otherwise it is 0. Also called an emission function or node function, it only determines the prediction result of a single character and does not involve the dependency between labels. , if the conditional probability distribution of the predicted sequence Y satisfies the Markov property of the random variable Y, then Only its preceding label Related, and together with the input X, form a conditional random field. Therefore, the following transition function needs to be added: ;

[0128] Where X represents the observation sequence, i represents a character in the observation sequence, and Represents the predicted labels of the i-th and i-1-th characters respectively. If the current label prediction result meets the expected transfer rule of the previous label, the feature value is 1; otherwise it is 0. The sequence score predicted by CRF is calculated using the following formula: ;

[0129] in, represents the result of the prediction feature function calculation of the i-th character in the sequence input to the CRF layer, It represents the calculation result of the transfer function of the character after this character. Since there are many possibilities for the predicted sequence, and only one of them is the most correct, all possible sequences need to be globally normalized to generate the probability from the correct sequence to the predicted sequence: ;

[0130] in, represents the score of predicting the correct label sequence, and represents the sum of the scores of all possible prediction sequence results. Then, during the training process, a loss function needs to be designed to calculate the maximum probability. The present invention adopts the Viterbi Algorithm to reduce the time complexity from Optimize to .

[0131] Finally, after fine-tuning, a BERT-CRF-based annotation model was obtained. The label combination results were extracted from the corpus, and the hazard set and accident cause set were constructed based on HZ (hazard) and AC (accident cause) in the BIOES label respectively.

[0132] 3. Coupling analysis

[0133] Based on the BERT pre-trained model, the word vector fine-tuned by combining expert experience and CRF annotation is a more accurate semantic expression of the danger point. There is correlation between danger points, such as "crew" and "captain", "second mate" and "pilot", which often express similar semantics and have similar contexts. This is represented as a closer distance in the word vector space, that is, the absolute value of the modulus is smaller. The coupling between the causes of the accident and the danger points is calculated by cosine similarity. The calculation formula of cosine similarity is as follows: ;

[0134] Among them, vector A i represents the i-th vector A, vector Bi represents the i-th vector B, and n represents the total number of vectors. The calculated coupling degree is saved in the form of a matrix, and a threshold C is set. It is assumed that two keywords with a coupling degree exceeding C will establish a coupling connection, and a chord diagram is used to display the coupling connection diagram. After statistics, the number of coupling connections (L) for each word can be obtained.

[0135] 4. Importance Analysis

[0136] Word frequency statistical analysis is the most commonly used text analysis method. After extracting the accident causes and danger points from the accident corpus, the present invention first conducts word frequency statistical analysis. It is necessary to perform keyword word frequency statistics against the accident cause list and danger point list respectively, and return to the accident cause and danger point word frequency parameters (TF), as follows: ;

[0137] Expressive words In the documentation Medium word frequency, The total length of a word. TF means that the more times a word appears in a document, the greater the TF value will be. After that, the words are sorted in descending order according to the word frequency parameter and displayed in a bar chart. Combined with the statistical coupling connection number (L), as the coordinates of the word (L, F), the importance of the word is displayed in a scatter plot. The closer the word is to the upper right corner, the higher its importance is considered.

[0138] 5. Scenario Analysis

[0139] The purpose of scenario analysis is to construct an accident scenario set and find the accident causes and risk points contained in different scenarios. The present invention uses an improved LDA topic model algorithm to perform accident scenario analysis. The accident scenario set is a trained LDA model, including topics and keywords. The topic is the scenario, and the keyword is the cause of the accident.

[0140] ; ;

[0141] in, represents the accident scene set, Indicates an accident scene.

[0142] The number of topics in the LDA model determines its perplexity, and the two are generally inversely proportional. Therefore, when conducting scene analysis experiments, the number of scenes to be trained is determined by observing the changes in perplexity.

[0143] The main steps to build a scene set are as follows:

[0144] Step 1. Text preprocessing. The corpus needs to be preprocessed, and the 187 accident reports are saved as 187 continuous text segments, and irrelevant content such as punctuation, spaces, and line breaks are removed.

[0145] Step 2. Parameter setting: Customize a set of parameters for LDA topic model training, including the word frequency amplification factor, the number of training topics, and the number of topic word displays.

[0146] Step 3. Set keywords and stop words. Keywords refer to words or phrases that need special classification. Because the frequency of accident causes is not the highest in accident reports, in order to obtain a more accurate LDA topic model, these words need to be amplified. Therefore, all accident causes extracted by BERT are used as keywords for frequency amplification. At the same time, some stop words with high frequency but little meaning need to be removed, such as accident, collision, safety, etc.

[0147] Step 4. Amplify the word frequency. Amplify the word frequency of the cause of the accident. As shown in the following formula: ;

[0148] In the above formula, Indicates the updated word The word frequency, represents its original word frequency, Indicates the frequency increase coefficient of a certain word (expand number), Represents the cause set of accidents. When a word belongs to When the word frequency increases times, otherwise it will not increase. The value of will also affect the calculation results of Perplexity, and the optimal value will be selected through perplexity experiments.

[0149] Step 5. Model training: Perform LDA topic model training on the corpus to obtain a topic model.

[0150] Step 6. Calculate the perplexity. The perplexity determines the accuracy of the topic model for topic classification. The lower the perplexity, the higher the discrimination of the topic model and the better the classification effect. In this paper, the lower the perplexity, the clearer the accident scene. According to the number of different scenes (number of topics) and the magnification factor , to combine experiments and calculate perplexity. Observe and select the most appropriate combination of scene number and magnification factor, as well as the trained topic model, which is the constructed scene set. The perplexity calculation formula is as follows: ;

[0151] Where D represents the test set of the corpus, which contains M documents; represents the number of words in document d, and represents the probability of generating word w in document d.

[0152] Step 7. Analyze the scene set and classify and name the scenes.

[0153] Step 8. Scenario-Accident Cause-Hazard Point Correlation Analysis: Calculate the coupling degree between the accident cause and the hazardous point in a scenario, set a threshold to establish the coupling connection, and display it through a Sankey diagram.

[0154] Key technologies include:

[0155] (1) Select a suitable keyword extraction algorithm and a suitable corpus for training. Use the BERT pre-trained model and the CRF algorithm to pre-train the corpus. BERT is designed to pre-train unlabeled text by combining the left and right contexts to obtain a deep bidirectional representation. Therefore, only one additional output layer is needed to fine-tune the pre-trained BERT model, which is very suitable for use in this project. CRF is a conditional random field throughout, combining the characteristics of the maximum entropy model and the hidden Markov model. It has achieved good results in sequence labeling tasks such as word segmentation, part-of-speech tagging, and named entity recognition.

[0156] (2) Calculate the risk coupling degree through cosine similarity. Cosine similarity, also known as cosine similarity, evaluates the similarity of two vectors by calculating the cosine value of the angle between them. Cosine similarity plots vectors into a vector space, such as the most common two-dimensional space, based on their coordinate values. Cosine similarity measures the size of the angle between two vectors, and the result is expressed by the cosine value of the angle. Therefore, the cosine similarity of two vectors is: ;

[0157] The numerator is the dot product of vector A and vector B, and the denominator is the L2 product of the two, that is, the square root of all dimension values ​​is added together. The value of cosine similarity is [-1, 1], and the larger the value, the more similar it is.

[0158] (3) Use the LDA model for training so that it can achieve scene classification. LDA is the abbreviation of two commonly used models: Linear Discriminant Analysis and Latent Dirichlet Allocation. LDA plays a very important role in topic models and is often used for text classification. In the LDA model, a document is generated as follows:

[0159] Sample from the Dirichlet distribution to generate the topic distribution of document i. Sample from the multinomial distribution of topics to generate the topic of a word in the document. Sample from the Dirichlet distribution to generate the word distribution corresponding to the topic. Sample from the multinomial distribution of words to finally generate the word.

[0160] (4) Use Sankey diagrams to show the connections between topics, keywords, and risk points. Sankey diagrams can very well show many-to-many dependencies.

[0161] (5) Use a platform-based platform (such as Qt platform) to build the software front end. Qt is a cross-platform C++ graphical user interface library.

[0162] Technical route taken

[0163] 1. The software front-end interface is under development. The overall design adopts modular design, and the four functions of keyword extraction, risk coupling analysis, exhaustive analysis, and scenario classification and identification are inserted into separate modules to reduce coupling and adhesion. This will facilitate future function addition and maintenance.

[0164] 2. Get in touch with shipyards, shipping companies and other units to collect as much actual data as possible. Combined with open source data sets on the Internet, build a relatively complete corpus.

[0165] 3. Combine the existing corpus to train the model. According to the characteristics of ship traffic accidents, adjust the parameters of the trained network.

[0166] The present invention includes four functions: (1) keyword extraction. (2) coupling analysis. (3) importance analysis. (4) scenario analysis. Using a keyword extraction model trained based on prior knowledge, the set of accident causes and risk points extracted from the Marine Accident Basic Corpus (MABC) is the basis for the subsequent three analyses. Coupling analysis is a method used to find the correlation features between accident causes and dangerous points, and it is necessary to build a coupling connection relationship based on the distance between word vectors; importance analysis is used to evaluate the importance of accident causes or dangerous points in a certain type of accident, relying on word frequency statistics and coupling connection numbers; scenario analysis is used to identify and classify accident scenarios, and study the relationship between accident causes and risk points in a certain accident scenario.

[0167] The initial marine accident prior knowledge base (MAKB) will be used as the prior knowledge of the present invention. Its construction process mainly uses an expert group combined with literature research to pre-sort out the possible causes and danger points of ship collision accidents, and then randomly selects 20 accident investigation reports for analysis and demonstration, supplements and describes and reviews the causes and danger points of accidents, and finally forms an expert list of accident causes and risk points. The causes and danger points in some expert lists are shown in Table 1.

[0168] Table 1 List of causes and hazards of expert experience reason harm Class (c=crew, s=vessel, e=environment, m=management) Neglect of Lookout captain c Failure to verify the position of the vessel to be encountered sailor c Not using radar correctly Pilot c Fatigue duty Physical reactions c Unfamiliar with system documents attention c Insufficient sailing skills Judgment c Inexperience Security Awareness c Poor visibility rain e Strong wind and waves fog e Effects of nighttime shore lights wind e High traffic density wave e Incomplete staffing Collision avoidance rules m Poor monitoring Sailing rules m Illegal operation Check visa m Hiring undocumented personnel Duty schedule m Management absence Shift handover regulations m Host failure light s Reverse failure Sound signal s Failure to perform routine maintenance on the host as planned high frequency s Navigation equipment failure whistle s overload radar s

[0169] The corpus content-aware marine accident analysis model (CAM-MA model) proposed in this invention is applied to the ship collision accident set to mine the causes and danger points of the accident and analyze the characteristics of the collision accident. The training parameters of the keyword extraction model based on BERT-CRF are shown in Table 2. The parameters of the two types of keywords are the same. The accuracy (Accuracy) of directly using the BERT pre-trained model for extraction is 85.95%. After fine-tuning, the model accuracy is 98.19%, which has been significantly improved.

[0170] Table 2 Training parameters of keyword extraction model based on BERT-CRF Training parameters Cause Extraction Model Hazard Extraction Model Training set size 6141 6141 Batch size 32 32 Batch Quantity 192 192 Round 4 4 Loss Function Cross Entropy Cross Entropy Hardware Environment Quadro RTX 6000 (architecture = 7.5)

[0171] The causes of accidents are concentrated, and a total of 139 causes of accidents are extracted. According to the classification of man-machine-environment-management, there are 90 human factors accounting for about 45.39%, 14 ship factors accounting for about 17.34%, 15 environmental factors accounting for about 15.50%, 20 management factors accounting for about 21.77%, and 2 other factors accounting for about 0.6%. Through comparative analysis, it can be seen that since the main cause of ship collision accidents is human factors, the description of human factors in the accident investigation report is richer and more specific, which determines that the largest number of accident causes can be extracted. According to the frequency of dangerous points, the top 20 accident causes with the highest frequency are shown in Tables 3 and Figure 4 .

[0172] Table 3 List of accident causes (top 20 / 139) # name Word frequency 1 Poor visibility 244 2 Neglect of Lookout 92 3 Failure to maintain a regular lookout 80 4 Let him clear his boat 54 5 fatigue 26 6 Vessel is not seaworthy 24 7 Failure to sound the horn in accordance with regulations 20 8 High traffic density 16 9 Driving on the outer edge 16 10 subjective 12 11 Failure to exercise good seamanship 10 12 Lack of safety awareness 10 13 Not sailing at a safe speed 10 14 Sailing beyond the approved navigation area 10 15 Not equipped as required 10 16 Security Check 8 17 Not discovered in time 8 18 Carelessness 8 19 Competency Assessment 8 20 No effective avoidance measures were taken 6

[0173] The risk points are concentrated, and the accuracy (Accuracy) of directly using the BERT pre-trained model for extraction is 85.95%. After fine-tuning, the model accuracy is 98.19%, which has been significantly improved. A total of 132 danger points were extracted and classified according to people, ships, environment and management. There are 33 danger points for people, accounting for about 25%, 33 danger points for ships, accounting for about 25%, 27 danger points for the environment, accounting for about 20.45%, and 20 danger points for management, accounting for about 29.55%. According to the frequency of the danger points, the top 10 words with the highest frequency are International Regulations for Preventing Collisions at Sea, crew, regulations, captain, company, anchor, radar, ship manager, fishermen's safety awareness, and driver. Most of them are related to human factors and management factors, see Table 4 and Figure 5 .

[0174] Table 4 Danger list (top 19 / 280) # name Word frequency 1 International Regulations for Preventing Collisions at Sea 1242 2 crew 908 3 Regulation 846 4 captain 784 5 company 612 6 anchor 378 7 radar 318 8 Ship Manager 306 9 Fishermen's safety awareness 288 10 Driver 282 11 Pilot Station 266 12 Second Mate 182 13 Vision 176 14 Seamanship 168 15 Bridge watch officer 152 16 Maritime Traffic Safety Law 142 17 Duty driver 142 18 Pilot 142 19 Electronic Navigational Charts 116

[0175] In the coupling degree, the coupling degree is calculated based on the cosine similarity, and the connections of the top 20 coupling degrees are arranged in descending order, as shown in Table 5. The selection of thresholds from low to high determines the scale of the overall coupling links. For example, 0.7 means 1375 coupling links, while 0.8 means 148 coupling links, and 0.85 means 30 coupling links.

[0176] Table 5 Dangerous coupling links (top 20)

[0177] The display of coupling degree, such as Figure 6 As shown, Figure 6 The horizontal axis represents the number of topics, and the vertical axis represents the perplexity. At the same time, in order to train a more accurate scene analysis model, a perplexity calculation experiment was conducted in the accident scene concentration. The experiments were conducted according to e=(1, 10, 100, 500, 1000, 2000), and the maximum number of topics was set to 60. The results show that the value of e affects the upper limit of the perplexity. As e increases, the perplexity decreases significantly. This also reflects the role of improving LDA and increasing the amplification factor. When e exceeds 500, the decrease in the upper limit of the perplexity slows down. In addition, although e affects the upper limit of the perplexity, the downward trend of the perplexity with the number of topics is basically the same. When the number of topics = 40, the rate of decrease of the perplexity begins to decrease significantly, and even rebounds. When the number of topics and the amplification factor are too large, overfitting is prone to occur. Therefore, the two parameters of the number of topics 40 and the amplification factor 500 were finally selected to train the topic model. The perplexity calculation results are shown in Figure 7 shown.

[0178] The present invention is summarized as follows:

[0179] 1. An accident analysis model CAM-MA was constructed, and a ship traffic accident risk and scenario analysis platform software was developed, which was applied to the ship collision accident text dataset. The BERT-CRF model was trained with expert prior knowledge to extract 280 risk points and 320 accident causes. At the same time, word frequency analysis, coupling degree analysis based on cosine similarity, and scenario analysis based on the improved LDA topic model were performed on the accident causes and risk points. The final result obtained a more comprehensive set of accident causes, risk point sets, coupling connection sets, and accident scenario sets. Based on the results, the characteristics of coastal ship collision accidents were analyzed, and risk classification standard recommendations for ship collision accidents were proposed.

[0180] 2. Construct a ship navigation safety structural network model based on STAMP. Aiming at the set of ship collision accidents, analyze its complex network characteristics from different accident stage scenarios such as "collision danger", "urgent situation" and "urgent danger", and study the main risk characteristics in the ship navigation safety structure under the collision scenario.

[0181] 3. A quantitative assessment model for ship traffic accident risk based on risk entropy was constructed. Key risk nodes and edges were quantitatively assessed, and risk prevention and control strategies were proposed.

[0182] In addition, in terms of model training, we can try to use other language models based on deep learning, such as the GPT (Generative Pretrained Transformer) series of models, and combine them with specific transfer learning techniques to optimize training based on the characteristics of water traffic accident reports. This will achieve similar or better results in some aspects, but the model structure and training strategy need to be redesigned and adjusted. In terms of risk coupling analysis algorithms, in addition to cosine similarity, we can also explore the use of graph neural network-based methods to build a relationship graph of accident factors and analyze the coupling relationship through feature learning of the nodes and edges of the graph, but this requires more complex algorithm design and computing resource support, and the effect in actual applications needs further verification and optimization.

[0183] The key points of the present invention are:

[0184] (1) A method and model structure for accurately extracting the causes and risk points of water traffic accident reports using large language models such as BERT combined with expert experience and the CRF algorithm.

[0185] (2) A unique algorithm and parameter setting for coupling analysis and scene classification and identification of accident causes and risk points based on cosine similarity and improved LDA topic model.

[0186] (3) Close collaboration among modules, efficient data transmission and processing procedures, and the overall system architecture design ensure comprehensive, rapid, and accurate analysis of accident reports.

[0187] The present invention solves the following technical problems:

[0188] First, improve the analysis efficiency of water traffic accident investigation reports and realize automated and intelligent processing; second, enhance the accuracy of identifying accident causes and risk points and comprehensively extract key information; third, accurately assess the degree of risk and coupling relationship through quantitative analysis methods; fourth, provide stable and reliable analysis results to provide strong support for water traffic safety management and decision-making.

[0189] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 3 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store static information and dynamic information data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the steps in the above method embodiment are implemented.

[0190] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0191] In addition, the present invention also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiment when executing the computer program.

[0192] In addition, the present invention further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiment are implemented.

[0193] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0194] The present invention is not limited to the structures which have been described above and shown in the drawings, and various modifications and changes may be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A method for processing and analyzing a water traffic accident investigation report, characterized in that: include: The language model is pre-trained using the water traffic accident text data, so that the language model can obtain the semantic features and language patterns in the water traffic accident text data, and a keyword extraction model is obtained; the output of the keyword extraction model is annotated using the conditional random field; Based on the keyword extraction model, the causes and risk points of the water traffic accident reports to be analyzed are extracted; the similarity algorithm is used to calculate the correlation between the causes of the accidents and the risk points, and a coupling relationship matrix is ​​constructed; Using the coupling relationship matrix and coupling degree threshold, the strong coupling relationship pairs of accident causes and risk points are screened out; According to the word frequency statistical analysis technology, the word frequency analysis of the accident causes and risk points is carried out to obtain the importance of the accident causes and risk points; Using the topic model and combining the causes of accidents, the water traffic accident reports to be analyzed are classified into scenarios and an accident scenario set is constructed.

2. A method for processing and analyzing water traffic accident investigation reports according to claim 1, characterized in that: Before the language model is pre-trained using the water traffic accident text data, the following steps are also included: Collect water traffic accident investigation reports and convert them into text data in a unified format; Clean the text data in the water traffic accident investigation report and remove noise data, irrelevant symbols and repeated information; The text data in the water traffic accident investigation report is subjected to lexical and syntactic analysis, and the professional terms are standardized to obtain the water traffic accident text data.

3. A method for processing and analyzing water traffic accident investigation reports according to claim 1, characterized in that: The method of labeling the output of the keyword extraction model using the conditional random field includes: Construct the characteristic function and transfer function of the conditional random field; use the conditional random field and combine it with expert experience to annotate the output of the keyword extraction model.

4. A method for processing and analyzing water traffic accident investigation reports according to claim 1, characterized in that: The similarity algorithm is used to calculate the correlation between the causes of accidents and the risk points, and to construct a coupling relationship matrix, including: The cosine similarity calculation formula is used to calculate the coupling degree between the causes of accidents and between the risk points, and the coupling degree is saved in the form of a matrix.

5. The method for processing and analyzing water traffic accident investigation reports according to claim 1, characterized in that: The coupling relationship matrix and the coupling degree threshold are used to screen out the accident cause strong coupling relationship pairs and the risk point strong coupling relationship pairs, including: A coupling degree threshold is set in advance, and a coupling connection is established between two keywords that exceed the coupling degree threshold. The coupling connection status is displayed using a chord diagram, and the number of coupling connections for each keyword is statistically obtained. The keywords include accident causes and risk points. According to the coupling connections established between the keywords, the strongly coupled relationship pairs of accident causes and the strongly coupled relationship pairs of risk points are obtained.

6. A method for processing and analyzing water traffic accident investigation reports according to claim 1, characterized in that: According to the word frequency statistical analysis technology, the word frequency analysis is performed on the accident causes and risk points to obtain the importance of the accident causes and risk points, including: Through the accident cause list and risk point list, the word frequency statistics of accident cause and risk point are carried out to obtain the word frequency parameters of each accident cause and risk point; Arrange in descending order according to the word frequency parameter and display it in a bar graph; determine the coordinates of the corresponding keyword in the scatter plot based on the statistical number of coupling connections for each keyword, and determine the importance of the corresponding accident causes and risk points based on the coordinates of each keyword in the scatter plot.

7. A method for processing and analyzing water traffic accident investigation reports according to claim 1, characterized in that: The topic model is used to classify the water traffic accident reports to be analyzed in combination with the causes of the accident, and the accident scene set is constructed, including: Determine the training parameters of the topic model, including the frequency amplification coefficient, the number of training topics, and the number of topic word displays; amplify the frequency of the cause of the accident and remove stop words; Through the perplexity experiment, the required number of scenes and the magnification factor are selected and observed; the topic model is trained using the corpus to obtain a trained topic model; Combining the number of scenes, the magnification factor and the trained topic model, the accident scene set is obtained.

8. A method for processing and analyzing water traffic accident investigation reports according to claim 7, characterized in that: When the frequency of the word causing the accident is amplified, the frequency amplification formula is: ； In the formula, Indicates the updated word k The word frequency, represents its original word frequency; e represents the frequency increase coefficient of a certain word, C represents the set of accident causes, k Words that represent incident reports.

9. A method for processing and analyzing water traffic accident investigation reports according to claim 7, characterized in that: The construction of the accident scenario set includes: Classify and name each scene in the accident scene set; calculate the coupling degree between the accident cause and the risk point in a scene through the correlation analysis of scenes, accident causes and risk points, establish the coupling connection through the preset threshold, and display it through the Sankey diagram.

10. A water traffic accident investigation report processing and analysis system, characterized in that: include: The keyword extraction model training module is used to pre-train the language model using the water traffic accident text data, so that the language model can obtain the semantic features and language patterns in the water traffic accident text data, and obtain the keyword extraction model; the output of the keyword extraction model is annotated using the conditional random field; The coupling analysis module is used to extract the causes and risk points of the water traffic accident reports to be analyzed based on the keyword extraction model; the similarity algorithm is used to calculate the correlation between the causes of the accidents and the risk points, and to construct a coupling relationship matrix; Using the coupling relationship matrix and coupling degree threshold, the strong coupling relationship pairs of accident causes and risk points are screened out; Importance analysis module, used to perform word frequency analysis on accident causes and risk points based on word frequency statistical analysis technology to obtain the importance of accident causes and risk points; The scene classification module is used to utilize the topic model and combine the causes of the accident to classify the water traffic accident reports to be analyzed and construct an accident scene set.

Images