Aviation safety report general information intelligent extraction method based on accident cause table

Abstract

The invention belongs to the technical field of aviation safety, natural language processing and artificial intelligence, and particularly relates to an intelligent extraction method for general information of an aviation safety report based on an accident cause table. The method aims at solving the problem that an existing aviation safety information risk identification technology cannot efficiently and accurately identify the risk type and the risk reason of an aviation report. The method comprises the steps that an aviation accident report is acquired, translated and corrected, and a Chinese aviation data set is constructed; preprocessing data in the data set; converting the preprocessed text into numerical vectors and clustering the numerical vectors to generate accident cause clusters; performing subject naming and hierarchical division on the accident cause cluster to obtain various types of accident cause tables; performing fine tuning training on the BERT classification model; and acquiring a newly input aviation safety information text and a cause table corresponding to the newly input aviation safety information text, and obtaining an optimal accident matching cause by using a Ball-tree nearest neighbor algorithm. According to the invention, efficient and accurate identification of the risk type and the risk reason of the aviation report is realized.

Description

Technical Field

[0001] The present invention belongs to the field of aviation safety, natural language processing and artificial intelligence technology, and specifically relates to a method, system, electronic equipment and storage medium for intelligently extracting general information from aviation safety reports based on an accident cause table. Background Art

[0002] my country's civil aviation passenger volume continues to rise. In aviation safety management, most accident information is recorded in the form of accident reports. There is a lot of important information hidden in the massive accident report data waiting for us to mine, such as aviation accident cause analysis and accident chain, which plays a vital role in aviation safety management.

[0003] Traditional aviation accident cause analysis often adopts a coarse-grained unified classification method, using the same cause classification standard for all accident types. The same is true for the cause classification method of the National Transportation Safety Board (NTSB) of the United States. However, in reality, each type of accident has its own particularity, and a rough and unified classification method cannot effectively identify each risk accurately and take corresponding training and preventive measures. Therefore, a method is needed to efficiently collect and summarize the accident causes of various accident types in massive accident data, so as to improve the accuracy of aviation safety prevention.

[0004] In addition, during daily aviation operations, airlines also generate a large number of aviation safety information reports every day, which require human analysis and interpretation, identification and classification of general information, and risk screening and prevention. The traditional aviation safety report classification method mainly relies on manual reading and manual coding. In the actual operation process, there are problems such as large workload and high time cost. In addition, the accuracy of the classification results depends largely on the experience and knowledge of the analysts and is easily affected by subjective judgment.

[0005] Existing aviation safety information risk identification technologies can be divided into two categories. One method is to obtain keywords or sensitive words through text features for matching. However, simple matching often fails to capture the deep semantic information in the text, resulting in insufficient accuracy and comprehensiveness of information extraction. The other method is to build a model based on existing corpus data, and obtain text classification results after fine-tuning the model. This method has good classification effects, but is often limited by the basic classification structure of the existing corpus, and the data sources are different, and the definition and basis of classification are unclear, resulting in some important but uncommon information being ignored and concealed, and the obtained classified information cannot be directly applied in actual aviation operations. Therefore, there is an urgent need for an automated information extraction method that can efficiently process large-scale text data and has deep semantic understanding capabilities to replace traditional manual analysis and improve the efficiency and accuracy of aviation safety report analysis. Summary of the invention

[0006] In order to solve the above problems in the prior art, that is, to solve the problem that the existing aviation safety information risk identification technology cannot efficiently and accurately identify the risk type and risk cause of aviation reports, the first aspect of the present invention proposes an intelligent extraction method for general information of aviation safety reports based on an accident cause table, which is used to automatically obtain the best matching cause of aviation accidents. The method includes:

[0007] Step S1, obtaining aviation accident reports, translating and correcting the aviation accident reports, and constructing a Chinese aviation dataset;

[0008] Step S2, preprocessing the data in the Chinese aviation data set to obtain a preprocessed text;

[0009] Step S3, converting the preprocessed text into a numerical vector through the SBERT model, and clustering it using the Single-pass incremental clustering algorithm to generate an accident cause cluster;

[0010] Step S4, subject-naming and hierarchical division of the accident cause clusters to obtain a table of causes of various types of accidents;

[0011] Step S5, fine-tuning the BERT classification model;

[0012] Step S6, classify the newly input aviation safety information text through the fine-tuned BERT classification model to obtain the accident type, and combine the cause tables of each type of accident to obtain the cause table corresponding to the newly input aviation safety information text; based on the newly input aviation safety information text and its corresponding cause table, use the Ball-tree nearest neighbor algorithm to obtain the best matching cause of the accident.

[0013] In some preferred implementations, the aviation accident reports are translated and corrected to construct a Chinese aviation dataset, and the method is as follows:

[0014] converting the aviation accident report from a Microsoft Access format to an SQL format;

[0015] Selecting basic data corresponding to each accident type from the aviation accident report after format conversion, saving it as a data set and converting it into CSV format; the basic data includes a narrative text of the cause of the accident;

[0016] The dataset was translated and corrected from English to Chinese to obtain a Chinese aviation dataset.

[0017] In some preferred embodiments, the preprocessing includes data cleaning, sentence segmentation, and text cleaning;

[0018] Among them, sentence division is as follows: the continuous long text is divided into shorter text units by the original existing commas and periods to form a sentence sequence;

[0019] Text cleaning includes removing HTML tags, deleting punctuation marks, special characters and numbers in the text, and removing stop words.

[0020] In some preferred embodiments, the preprocessed text is converted into a numerical vector by an SBERT model, and clustered using a Single-pass incremental clustering algorithm to generate an accident cause cluster, the method being:

[0021] Encoding each text unit of the accident cause narrative text in the aviation accident report through the SBERT model, thereby obtaining a vector representation of each text unit;

[0022] Select the vector corresponding to the first text in the Chinese aviation dataset as the center of the first cluster;

[0023] For each newly input text vector, calculate its similarity with the center of each cluster. If the similarity exceeds the set similarity threshold, add the text vector to the cluster and update the cluster center vector. If all similarities are lower than the set similarity threshold, create a new cluster with the newly input text vector as the center of the cluster.

[0024] The Chinese aviation data set is traversed until all text vectors have completed clustering processing, thereby obtaining accident cause clusters.

[0025] In some preferred implementations, the accident cause clusters are subject-named and hierarchically divided to obtain a table of causes of various types of accidents, and the method is as follows:

[0026] Based on the text content of each accident cause cluster, determine the theme of the cluster;

[0027] The themes of the clusters are hierarchically divided according to the set dimensions to form a table of causes of various types of accidents.

[0028] In some preferred implementations, the BERT classification model is fine-tuned and trained by:

[0029] Construct training data based on the obtained aviation safety information text and its corresponding event type labels;

[0030] Inputting the aviation safety information text into the BERT classification model to obtain a classification result;

[0031] Calculate a loss function based on the classification result and the event type label, and then adjust the parameters of the BERT classification model;

[0032] The BERT classification model is looped until a trained BERT classification model is obtained.

[0033] In some preferred embodiments, based on the newly input aviation safety information text and its corresponding cause table, the best matching cause of the accident is obtained using the Ball-tree nearest neighbor algorithm, and the method is as follows:

[0034] Splitting the newly input aviation safety information text and its corresponding cause table;

[0035] After the text is split, the cause table corresponding to the newly input aviation safety information text is vectorized item by item and sentence by sentence through the SBERT model to obtain a cause table word vector matrix, and the newly input aviation safety information text is vectorized sentence by sentence to obtain an input text vector matrix;

[0036] Standardizing the text vectors in the cause table word vector matrix and the input text vector matrix;

[0037] A ball-tree is constructed based on the standardized word vector matrix of the cause table. The standardized input text vector matrix is ​​used as input. The ball-tree formed by each input text vector and a cause table word vector performs a nearest neighbor search. The n causes with the closest Euclidean distance are selected from the output of the nearest neighbor search as the best matching causes of the accident.

[0038] The second aspect of the present invention proposes an intelligent extraction system for general information of aviation safety reports based on an accident cause table, which is used to automatically obtain the best matching cause of aviation accidents. The system includes:

[0039] A data acquisition module, configured to acquire aviation accident reports, translate and correct the aviation accident reports, and construct a Chinese aviation data set;

[0040] A preprocessing module, configured to preprocess the data in the Chinese aviation data set to obtain a preprocessed text;

[0041] A clustering module, configured to convert the preprocessed text into a numerical vector through an SBERT model, and perform clustering using a single-pass incremental clustering algorithm to generate an accident cause cluster;

[0042] A hierarchical division module is configured to perform subject naming and hierarchical division on the accident cause clusters to obtain a table of causes of various types of accidents;

[0043] A fine-tuning training module, configured to perform fine-tuning training on the BERT classification model;

[0044] The cause acquisition module is configured to classify the newly input aviation safety information text through the fine-tuned BERT classification model to obtain the accident type, and combine the cause tables of each type of accident to obtain the cause table corresponding to the newly input aviation safety information text; based on the newly input aviation safety information text and its corresponding cause table, use the Ball-tree nearest neighbor algorithm to obtain the best matching cause of the accident.

[0045] According to a third aspect of the present invention, an electronic device is provided, comprising:

[0046] At least one processor; and a memory in communication with at least one of the processors; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to implement the above-mentioned method for intelligently extracting general information from aviation safety reports based on an accident cause table.

[0047] In a fourth aspect of the present invention, a computer-readable storage medium is proposed, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned method for intelligently extracting general information from aviation safety reports based on the accident cause table.

[0048] Beneficial effects of the present invention:

[0049] The present invention realizes efficient and accurate identification of risk types and risk causes of aviation reports.

[0050] 1) Starting from the accident report text, the present invention can efficiently collect and summarize the accident causes of various accident types in massive accident data, thereby accurately improving the accuracy of aviation safety prevention of various risks. The construction of the accident cause table clusters the causes of each accident type, and defines and explains the causes, breaking through the limitations of the existing data corpus structure and classification and the ambiguity of the definition. Based on the accident cause table, the similarity of the new text is calculated to obtain the Top-n causes with the highest similarity, thereby realizing the intelligent extraction of general information of aviation safety reports, helping to identify the main risk factors and trends in the actual aviation operation process, providing data support for aviation safety management and decision-making, and realizing the rapid management of future aviation safety reports while ensuring management accuracy;

[0051] 2) The present invention replaces the traditional method with SBERT's text vectorization technology, and proposes a method of dividing the text sentence by sentence and performing semantic similarity clustering, thereby realizing a more fine-grained single-pass incremental clustering vector set, thereby improving the accuracy of the clustering process at the semantic level;

[0052] 3) Different from the previous intelligent risk identification methods for aviation safety information reports, this invention adopts the accident cause table generated by single-pass clustering as the basis for intelligent identification tasks. Different from the traditional unified classification of accident causes, this invention constructs a cause table for each accident type separately, and has a more detailed definition and explanation of each type of risk and cause, which greatly improves the accuracy of risk type and corresponding hazard source identification;

[0053] 4) Aiming at the actual scenario of combining cause table analysis, the present invention adopts commas and periods to segment the text in the similarity matching link, and forms a word vector matrix by vectorizing shorter text units. This operation separates invalid information from valid information as much as possible, improves the quality of the formed word vector, and thus greatly improves the matching accuracy;

[0054] 5) In the stage of matching text and cause table, the present invention uses the Ball-tree nearest neighbor algorithm to accelerate the matching process of high-dimensional vectors, which is different from the previous matrix traversal matching process. And the strong correlation between Euclidean distance and cosine similarity on the standardized vector is used to ensure that the cause matching result still follows the optimal cosine similarity;

[0055] 6) The present invention divides the cause table into multiple levels, and forms the cause table based on different risk types, the attribution of risks in human factors, environment, machinery, organization, etc., and the specific description of risks. The hierarchical cause table significantly improves the matching accuracy of events, risks and corresponding hazards. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting embodiments made with reference to the following drawings.

[0057] Figure 1 It is a flow chart of a method for intelligently extracting general information from aviation safety reports based on an accident cause table according to an embodiment of the present invention;

[0058] Figure 2 is a flow chart of NTSB data acquisition according to an embodiment of the present invention;

[0059] Figure 3 This is an example diagram of querying NTSB data based on Navicat Premium 15.0 tool according to an embodiment of the present invention;

[0060] Figure 4 This is a comparison result diagram of different text vectorization methods based on the silhouette coefficient evaluation index according to an embodiment of the present invention;

[0061] Figure 5This is a comparison result diagram of different clustering models based on intra-cluster and inter-cluster average similarity evaluation indicators according to an embodiment of the present invention;

[0062] Figure 6 This is an example diagram of an accident cause representation based on natural language processing according to an embodiment of the present invention;

[0063] Figure 7 It is an example diagram of obtaining accident text data attribution (i.e., the best matching cause of the accident) based on the method of the present invention in one embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0066] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other.

[0067] The first embodiment of the present invention is a method for intelligently extracting general information from aviation safety reports based on an accident cause table, which is used to automatically obtain the best matching cause of an aviation accident, such as Figure 1 As shown, the following steps are included:

[0068] Step S1, obtaining aviation accident reports, translating and correcting the aviation accident reports, and constructing a Chinese aviation dataset;

[0069] Step S2, preprocessing the data in the Chinese aviation data set to obtain a preprocessed text;

[0070] Step S3, converting the preprocessed text into a numerical vector through the SBERT model, and clustering it using the Single-pass incremental clustering algorithm to generate an accident cause cluster;

[0071] Step S4, subject-naming and hierarchical division of the accident cause clusters to obtain a table of causes of various types of accidents;

[0072] Step S5, fine-tuning the BERT classification model;

[0073] Step S6, classify the newly input aviation safety information text through the fine-tuned BERT classification model to obtain the accident type, and combine the cause tables of each type of accident to obtain the cause table corresponding to the newly input aviation safety information text; based on the newly input aviation safety information text and its corresponding cause table, use the Ball-tree nearest neighbor algorithm to obtain the best matching cause of the accident.

[0074] In order to more clearly illustrate the method for intelligently extracting general information from aviation safety reports based on an accident cause table of the present invention, each step in an embodiment of the method of the present invention is described in detail below with reference to the accompanying drawings.

[0075] The present invention provides a method for intelligently extracting general information from aviation safety reports based on an accident cause table, aiming to improve the analysis efficiency and accuracy of aviation safety reports through natural language processing technology, and comprising the following steps: the invention collects aviation reports (i.e., aviation accident reports), translates and proofreads them by experts, constructs a Chinese aviation data set, and performs data cleaning and text preprocessing steps; the invention adopts the text vectorization technology of the SBERT (Sentence-Bidirectional Encoder Representation from Transformer) model combined with the Single-pass clustering method to obtain the causal factors; and the definition of the causal factors is explained, and the hierarchical division of the causal factors is performed to form a structured accident cause table; further, the invention adopts BERT (Bidirectional Encoder Representation from Transformer) model to extract the causal factors; the invention further adopts the ... The AI ​​​​Transformer) model is used to classify new texts and determine the type of accident, thus completing the intelligent identification of major risks in aviation safety information reports generated during actual navigation; further, the text similarity is calculated based on the locked cause table of the accident type, and the text processing is divided sentence by sentence to obtain the word vector matrix of the input and the cause table, and the Ball-tree nearest neighbor algorithm is used to extract the Top-n causes with the highest cosine similarity, thus achieving intelligent identification of the risks of aviation safety information reports and the corresponding risk causes. The present invention can not only process large-scale text data, but also accurately capture the deep semantic information in text classification, realize the intelligent extraction of general information in aviation safety reports, and provide data support for aviation safety improvements. The details are as follows:

[0076] Step S1, obtaining aviation accident reports, translating and correcting the aviation accident reports, and constructing a Chinese aviation dataset;

[0077] In this embodiment, the aviation accident report is preferably collected from NTSB aviation reports, and the original English data is machine translated and proofread by experts to construct a practical Chinese aviation data set. Specifically:

[0078] Step S101, obtain the original English data from the NTSB official website, and use Navicat Premium 15.0 graphical management software to convert the NTSB data from Microsoft Access format to SQL format to facilitate data query and analysis. Figure 2 shown.

[0079] Step S102: In the NTSB database, according to the code number corresponding to the accident type (also called event type), select the aviation accident report related to "Runway excursion" and not empty as the basic data. The data set of the basic data includes ev_id (event number), narr_cause (event cause description (also called accident cause description)), such as Figure 3 As shown in the figure, narr_cause is the corpus of causal analysis and extraction research, which is finally saved in csv format.

[0080] Step S103: The original English data is translated into Chinese text by machine translation, and the Chinese text is corrected by experts to obtain accurate and usable Chinese text for subsequent analysis.

[0081] Step S2, preprocessing the data in the Chinese aviation data set to obtain a preprocessed text;

[0082] In this embodiment, the preprocessing includes data cleaning, sentence segmentation, and text cleaning, that is, data cleaning, sentence segmentation, and text cleaning are performed on the corrected Chinese text.

[0083] Sentence segmentation: Split long continuous text into shorter text units using the original commas and periods to form a sequence of sentences, which helps to process text in a more fine-grained manner;

[0084] Text cleaning: remove HTML tags (such as , etc.), delete punctuation marks, special characters (such as %, *, <, etc.) and numbers in the text. Removing stop words aims to remove high-frequency words that contribute little to the semantic analysis of the text (such as "of, is, at, we, you...", prepositions, personal pronouns, auxiliary words, conjunctions, etc.), so as to increase the proportion of target information in the overall text.

[0085] Step S3, convert the preprocessed text into a numerical vector through the SBERT model, and perform clustering using the Single-pass incremental clustering algorithm to generate accident cause clusters;

[0086] In this embodiment, the specific process of generating accident cause clusters is as follows:

[0087] Step S301, encode each text unit of the cause narrative text in the aviation accident report through the SBERT (Sentence-BERT) model, so as to obtain the vector representation of each text unit.

[0088] Step S302, select the vector corresponding to the first text (i.e., the first aviation accident cause statement) in the Chinese aviation dataset as the center of the first cluster;

[0089] Step S303, for each newly input text vector, calculate its similarity score with the centers of existing clusters (in the present invention, preferably, the similarity score uses the cosine similarity calculation method to calculate the text similarity between different sentence vectors).

[0090] The calculation formula based on cosine similarity (Cosine Similarity) is the dot product of two vectors divided by the product of their magnitudes, and the process is as follows:

[0091] d = (w d0 , w d1 , …, w dk ) (1)

[0092] q = (w q0 , w q1 , …, w qk ) (2)

[0093]

[0094] wherein, Equation (1) and Equation (2) obtain the corresponding sentence vectors through text feature representation, w di represents the i-th component of the sentence d vector; w qi represents the i-th component of the sentence q vector; k represents the dimension of the vector. Based on the cosine similarity calculation formula, the cosine value between two sentence vectors can be defined as calculation formula (3).

[0095] Step S304, set 0.7 as the similarity threshold. The cluster i with the maximum similarity to the input text vector is obtained by calculation. If the maximum similarity exceeds the threshold, the new text is added to the current cluster i and the cluster center vector is updated; if all similarities are lower than the threshold, a new cluster is created with the new text vector as the center of the cluster.

[0096] Step S305, traverse all texts and repeat the above steps until all aviation accident cause statement data are processed. The final clustering result contains a number of similar cause texts in each cluster, namely, the accident cause cluster.

[0097] After step S305, the method further includes step S306, in which the model result is evaluated.

[0098] This paper establishes evaluation indicators for measuring clustering effects from different perspectives. By comparing different text vectorization methods (TF-IDF, Word2Vec) and different clustering algorithms (K-means clustering), and by using multiple evaluation indicators in combination, the clustering model can be evaluated and optimized more accurately to ensure its effectiveness in practical applications.

[0099] First, we use TF-IDF to establish an aviation accident cause extraction model, which is a traditional text vectorization model. We use TF-IDF to calculate the frequency (term frequency, TF) of a given word in the document, as shown in formula (4):

[0100]

[0101] Where count(t) is the number of times word t appears in document D. i is the total number of words in document D.

[0102] Calculate the number of documents in the corpus that contain word t, and calculate the inverse document word frequency, as shown in equations (5) and (6):

[0103]

[0104]

[0105] Where N is the total number of documents in the corpus, ∑I(t, D i ) is the total number of documents containing word t.

[0106] Filter out common words and retain key words. TF-IDF calculation is shown in formula (7):

[0107]

[0108] An aviation accident cause extraction model based on Word2Vec is constructed to map words into a low-dimensional vector space so that similar words are close to each other in the vector space.

[0109] The Silhouette Coefficient is used to evaluate different text vectorization methods. The Silhouette Coefficient combines the closeness and separation to measure the distance between each data point and its cluster and the nearest other cluster. The value of the Silhouette Coefficient is between -1 and 1. The larger the value, the better the clustering effect. The specific calculation formula is:

[0110]

[0111] Among them, a(i) is the average distance from data point i to other points in the same cluster, and b(i) is the average distance from data point i to the nearest other cluster.

[0112] The silhouette coefficient evaluation results using different text vectorization methods are shown in Figure 2. Figure 4 As shown in the figure, compared with other text vectorization methods such as TF-IDF and Word2Vec, the use of the pre-trained SBERT model for text vectorization has a better improvement, which proves that the SBERT model is applied to incremental text clustering and improves the accuracy of similarity calculation and clustering effect.

[0113] The Average Intra-Cluster Similarity evaluation index is established to measure the average distance between data points and the cluster center in the same cluster. A smaller average intra-cluster similarity indicates a better compactness of clusters. A larger average inter-cluster similarity indicates a better separation between clusters.

[0114] The average similarity evaluation results of clusters within and between clusters using different clustering models are as follows: Figure 5 As shown in the figure, compared with the K-means clustering model, the Single-Pass clustering model has a higher average similarity within the cluster and a lower average similarity between clusters, which improves the quality and efficiency of text processing, is suitable for large-scale text data, and can quickly process and analyze a large number of aviation safety reports.

[0115] Step S4, subject-naming and hierarchical division of the accident cause clusters to obtain a table of causes of various types of accidents;

[0116] In this embodiment, based on the text content of each accident cause cluster, the theme of the cluster is determined; the theme of the cluster is hierarchically divided according to the set dimension to form a table of various types of accident causes. Specifically:

[0117] After clustering is completed, the text content of each cluster is combined with expert knowledge to determine the theme of the cluster. It is preferred to divide the cluster themes into levels according to the four dimensions of "man-machine-environment-management" to form a clear hierarchical structure (that is, according to the risk type, the risk attribution in human factors, environment or machinery, and the definition of each cause category is explained). The sorted accident causes are summarized and organized to form a structured accident cause table, such as Figure 6 Taking the runway deviation accident type as an example, this figure shows the clustering results of some specific accident causes at different levels of this accident type, which is convenient for subsequent analysis and application.

[0118] Step S5, fine-tuning the BERT classification model;

[0119] In this embodiment, the BERT classification model is trained and fine-tuned according to the event type label. Accident text data with event type labels is used as a training data set, and the BERT model is fine-tuned through this own data set so that the model can classify new input text. Specifically:

[0120] Based on the obtained aviation safety information text and its corresponding event type label, training data is constructed; the aviation safety information text is input into the BERT classification model to obtain a classification result; according to the classification result and the event type label, a loss function is calculated, and then the parameters of the BERT classification model are adjusted; the BERT classification model is looped until a trained BERT classification model is obtained.

[0121] Step S6, classify the newly input aviation safety information text through the fine-tuned BERT classification model to obtain the accident type, and combine the cause tables of each type of accident to obtain the cause table corresponding to the newly input aviation safety information text; based on the newly input aviation safety information text and its corresponding cause table, use the Ball-tree nearest neighbor algorithm to obtain the best matching cause of the accident.

[0122] In this embodiment, the specific process of obtaining the best matching cause of the accident is as follows:

[0123] S601, pre-processing the newly input aviation safety information text.

[0124] S602: Classify the pre-processed aviation safety information text using the fine-tuned BERT classification model to obtain the accident type.

[0125] S603, retrieve the corresponding cause table according to the confirmed accident type. At this time, for the input event text (ie, aviation safety information text) and the cause table text to be matched, each text is split into a list of short texts according to periods and commas.

[0126] S604, vectorizing the cause table corresponding to the newly input aviation safety information text line by line and sentence by sentence through the SBERT model to obtain a cause table word vector matrix, and vectorizing the newly input aviation safety information text sentence by sentence to obtain an input text vector matrix;

[0127] S605, normalizing the text vectors in the cause table word vector matrix and the input text vector matrix;

[0128] S606, a ball-tree is constructed based on the standardized cause table word vector matrix, and the standardized input text vector matrix is ​​used as input. The ball-tree formed by each input text vector and a cause table word vector performs a nearest neighbor search to obtain a cause with the closest Euclidean distance. After completing the nearest neighbor search for all cause texts, the n causes with the closest Euclidean distance are selected from the output of the nearest neighbor search as the best matching cause of the accident.

[0129] For the normalized vectors d, q, the Euclidean distance E dis (d, q) and cosine similarity COS(d, q) are related:

[0130]

[0131] Convert Euclidean distance to cosine similarity, and the final output is as follows Figure 7 shown.

[0132] In summary, the present invention: 1) uses the SBERT model to capture deep semantic information at the sentence level, represents the text features of the accident cause text in the aviation safety report, generates high-quality sentence vectors, and combines the incremental clustering algorithm and the Ball-tree nearest neighbor search algorithm to improve the accuracy of clustering analysis and similarity calculation. Similar causes in aviation safety reports can be classified into the same category, thereby forming a cause table for each type of accident, which has a clear definition and basis for the cause classification, helps to identify the main risk factors and trends of aviation accidents, and provides data support for aviation safety improvements. When processing and analyzing aviation safety reports, it can provide more efficient and accurate cause extraction and identification, and provide strong support for aviation safety management and decision-making; 2) The BERT model is fine-tuned to achieve classification of a variety of common event types. And by matching the cause table of each accident type separately, the accuracy of matching SBERT model events with causes is improved. That is, the present invention uses a pre-trained model to process complex similarity text analysis, which makes calculations more convenient and the results more accurate and stable. 3) This method establishes a hierarchical cause table for different accident types, so that the explanation of cause division and hazard source definition is more detailed and accurate. The general information of the aviation safety report is intelligently extracted based on the accident cause table, which reduces the time and energy cost of manual reading of the text, avoids problems such as inconsistent and inaccurate manual reading comprehension, and significantly improves the accuracy of intelligent identification of risk types and risk causes in the report, providing a convenient and controllable direction for aviation safety text research and flight safety management.

[0133] A second embodiment of the present invention is an intelligent extraction system for general information of aviation safety reports based on an accident cause table, which is used to automatically obtain the best matching cause of aviation accidents. The system includes:

[0134] A data acquisition module, configured to acquire aviation accident reports, translate and correct the aviation accident reports, and construct a Chinese aviation data set;

[0135] A preprocessing module, configured to preprocess the data in the Chinese aviation data set to obtain a preprocessed text;

[0136] A clustering module, configured to convert the preprocessed text into a numerical vector through an SBERT model, and perform clustering using a single-pass incremental clustering algorithm to generate an accident cause cluster;

[0137] A hierarchical division module is configured to perform subject naming and hierarchical division on the accident cause clusters to obtain a table of causes of various types of accidents;

[0138] A fine-tuning training module, configured to perform fine-tuning training on the BERT classification model;

[0139] The cause acquisition module is configured to classify the newly input aviation safety information text through the fine-tuned BERT classification model to obtain the accident type, and combine the cause tables of each type of accident to obtain the cause table corresponding to the newly input aviation safety information text; based on the newly input aviation safety information text and its corresponding cause table, use the Ball-tree nearest neighbor algorithm to obtain the best matching cause of the accident.

[0140] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0141] It should be noted that the above embodiment provides an intelligent extraction system for general information of aviation safety reports based on the accident cause table, and only uses the division of the above functional modules as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the modules or steps in the embodiments of the present invention can be decomposed or combined. For example, the modules in the above embodiments can be combined into one module, or further divided into multiple sub-modules to complete all or part of the functions described above. The names of the modules and steps involved in the embodiments of the present invention are only for distinguishing the modules or steps, and are not regarded as improper limitations of the present invention.

[0142] An electronic device according to the third embodiment of the present invention comprises: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the above-mentioned method for intelligently extracting general information from aviation safety reports based on an accident cause table.

[0143] A computer-readable storage medium according to a fourth embodiment of the present invention stores computer instructions, and the computer instructions are used to be executed by the computer to implement the above-mentioned method for intelligently extracting general information from aviation safety reports based on an accident cause table.

[0144] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and related instructions of the electronic device and computer-readable storage medium described above can refer to the corresponding process in the aforementioned method example and will not be repeated here.

[0145] Those skilled in the art should be able to appreciate that the modules and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software or a combination of the two, and the programs corresponding to the software modules and method steps can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the technical field. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0146] The terms "first", "second", etc. are used to distinguish similar objects rather than to describe or indicate a particular order or sequence.

[0147] The term "comprise" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that includes a list of elements includes not only those elements but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or apparatus / device.

[0148] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

Claims

1. An intelligent extraction method for general information of aviation safety reports based on accident cause table, used to automatically obtain the best matching cause of aviation accidents, characterized in that: The method comprises the following steps: Step S1, obtaining aviation accident reports, translating and correcting the aviation accident reports, and constructing a Chinese aviation dataset; Step S2, preprocessing the data in the Chinese aviation data set to obtain a preprocessed text; Step S3, converting the preprocessed text into a numerical vector through the SBERT model, and clustering it using the Single-pass incremental clustering algorithm to generate an accident cause cluster; Step S4, subject-naming and hierarchical division of the accident cause clusters to obtain a table of causes of various types of accidents; Step S5, fine-tuning the BERT classification model; Step S6, classify the newly input aviation safety information text through the fine-tuned BERT classification model to obtain the accident type, and combine the cause tables of each type of accident to obtain the cause table corresponding to the newly input aviation safety information text; based on the newly input aviation safety information text and its corresponding cause table, use the Ball-tree nearest neighbor algorithm to obtain the best matching cause of the accident.

2. The intelligent extraction method of general information of aviation safety report based on accident cause table according to claim 1 is characterized in that: The aviation accident reports are translated and corrected to construct a Chinese aviation dataset. The method is as follows: converting the aviation accident report from a Microsoft Access format to an SQL format; Selecting basic data corresponding to each accident type from the aviation accident report after format conversion, saving it as a data set and converting it into CSV format; the basic data includes a narrative text of the cause of the accident; The dataset was translated and corrected from English to Chinese to obtain a Chinese aviation dataset.

3. The intelligent extraction method of general information of aviation safety report based on accident cause table according to claim 2 is characterized in that: The preprocessing includes data cleaning, sentence segmentation, and text cleaning; Among them, sentence division is as follows: the continuous long text is divided into shorter text units by the original existing commas and periods to form a sentence sequence; Text cleaning includes removing HTML tags, deleting punctuation marks, special characters and numbers in the text, and removing stop words.

4. The intelligent extraction method of general information of aviation safety report based on accident cause table according to claim 3 is characterized in that: The preprocessed text is converted into a numerical vector through the SBERT model, and clustered using the Single-pass incremental clustering algorithm to generate accident cause clusters, the method is as follows: Encoding each text unit of the narrative text of the cause of the accident in the aviation accident report through the SBERT model, thereby obtaining a vector representation of each text unit; Select the vector corresponding to the first text in the Chinese aviation dataset as the center of the first cluster; For each newly input text vector, calculate its similarity with the center of each cluster. If the similarity exceeds the set similarity threshold, add the text vector to the cluster and update the cluster center vector. If all similarities are lower than the set similarity threshold, create a new cluster with the newly input text vector as the center of the cluster. The Chinese aviation data set is traversed until all text vectors have completed clustering processing, thereby obtaining accident cause clusters.

5. The intelligent extraction method of aviation safety report general information based on accident cause table according to claim 4 is characterized in that: The accident cause clusters are named and divided into levels to obtain a table of causes of various types of accidents. The method is as follows: Based on the text content of each accident cause cluster, determine the theme of the cluster; The themes of the clusters are hierarchically divided according to the set dimensions to form a table of causes of various types of accidents.

6. The intelligent extraction method of general information of aviation safety report based on accident cause table according to claim 5 is characterized in that: Fine-tune the BERT classification model as follows: Construct training data based on the obtained aviation safety information text and its corresponding event type labels; Inputting the aviation safety information text into the BERT classification model to obtain a classification result; Calculate the loss function according to the classification result and the event type label, and then adjust the parameters of the BERT classification model; The BERT classification model is looped until a trained BERT classification model is obtained.

7. The intelligent extraction method of general information of aviation safety report based on accident cause table according to claim 6 is characterized in that: Based on the newly input aviation safety information text and its corresponding cause table, the best matching cause of the accident is obtained using the Ball-tree nearest neighbor algorithm, and the method is as follows: Splitting the newly input aviation safety information text and its corresponding cause table; After the text is split, the cause table corresponding to the newly input aviation safety information text is vectorized item by item and sentence by sentence through the SBERT model to obtain a cause table word vector matrix, and the newly input aviation safety information text is vectorized sentence by sentence to obtain an input text vector matrix; Standardizing the text vectors in the cause table word vector matrix and the input text vector matrix; A ball-tree is constructed based on the standardized word vector matrix of the cause table. The standardized input text vector matrix is ​​used as input. The ball-tree formed by each input text vector and a cause table word vector performs a nearest neighbor search. The n causes with the closest Euclidean distance are selected from the output of the nearest neighbor search as the best matching causes of the accident.

8. An intelligent extraction system for general information of aviation safety reports based on accident cause table, used to automatically obtain the best matching cause of aviation accidents, characterized in that: The system includes: A data acquisition module, configured to acquire aviation accident reports, translate and correct the aviation accident reports, and construct a Chinese aviation data set; A preprocessing module, configured to preprocess the data in the Chinese aviation data set to obtain a preprocessed text; A clustering module, configured to convert the preprocessed text into a numerical vector through an SBERT model, and perform clustering using a single-pass incremental clustering algorithm to generate an accident cause cluster; A hierarchical division module is configured to perform subject naming and hierarchical division on the accident cause clusters to obtain a table of causes of various types of accidents; A fine-tuning training module, configured to perform fine-tuning training on the BERT classification model; The cause acquisition module is configured to classify the newly input aviation safety information text through the fine-tuned BERT classification model to obtain the accident type, and combine the cause tables of each type of accident to obtain the cause table corresponding to the newly input aviation safety information text; based on the newly input aviation safety information text and its corresponding cause table, use the Ball-tree nearest neighbor algorithm to obtain the best matching cause of the accident.

9. An electronic device, characterized in that: include: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor, the instructions being used to be executed by the processor to implement the method for intelligently extracting general information from aviation safety reports based on an accident cause table as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to be executed by the computer to implement the method for intelligently extracting general information from aviation safety reports based on an accident cause table as described in any one of claims 1 to 7.

Images