Method and device for generating sailing notice sample, electronic device and storage medium

Abstract

The application discloses a navigation notice sample generation method and device, electronic equipment and a storage medium, relates to the navigation notice data processing field or other related technical fields, and the method comprises the following steps: carrying out multi-level named entity recognition on navigation notice data in a navigation notice data set to obtain entity recognition results and type labels; constructing a label sequence based on the entity recognition results and the type labels, inputting the label sequence into an embedding model, and outputting an embedding vector corresponding to the navigation notice data; performing cluster analysis on all navigation notice messages based on the embedding vector corresponding to the navigation notice data, obtaining a plurality of clustering clusters, and constructing navigation notice samples based on the clustering clusters to obtain a navigation notice sample set. Intelligent analysis of navigation notices is highly dependent on high-quality samples, and the application solves the technical problems of low sample library construction efficiency and unreasonable distribution in the related art, which mainly relies on manual selection of samples for processing of navigation notices.

Description

Technical Field

[0001] This invention relates to the field of NOTAM data processing or other related technical fields. Specifically, it relates to a method, apparatus, electronic device and storage medium for generating NOTAM samples. Background Technology

[0002] With the rapid development of the global air transport industry, the volume of aeronautical intelligence data has exploded. According to statistics from the International Civil Aviation Organization (ICAO), millions of NOTAMs are issued globally each year, and this number continues to increase annually. NOTAMs are issued via telecommunications to inform flight personnel and those involved in flight operations about the establishment, status, or changes of aviation facilities, services, and procedures, as well as the occurrence and changes of hazardous situations. NOTAMs cover critical information such as airport operational status, airspace restrictions, changes in navigation facilities, and weather conditions; the accuracy and timeliness of their interpretation directly affect flight safety and operational efficiency.

[0003] However, due to the different writing habits of intelligence personnel in different countries and regions, and the lack of globally unified mandatory text standards, Notices to Airmen (NOANC) texts exhibit a high degree of heterogeneity and unstructured characteristics. This heterogeneity is mainly reflected in the following aspects: First, the forms of entity expression are diverse, and the same entity may have multiple abbreviations, full names, or codes; second, the sentence structure is complex and varied, and the same semantic meaning may be expressed through different grammatical structures; and third, there are a large number of highly ambiguous words.

[0004] Modern AI technology can effectively improve parsing efficiency, but its core capabilities rely on a large number of high-quality labeled samples. Among related technologies, the construction of traditional sample libraries mainly relies on manual selection, which is inefficient and has insufficient sample coverage, and cannot effectively support the needs of navigation notice parsing scenarios.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This invention provides a method, apparatus, electronic device, and storage medium for generating NOTAM samples, to at least solve the technical problems in related technologies where the processing of NOTAMs mainly relies on manual sample selection, resulting in low efficiency in sample library construction and unreasonable distribution.

[0007] According to one aspect of the present invention, a method for generating NOTAM samples is provided, comprising: acquiring a set of NOTAM messages and preprocessing the NOTAM messages in the set of NOTAM messages to obtain a set of NOTAM data; performing multi-level named entity recognition on the NOTAM data in the NOTAM data set to obtain entity recognition results and type labels; constructing a label sequence based on the entity recognition results and the type labels, and inputting the label sequence into an embedding model to output an embedding vector corresponding to the NOTAM data; performing cluster analysis on all the NOTAM messages based on the embedding vector corresponding to the NOTAM data to obtain multiple clusters, and constructing NOTAM samples based on each cluster to obtain a set of NOTAM samples.

[0008] Furthermore, the step of performing multi-level named entity recognition on the NOTAM data in the NOTAM data set to obtain entity recognition results and type labels includes: loading a geographic location library, a facility and equipment library, an action status library, a time expression library, a common phrase library, and a unit and value library to construct a domain knowledge base; based on the domain knowledge base, using a forward maximum matching algorithm and regular expressions to perform entity matching on the NOTAM data, identifying deterministic entities, and obtaining the type labels of the deterministic entities.

[0009] Furthermore, the step of performing entity matching on the NOTAM data using the forward maximum matching algorithm and regular expressions based on the domain knowledge base includes: constructing an index tree based on the domain knowledge base; performing forward maximum matching on the NOTAM data based on the index tree to identify term entities; and scanning the NOTAM data using a predefined regular expression rule base to identify fixed-format sentence entities.

[0010] Further, after performing entity matching on the NOTAM data using the forward maximum matching algorithm and regular expressions, the process includes: marking words that simultaneously match multiple named entities as ambiguous candidate words; extracting context data of the ambiguous candidate words, and constructing an input sequence based on the ambiguous candidate words and the context data; inputting the input sequence into a pre-trained bidirectional encoding model, and outputting the named entity probability distribution of the ambiguous candidate words through the bidirectional encoding model; selecting the named entity with the highest probability in the named entity probability distribution as the deterministic entity of the ambiguous candidate word, and combining the type label of the deterministic entity to obtain the entity recognition result and type label.

[0011] Further, the step of constructing a label sequence based on the entity recognition result and the type label, and inputting the label sequence into an embedding model to output the embedding vector corresponding to the NOTAM data includes: generating a preliminary label sequence based on the type label corresponding to the named entity in the entity recognition result; performing redundant compression on consecutively occurring identical type labels in the preliminary label sequence to obtain a standardized label sequence; inputting the standardized label sequence into a pre-trained embedding model, and using the feature learning capability of the embedding model to map the standardized label sequence into a high-dimensional dense vector to obtain the embedding vector corresponding to the NOTAM data.

[0012] Further, the step of performing cluster analysis on all the NOTAM messages based on the embedding vectors corresponding to the NOTAM data to obtain multiple clusters includes: Step 1, setting a candidate range for the number of clusters, and obtaining the number of candidate clusters by sequentially increasing the number of candidate clusters with a preset step size, executing a clustering algorithm for each candidate cluster number, and calculating the silhouette coefficient and the sum of squared errors within the cluster; Step 2, determining the target number of clusters K based on the silhouette coefficient and the sum of squared errors within the cluster, and selecting K cluster centroids using a preset initialization strategy, where K is a positive integer; Step 3, calculating the Euclidean distance from each embedding vector to each cluster centroid, assigning the embedding vector to the cluster corresponding to the cluster centroid with the smallest Euclidean distance, and updating each cluster centroid; repeating Step 3 until the maximum number of iterations is reached to obtain the converged K clusters.

[0013] Further, the step of constructing a NOTAM sample set based on each of the clusters includes: for each cluster, calculating the Euclidean distance between all embedding vectors in the cluster and the centroid of the cluster; selecting the NOTAM message corresponding to the embedding vector with the smallest Euclidean distance as the NOTAM sample of that cluster, and constructing the NOTAM sample set based on the NOTAM samples of all the clusters.

[0014] According to another aspect of the present invention, an apparatus for generating NOTAM samples is also provided, comprising: an acquisition unit, configured to acquire a set of NOTAM messages and preprocess the NOTAM messages in the set of NOTAM messages to obtain a set of NOTAM data; an identification unit, configured to perform multi-level named entity recognition on the NOTAM data in the set of NOTAM data to obtain entity recognition results and type labels; an output unit, configured to construct a label sequence based on the entity recognition results and the type labels, input the label sequence into an embedding model, and output an embedding vector corresponding to the NOTAM data; and a construction unit, configured to perform cluster analysis on all the NOTAM messages based on the embedding vectors corresponding to the NOTAM data to obtain multiple clusters, and construct NOTAM samples based on each cluster to obtain a set of NOTAM samples.

[0015] Furthermore, the identification unit includes: a first construction module, used to load a geographic location library, a facility and equipment library, an action status library, a time expression library, a common phrase library, and a unit and value library to construct a domain knowledge base; and a first matching module, used to perform entity matching on the NOTAM data based on the domain knowledge base using a forward maximum matching algorithm and regular expressions, identify deterministic entities, and obtain the type label of the deterministic entity.

[0016] Furthermore, the first matching module includes: a first construction submodule, used to construct an index tree based on the domain knowledge base; a first identification submodule, used to perform positive maximum matching on the NOTAM data based on the index tree to identify term entities; and a first scanning submodule, used to scan the NOTAM data using a predefined regular expression rule library to identify fixed-format sentence entities.

[0017] Furthermore, the first matching module further includes: a first tagging submodule, used to tag words that simultaneously match multiple named entities as ambiguous candidate words; a first extraction submodule, used to extract context data of the ambiguous candidate words and construct an input sequence based on the ambiguous candidate words and the context data; a first output submodule, used to input the input sequence into a pre-trained bidirectional encoding model and output the named entity probability distribution of the ambiguous candidate words through the bidirectional encoding model; and a first summarization submodule, used to select the named entity with the highest probability in the named entity probability distribution as the deterministic entity of the ambiguous candidate word, and summarize the entity recognition result and type label by combining the type label of the deterministic entity.

[0018] Further, the output unit includes: a first generation module, used to generate a preliminary label sequence based on the type labels corresponding to the named entities in the entity recognition result; a first compression module, used to perform redundant compression on the consecutively occurring identical type labels in the preliminary label sequence to obtain a standardized label sequence; and a first mapping module, used to input the standardized label sequence into a pre-trained embedding model, and through the feature learning capability of the embedding model, map the standardized label sequence into a high-dimensional dense vector to obtain the embedding vector corresponding to the navigation notification data.

[0019] Further, the construction unit includes: a first calculation module, used in step one to set a candidate range for the number of clusters, and to obtain the number of candidate clusters by sequentially increasing the number of candidate clusters with a preset step size, and to execute a clustering algorithm for each candidate cluster number to calculate the silhouette coefficient and the sum of squared errors within the cluster; a first determination module, used in step two to determine the target number of clusters K based on the silhouette coefficient and the sum of squared errors within the cluster, and to select K cluster centroids using a preset initialization strategy, where K is a positive integer; a first update module, used in step three to calculate the Euclidean distance from each embedding vector to each cluster centroid, to assign the embedding vector to the cluster corresponding to the cluster centroid with the smallest Euclidean distance, and to update each cluster centroid; and a first repetition module, used to repeatedly execute step three until the maximum number of iterations is reached to obtain the converged K clusters.

[0020] Further, the construction unit includes: a second calculation module, used to calculate the Euclidean distance between all embedding vectors in the cluster and the centroid of the cluster for each cluster; and a second construction module, used to select the NOTAM message corresponding to the embedding vector with the smallest Euclidean distance as the NOTAM sample of the cluster, and construct the NOTAM sample set based on the NOTAM samples of all the clusters.

[0021] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform any of the above-described methods for generating a NOTAM sample.

[0022] According to another aspect of the present invention, an electronic device is also provided, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any of the above-described methods for generating a NOTAM sample.

[0023] In this application, the following steps are performed: obtaining a set of NOTAM messages, preprocessing the NOTAM messages in the NOTAM message set to obtain a NOTAM data set, performing multi-level named entity recognition on the NOTAM data set to obtain entity recognition results and type labels, constructing a label sequence based on the entity recognition results and type labels, inputting the label sequence into an embedding model, outputting the embedding vector corresponding to the NOTAM data, performing cluster analysis on all NOTAM messages based on the embedding vector corresponding to the NOTAM data to obtain multiple clusters, and constructing NOTAM samples based on each cluster to obtain a set of NOTAM samples.

[0024] This application constructs a fully automated technical chain encompassing preprocessing, multi-level named entity recognition, structured tag sequence generation, vectorization, and cluster mining. First, it utilizes a multi-level named entity recognition mechanism combining rule-based pre-identification based on a domain knowledge base with deep learning-based ambiguity resolution to efficiently and accurately extract entities and eliminate polysemy. Then, the identified entities are converted into type labels, constructing a tag sequence to highlight the deep structural features of the text. Subsequently, an embedding model maps the tag sequence to a vector space and performs cluster analysis, automatically mining a set of highly structurally representative typical samples from massive datasets. This achieves efficient, low-cost, and high-precision automated generation of typical NOTAM samples, improving processing efficiency and ensuring high coverage and parsing accuracy of samples across various message structure patterns. It also enhances the efficiency of sample library construction and the rationality of sample distribution. This solves the technical problems in related technologies where NOTAM processing mainly relies on manual sample selection, resulting in low sample library construction efficiency and unreasonable distribution. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0026] Figure 1 This is a flowchart of an optional method for generating a NOTAM sample according to an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of an optional NOTAM sample generation process according to an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of an optional clustering process according to an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of an optional NOTAM sample generation apparatus according to an embodiment of the present invention;

[0030] Figure 5 This is a hardware structure block diagram of an electronic device (or mobile device) that performs a method for generating NOTAM samples according to an embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] To facilitate understanding of the present invention by those skilled in the art, some terms or nouns involved in the various embodiments of the present invention are explained below:

[0034] BERT is a pre-trained language model based on the Transformer architecture. Its core mechanism is to capture the bidirectional dependencies between all lexical units in the input sequence through a self-attention layer, thereby constructing a deep representation rich in contextual semantic information.

[0035] It should be noted that the method and apparatus for generating NOTAM samples in this application can be used in the field of NOTAM data processing when generating NOTAM samples, and can also be used in any field other than the field of NOTAM data processing when generating NOTAM samples. This application does not limit the application field of the method and apparatus for generating NOTAM samples.

[0036] It should be noted that the relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) involved in this application are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of such data all comply with the relevant laws, regulations, and standards of the relevant regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, this system has interfaces with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface, and the relevant information is obtained only after receiving consent from the aforementioned user or organization.

[0037] It should be noted that in this application, when collecting and analyzing customer information, users are provided with corresponding operation entry points to choose whether to agree to or reject the automated decision-making results; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0038] The following embodiments of the present invention can be applied to various systems / applications / devices for generating NOTAM samples. The present invention constructs a fully automated technical chain encompassing preprocessing, multi-level named entity recognition, structured tag sequence generation, vectorization, and cluster mining. First, it utilizes a multi-level named entity recognition mechanism combining rule-based pre-identification based on a domain knowledge base with deep learning ambiguity resolution to efficiently and accurately extract entities and eliminate polysemy. Then, the identified entities are converted into type tags, constructing tag sequences to highlight the deep structural features of the text. Subsequently, an embedding model maps the tag sequences to a vector space and performs cluster analysis, automatically mining a set of highly structurally representative typical samples from massive datasets. This achieves efficient, low-cost, and high-precision automated generation of typical NOTAM samples, improving processing efficiency and ensuring high coverage and parsing accuracy of samples across various message structure patterns, thereby enhancing the efficiency of sample library construction and the rationality of sample distribution.

[0039] The present invention will now be described in detail with reference to various embodiments.

[0040] Example 1

[0041] According to an embodiment of the present invention, an embodiment of a method for generating a NOTAM sample is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0042] Figure 1This is a flowchart of an optional method for generating a NOTAM sample according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0043] Step S101: Obtain the set of NOTAM messages and preprocess the NOTAM messages in the set to obtain the NOTAM data set.

[0044] In step S101 above, NOTAM messages are collected to obtain a set of NOTAM messages. NOTAM messages are issued by telecommunications to inform flight personnel and personnel related to flight operations about the establishment, status or changes of aviation facilities, services and procedures, as well as the occurrence and changes of dangerous situations for aviation.

[0045] After obtaining the set of NOTAM (Notices to Airmen) messages, preprocessing is required. First, the original messages undergo rigorous format validation, removing invalid data that does not conform to the standard NOTAM format, such as missing headers or footers, to ensure the validity of subsequent processing. Next, data cleaning is performed. This involves correcting line break errors caused by transmission or storage, repairing abnormally split words, and ensuring the integrity of text units. Simultaneously, the telegraphic code is translated into readable text, and for messages containing multiple languages, language identification is performed, retaining English content or performing standardized conversion, and removing non-target language interference. The cleaned, translated, and filtered message data is then integrated to form the NOTAM dataset for subsequent processing.

[0046] The above steps aim to eliminate noise, missing and redundant information in the original data, standardize the data format, and provide a high-quality, standardized input data foundation for subsequent high-precision named entity recognition and structural analysis, thereby improving the robustness of the overall technical solution.

[0047] Step S102: Perform multi-level named entity recognition on the NOTAM data in the NOTAM dataset to obtain entity recognition results and type labels.

[0048] In step S102 above, multi-level entity matching is used to perform multi-level named entity recognition on the NOTAM data in the NOTAM dataset. Specifically, this includes entity recognition based on regular expression matching and entity recognition based on bidirectional coding model, forming complete entity recognition results and type labels. The multi-level recognition mechanism ensures efficiency by processing deterministic entities through rules and ensures accuracy by solving the polysemy problem through deep learning, thereby providing accurate basic data for the subsequent construction of structured labels.

[0049] In one optional embodiment, the steps of performing multi-level named entity recognition on the NOTAM data in the NOTAM database to obtain entity recognition results and type labels include: loading a geographic location database, a facility and equipment database, an action status database, a time expression database, a common phrase database, and a unit and value database to construct a domain knowledge base; based on the domain knowledge base, using a forward maximum matching algorithm and regular expressions to perform entity matching on the NOTAM data, identifying deterministic entities, and obtaining type labels for the deterministic entities.

[0050] Furthermore, when performing entity recognition, a domain knowledge base is first constructed by loading a geolocation database, a facility and equipment database, an action state database, a time expression database, a common phrase database, and a unit and value database. The geolocation database includes information such as standard airport codes and navigation beacon codes; the facility and equipment database contains aviation facility terms such as runways and taxiways; the action state database summarizes state terms such as construction, maintenance, and malfunctions; the time expression database organizes standard time formats; the common phrase database summarizes fixed collocations; and the unit and value database includes units of measurement such as altitude and distance. By integrating the above data, a structured support system is formed, providing a basis for subsequent rule matching.

[0051] Subsequently, based on the domain knowledge base, the forward maximum matching algorithm and regular expressions are used to perform entity matching on the NOTAM data. The forward maximum matching algorithm prioritizes matching the longest term by constructing a trie index to ensure the integrity of technical terms, while regular expressions identify patterns with fixed formats but difficult to exhaustively enumerate, such as runway numbers, coordinates, and time ranges, thereby covering sentence structures not included in the knowledge base.

[0052] During the matching process, entities that match only a single rule or dictionary entry are directly identified as deterministic entities and their type labels are obtained. Candidate words that match multiple rules or dictionaries are marked as ambiguous candidate words and enter the subsequent processing level. This process quickly processes high-confidence entities through rule pre-identification and improves the recognition coverage and accuracy by leveraging the synergistic complementarity of dictionaries and regular expressions, providing a precise basis for subsequent ambiguity resolution.

[0053] In one optional embodiment, the steps of performing entity matching on NOTAM data using a forward maximum matching algorithm and regular expressions based on a domain knowledge base include: constructing an index tree based on the domain knowledge base; performing forward maximum matching on the NOTAM data based on the index tree to identify term entities; and scanning the NOTAM data using a predefined regular expression rule base to identify fixed-format sentence entities.

[0054] Furthermore, when performing entity matching, an index tree is first constructed based on the domain knowledge base. Standard terms and codes from the geographic location database, facility and equipment database, action status database, time expression database, common phrase database, and unit and value database are organized into a hierarchical index tree. This index tree stores the prefix features of terms through tree nodes, supports fast retrieval and matching, and ensures efficient location of potential entities when processing massive amounts of NOTAM data.

[0055] The text content is then scanned, and the index tree is used to find the longest matching path starting from the current character. Once the longest matching item is found, it is confirmed as a term entity and output. This matching mechanism prioritizes the integrity of professional terms such as airport codes and facility codes, and avoids incorrectly splitting compound words. For example, “RWY18 / 36” is identified as a runway number entity as a whole instead of being split into runway and number.

[0056] Simultaneously, a predefined regular expression rule base is applied to scan the NOTAM data to identify entities with fixed sentence structures. The regular expression rule base covers runway number patterns, coordinate patterns, time range patterns, and flight altitude layer patterns, etc. By pattern matching, entities with fixed grammatical structures that are not in the dictionary are identified, such as coordinate strings composed of numbers and letters or time intervals with specific formats, thereby making up for the shortcomings of dictionary matching in covering non-standard but formatted entities.

[0057] By employing a dual mechanism of index tree matching and regular expression scanning, efficient and comprehensive identification of deterministic entities is achieved, providing accurate basic data for subsequent entity type labeling and ambiguity resolution.

[0058] In one optional embodiment, after performing entity matching on the NOTAM data using the forward maximum matching algorithm and regular expressions, the method includes: marking words that match multiple named entities simultaneously as ambiguous candidate words; extracting context data of the ambiguous candidate words and constructing an input sequence based on the ambiguous candidate words and context data; inputting the input sequence into a pre-trained bidirectional encoding model and outputting the named entity probability distribution of the ambiguous candidate words through the bidirectional encoding model; selecting the named entity with the highest probability in the named entity probability distribution as the deterministic entity of the ambiguous candidate word, and combining the type label of the deterministic entity to obtain the entity recognition result and type label.

[0059] Furthermore, after using the forward maximum matching algorithm and regular expressions to perform entity matching on the NOTAM data, words that match multiple named entities simultaneously during the matching process are marked as ambiguous candidate words. This mechanism aims to accurately locate text fragments where the rule matching results are uncertain due to polysemy or the generality of abbreviations, thereby limiting complex calculations to the minimum necessary range.

[0060] Subsequently, contextual data within a certain window range to the left and right of the ambiguous candidate word is extracted, and an input sequence is constructed by combining the ambiguous candidate word itself. This input sequence separates the context from the word to be disambiguated using a specific identifier to preserve complete semantic contextual information. The input sequence is then fed into a pre-trained bidirectional encoding model, which uses its bidirectional attention mechanism to deeply capture the semantic associations of the context. Through forward and backward propagation calculations, it outputs the probability distribution of the ambiguous candidate word belonging to each predefined named entity type. Finally, the named entity with the highest probability in the named entity probability distribution is selected as the definitive entity of the ambiguous candidate word, and its corresponding type label is associated with it. The labels after deep learning disambiguation are summarized with the labels determined by the previous rule matching to obtain the complete entity recognition result and type label.

[0061] This hierarchical processing strategy leverages the high efficiency of rule matching for deterministic entities while using deep learning models to resolve ambiguities in complex contexts, achieving a balance between recognition accuracy and processing efficiency.

[0062] Step S103: Construct a tag sequence based on the entity recognition results and type labels, input the tag sequence into the embedding model, and output the embedding vector corresponding to the navigation notification data.

[0063] In step S103 above, the specific entity values ​​in the identified and disambiguated NOTAM text are replaced with predefined structured type labels, thereby mapping the original text from the semantic space to the structural space. For example, specific airport codes or runway numbers are uniformly replaced with corresponding abstract labels. At the same time, compressed labeling is used for consecutively appearing entities of the same type to remove redundant information, forming a sequence composed of discrete type labels. Subsequently, the constructed label sequence is input into the embedding model, which uses its pre-trained feature learning ability to encode the label sequence. By transforming the discrete label sequence into a high-dimensional dense numerical vector, the NOTAM data that are structurally similar are close in distance in the vector space, thereby capturing the deep structural patterns of the text rather than the surface lexical differences. Finally, the embedding vector corresponding to the NOTAM data is output. This process provides a low-dimensional dense representation foundation that can reflect the essential structural features of the text for subsequent clustering analysis based on structural similarity.

[0064] In one optional embodiment, the steps of constructing a label sequence based on entity recognition results and type labels, and inputting the label sequence into an embedding model to output the embedding vector corresponding to the NOTAM data include: generating a preliminary label sequence based on the type labels corresponding to the named entities in the entity recognition results; performing redundant compression on consecutively occurring identical type labels in the preliminary label sequence to obtain a standardized label sequence; inputting the standardized label sequence into a pre-trained embedding model, and using the feature learning capability of the embedding model to map the standardized label sequence into a high-dimensional dense vector to obtain the embedding vector corresponding to the NOTAM data.

[0065] Furthermore, the specific entity values ​​identified in the original NOTAM text are replaced with their respective type labels. For example, specific airport codes are replaced with airport code labels, and specific runway numbers are replaced with runway number labels, thus forming a preliminary sequence composed of discrete labels.

[0066] Subsequently, redundant compression is performed on consecutively occurring labels of the same type in the initial label sequence to obtain a standardized label sequence. This process effectively reduces sequence length and redundant information by merging consecutively repeated labels of the same type into one label and marking the number, making the text representation more focused on structural patterns rather than repetitive details.

[0067] The standardized label sequence is input into a pre-trained embedding model, which is trained on a large-scale corpus and has a strong semantic capture capability. Through its internal multi-layer neural network structure, the model extracts and transforms the label sequence, mapping the discrete symbol sequence into a numerical representation in a continuous high-dimensional dense vector space, thus obtaining the embedding vector corresponding to the navigation notice data. This vector can preserve the structural similarity information of the original text, making navigation notices with the same structure close in distance in the vector space, providing a basis for subsequent clustering analysis.

[0068] Step S104: Based on the embedding vectors corresponding to the NOTAM data, perform cluster analysis on all NOTAM messages to obtain multiple clusters, and construct NOTAM samples based on each cluster to obtain a NOTAM sample set.

[0069] In step S104 above, geometric distance in vector space is used to characterize the similarity of text structures, and clustering algorithms in statistical learning are used to automatically group data and extract typical representatives. Since the preceding steps have replaced specific entities in the NOTAM texts with structured tag sequences and mapped them to high-dimensional dense vectors through an embedding model, NOTAMs with the same or similar structures but different specific vocabulary have a small Euclidean distance in vector space, while messages with significant structural differences are farther apart. The clustering analysis algorithm iteratively calculates the distance between sample vectors and cluster centers, automatically merging spatially close vectors into the same cluster, thus dividing massive heterogeneous NOTAM data into multiple subsets with inherent structural consistency. Samples within each cluster share common structural patterns. By selecting the sample closest to the centroid in each cluster as the core typical sample, one or more of the most representative messages can be extracted from each structural pattern group. This clustering and sampling mechanism based on structural similarity can cover the structural types of most NOTAMs with a very small number of samples, effectively filter redundant information, retain key structural features, and thus construct a NOTAM sample set that is both highly representative and has a high compression rate, providing high-quality basic data support for downstream automated parsing, model training, or rule construction.

[0070] In one optional embodiment, the steps of performing cluster analysis on all NOTAM messages based on the embedding vectors corresponding to NOTAM data to obtain multiple clusters include: Step 1, setting a candidate range for the number of clusters, and sequentially increasing the number of candidate clusters with a preset step size, performing a clustering algorithm on each candidate cluster number, and calculating the silhouette coefficient and the sum of squared errors within the cluster; Step 2, determining the target number of clusters K based on the silhouette coefficient and the sum of squared errors within the cluster, and selecting K cluster centroids using a preset initialization strategy, where K is a positive integer; Step 3, calculating the Euclidean distance from each embedding vector to each cluster centroid, assigning the embedding vector to the cluster corresponding to the cluster centroid with the smallest Euclidean distance, and updating each cluster centroid; repeating Step 3 until the maximum number of iterations is reached to obtain the converged K clusters.

[0071] Furthermore, when performing sample clustering, a candidate range for the number of clusters is first set, and the number of candidate clusters is obtained by sequentially increasing the number of candidate clusters with a preset step size. For each candidate cluster, a clustering algorithm is executed to calculate the silhouette coefficient and the sum of squared errors within the cluster. The silhouette coefficient is used to evaluate the density of the sample with its own cluster and the separation from other clusters in the clustering result. The sum of squared errors within the cluster reflects the sum of squared distances from the sample point within the cluster to the cluster center. By comprehensively considering these two indicators, the target number of clusters K is determined. The K value corresponding to the peak of the silhouette coefficient is selected first. If the silhouette coefficient changes slowly, the K value corresponding to the elbow point where the rate of decrease of the sum of squared errors within the cluster slows down is selected. If none of the above strategies give a clear optimal value, the K value with the silhouette coefficient closest to 1 is selected, thereby ensuring the rationality of the number of clusters.

[0072] Subsequently, a preset initialization strategy is used to select K cluster centroids. This strategy intelligently selects the initial cluster centroids from the dataset to avoid the algorithm getting trapped in local optima. The Euclidean distance from each embedding vector to each cluster centroid is calculated, and the embedding vector is assigned to the cluster corresponding to the centroid with the smallest Euclidean distance. The centroid of each cluster is then recalculated, which is the mean vector of all sample vectors within the cluster. The assignment and update steps are then repeated until the change in centroid position is less than the threshold or the maximum number of iterations is reached, resulting in K converged clusters.

[0073] The above steps automatically optimize the optimal cluster size and iteratively optimize the cluster structure, so that navigation notice data with similar structures are grouped into the same cluster, laying a solid foundation for the subsequent extraction of typical samples.

[0074] In one optional embodiment, the step of constructing a NOTAM sample set based on each cluster includes: for each cluster, calculating the Euclidean distance between all embedding vectors in the cluster and the centroid of the cluster; selecting the NOTAM message corresponding to the embedding vector with the smallest Euclidean distance as the NOTAM sample of that cluster, and constructing a NOTAM sample set based on the NOTAM samples of all clusters.

[0075] Furthermore, when constructing typical NOTAM samples, for each cluster, the Euclidean distance between all embedding vectors within the cluster and the centroid of the cluster is calculated. The Euclidean distance is used to quantify the degree of deviation between the structural features of a single NOTAM message and the overall structural features of the cluster. The smaller the distance, the higher the structural similarity between the message and other messages within the cluster, and the more representative it is of the core structural pattern of the cluster.

[0076] Subsequently, the NOTAM message corresponding to the embedding vector with the smallest Euclidean distance is selected as the NOTAM sample for this cluster. This selection mechanism ensures that the extracted samples are structurally the most typical and can reflect the common structural logic of all messages in the cluster to the greatest extent, avoiding structural distortion caused by selecting marginal or abnormal samples.

[0077] Finally, a set of NOTAM samples was constructed based on the NOTAM samples of all clusters. Typical samples selected from each cluster were summarized to form a high-quality sample set covering multiple structural patterns. This set, by selecting core representative samples, achieved dimensionality reduction and refinement from massive raw data to refined structural samples, providing a data foundation with clear structure and strong representativeness for subsequent downstream tasks.

[0078] Furthermore, the aforementioned NOTAM samples and their corresponding original messages (or standardized messages) are output to form a high-quality typical sample set. This sample set can be directly used for downstream tasks: as a high-quality knowledge source for retrieval enhancement generation, providing accurate structured references for large models; as high-quality training data for fine-tuning large models, improving the model's semantic understanding ability in the NOTAM domain; and as a benchmark example for constructing automated parsing rules.

[0079] Through the above steps, a set of NOTAM messages is obtained, and the NOTAM messages in the set are preprocessed to obtain a NOTAM data set. Multi-level named entity recognition is performed on the NOTAM data set to obtain entity recognition results and type labels. A label sequence is constructed based on the entity recognition results and type labels, and the label sequence is input into the embedding model to output the embedding vector corresponding to the NOTAM data. Cluster analysis is performed on all NOTAM messages based on the embedding vectors corresponding to the NOTAM data to obtain multiple clusters. NOTAM samples are constructed based on each cluster to obtain a set of NOTAM samples.

[0080] In this embodiment, a fully automated technical chain is constructed, encompassing preprocessing, multi-level named entity recognition, structured tag sequence generation, vectorization, and cluster mining. First, a multi-level named entity recognition mechanism combining rule-based pre-identification based on a domain knowledge base and deep learning disambiguation is utilized to efficiently and accurately extract entities and eliminate polysemy. Then, the identified entities are converted into type tags, constructing tag sequences to highlight the deep structural features of the text. Subsequently, an embedding model maps the tag sequences to a vector space and performs cluster analysis, automatically mining a set of highly structurally representative typical samples from massive datasets. This achieves efficient, low-cost, and high-precision automated generation of typical NOTAM samples, improving processing efficiency and ensuring high coverage and parsing accuracy of samples across various message structure patterns. It also enhances the efficiency of sample library construction and the rationality of sample distribution. This solves the technical problems in related technologies where NOTAM processing mainly relies on manual sample selection, resulting in low sample library construction efficiency and unreasonable distribution.

[0081] The following describes in detail another optional implementation method.

[0082] Figure 2 This is a schematic diagram illustrating an optional NOTAM sample generation process according to an embodiment of the present invention, such as... Figure 3 As shown, the process for generating a NOTAM sample includes:

[0083] Step 1: Obtain the NOTAM (Notice to Airmen) message;

[0084] Notices to Airmen (NOA) messages from the Aeronautical Information Region (AIR) were selected and categorized into four types based on their structural complexity and frequency of occurrence: common structure, complex and diverse structure, single structure, and rare structure. This ensures that the test sample comprehensively covers different scenarios. The specific categorization criteria are as follows: Common structure messages are regular messages with high frequency and fixed sentence structures, used to verify the method's ability to capture mainstream patterns; Complex and diverse structure messages are highly heterogeneous messages with rich sentence variations and numerous entity types, used to test the method's structural abstraction and generalization capabilities in complex scenarios; Single structure messages are simple messages with highly repetitive sentence structures and entity types, used to evaluate the method's ability to filter redundant information; Rare structure messages are marginal messages with extremely low frequency of occurrence, used to verify the method's coverage of long-tail patterns.

[0085] Step two, data processing;

[0086] Step 3: Multi-level named entities;

[0087] First, six core knowledge bases are constructed: a geolocation base containing standard geographic information such as airport four-letter codes, airport three-letter codes, navigation beacon codes, and information area codes; a facilities and equipment base containing standard terminology and codes for facilities such as runways, taxiways, navigation lights, and communication frequencies; an action status base summarizing action and status terms such as closed, active, under construction, and under maintenance; a time expression base organizing standard formats such as time formats, time ranges, and periodic expressions; a common phrase base summarizing phrase structures such as fixed collocations; and a unit and value base containing unit expressions such as altitude, distance, and frequency. Each knowledge base is stored using a hash table or index tree structure to support fast retrieval.

[0088] The first layer of the multi-level named entity is the knowledge base and rule pre-identification.

[0089] First, dictionary matching is performed. The system loads a geolocation database and scans the text. For example, the first "SAT" matches the airport three-letter code entry in the geolocation database and is identified as a candidate airport code; the second "SAT" also matches an airport three-letter code entry. Simultaneously, "AD" matches airport identifiers in the facilities and equipment database; "CLSD" matches the "closing" action in the action status database; "DUE TO" matches the reason-guided phrase in the common phrase database; and "MAINT" matches the maintenance reason in the action status database.

[0090] Then, regular expression matching is performed, with the system applying a predefined regular expression rule library for scanning. For example, the second "SAT" matches the weekday pattern rule and is identified as a weekday candidate; simultaneously, "1300-1700" matches the time range pattern \d{4}-\d{4} and is identified as a time range entity; the results are then processed separately. For "AD", "CLSD", "DUETO", "MAINT", and "1300-1700", all are unique matches, and the system directly outputs the entity label. For the two "SAT" entries, each matches multiple rules (dictionary matching for airport code, regular expression matching for weekday), the system marks them as ambiguity candidates and passes them along with their context information to the second layer.

[0091] The second layer of the multi-level named entity is deep learning ambiguity resolution.

[0092] The system processes the two ambiguous candidate "SAT" input from the first layer separately. First, it extracts the context window (3 to 5 words before and after) for each "SAT": For the first "SAT" in option E: the left context is "E"), and the right context is "AD CLSDDUE TO MAINT"; for the second "SAT" in option E: the left context is "MAINT.", and the right context is "1300-1700". This context is then used to construct the input sequence, which is fed into a pre-trained bidirectional coding model, such as the BERT model. The model outputs the probability distribution of each candidate type: For the first "SAT": the airport code has the highest probability and is determined as an airport code; for the second "SAT": the weekday has the highest probability and is determined as a weekday.

[0093] Step 4: Entity type annotation and label sequence construction;

[0094] All entities identified in the first and second layers are normalized and replaced with predefined type labels to construct a label sequence.

[0095] For consecutive occurrences of the same type of entity, a redundancy compression strategy is applied. For example, for the text "TWY T2,TWYT3,TWY L,TWY J CLOSED", after recognition, four consecutive {TWY_NO} tags are obtained, which the system compresses into "{TWY_NO}". "4 CLOSED" reduces text redundancy and highlights structural features.

[0096] An embedding model is selected, with the vector dimension set to 1024. The standardized label sequence is input into the embedding model to generate embedding vectors. For example, the standardized text "{RUN_NO}CLSD DUE TO{REASON}" generates vector V1=[0.68, 0.32,0.51,...,0.47]; the text with the same structure "{RUN_NO}UNAVAILABLE DUE TO{REASON}" generates vector V2=[0.65,0.35,0.49,...,0.45]. The two vectors are close in position in space.

[0097] Step 5: Sequence vectorization;

[0098] Step six, cluster analysis;

[0099] All vectorized texts are used as input, and clustering analysis is performed using a clustering algorithm.

[0100] Figure 3 This is a schematic diagram of an optional clustering process according to an embodiment of the present invention, such as... Figure 3 As shown, the cluster analysis process includes:

[0101] Step 1: Obtain the set of text embedding vectors;

[0102] Step 2, initial centroid selection;

[0103] First, set the candidate range of K value [K_min, K_max]. Based on the general characteristics of NOTAM data, K_min is set to 100, and K_max is set to the square root level of the data volume, increasing sequentially with a step size of step=100.

[0104] For each candidate K value, a clustering algorithm is executed. After the clustering converges, two evaluation metrics are calculated: the silhouette coefficient and the sum of squared errors within the cluster (SSE).

[0105] To determine the optimal K value, we prioritize the K value corresponding to the peak of the profile coefficient. If the profile coefficient changes gradually, we select the K value corresponding to the elbow point where the SSE descent rate slows down. If none of the above strategies provide a clear optimal value, we select the K value with the profile coefficient closest to 1.

[0106] The centroid is initialized based on a determined K value.

[0107] Step 3, Sample allocation;

[0108] Calculate the Euclidean distance from each sample vector to the cluster centroid, and assign the sample to the nearest cluster.

[0109] Step 4, Centroid update;

[0110] Based on the allocation results, the centroid of each cluster is recalculated, which is the mean vector of all sample vectors within the cluster.

[0111] Step 5: Iterative calculations yield the clusters.

[0112] Repeat the above steps until the change in the centroid position is less than the threshold or the maximum number of iterations is reached, thereby completing the sample clustering and obtaining multiple clusters.

[0113] Step 7: Construct typical samples.

[0114] For each cluster, calculate the Euclidean distance from all sample vectors within the cluster to the cluster centroid, select the sample with the smallest distance as the typical sample of the cluster, and construct the navigation notice sample set.

[0115] This invention constructs a fully automated technical chain encompassing preprocessing, multi-level named entity recognition, structured tag sequence generation, vectorization, and cluster mining. First, it utilizes a multi-level named entity recognition mechanism combining rule-based pre-identification based on a domain knowledge base with deep learning-based ambiguity resolution to efficiently and accurately extract entities and eliminate polysemy. Then, the identified entities are converted into type tags, constructing tag sequences to highlight the deep structural features of the text. Subsequently, an embedding model maps the tag sequences to a vector space and performs cluster analysis, automatically mining a set of highly structurally representative typical samples from massive datasets. This achieves efficient, low-cost, and high-precision automated generation of typical navigational notice samples, improving processing efficiency and ensuring high coverage and parsing accuracy of samples across various message structure patterns, thereby enhancing the efficiency of sample library construction and the rationality of sample distribution. A detailed description follows with another embodiment.

[0116] Example 2

[0117] The apparatus for generating a NOTAM sample provided in this embodiment includes multiple implementation units, each of which corresponds to a specific implementation step in Embodiment 1 above. The specific implementation method and its beneficial effects can be referred to the aforementioned method embodiment, and will not be repeated here.

[0118] Figure 4 This is a schematic diagram of an optional NOTAM sample generation apparatus according to an embodiment of the present invention, such as... Figure 4 As shown, the apparatus for generating the NOTAM sample may include: an acquisition unit 41, an identification unit 42, an output unit 43, and a construction unit 44, wherein,

[0119] The acquisition unit 41 is used to acquire a set of NOTAM messages and preprocess the NOTAM messages in the set of NOTAM messages to obtain a set of NOTAM data.

[0120] The identification unit 42 is used to perform multi-level named entity recognition on the NOTAM data in the NOTAM data set to obtain entity recognition results and type labels;

[0121] Output unit 43 is used to construct a label sequence based on entity recognition results and type labels, input the label sequence into the embedding model, and output the embedding vector corresponding to the navigation notification data;

[0122] The construction unit 44 is used to perform cluster analysis on all NOTAM messages based on the embedding vectors corresponding to the NOTAM data, obtain multiple clusters, and construct NOTAM samples based on each cluster to obtain a NOTAM sample set.

[0123] The aforementioned apparatus for generating NOTAM samples comprises: an acquisition unit 41 for acquiring a set of NOTAM messages and preprocessing the NOTAM messages in the set to obtain a NOTAM data set; an identification unit 42 for performing multi-level named entity recognition on the NOTAM data in the NOTAM data set to obtain entity recognition results and type labels; an output unit 43 for constructing a label sequence based on the entity recognition results and type labels, inputting the label sequence into an embedding model, and outputting the embedding vector corresponding to the NOTAM data; and a construction unit 44 for performing cluster analysis on all NOTAM messages based on the embedding vectors corresponding to the NOTAM data to obtain multiple clusters, and constructing NOTAM samples based on each cluster to obtain a set of NOTAM samples.

[0124] In this embodiment, a fully automated technical chain is constructed, encompassing preprocessing, multi-level named entity recognition, structured tag sequence generation, vectorization, and cluster mining. First, a multi-level named entity recognition mechanism combining rule-based pre-identification based on a domain knowledge base and deep learning-based ambiguity resolution is utilized to efficiently and accurately extract entities and eliminate polysemy. Then, the identified entities are converted into type labels, constructing a tag sequence to highlight the deep structural features of the text. Subsequently, an embedding model maps the tag sequence to a vector space and performs cluster analysis, automatically mining a set of highly structurally representative typical samples from massive datasets. This achieves efficient, low-cost, and high-precision automated generation of typical navigation notice samples, improving processing efficiency and ensuring high coverage and parsing accuracy of samples across various message structure patterns, thereby enhancing the efficiency of sample library construction and the rationality of sample distribution. Furthermore, the identification unit includes: a first construction module, used to load a geolocation library, a facility and equipment library, an action status library, a time expression library, a common phrase library, and a unit and value library to construct a domain knowledge base; and a first matching module, used to perform entity matching on the NOTAM data based on the domain knowledge base using a forward maximum matching algorithm and regular expressions, identify deterministic entities, and obtain type labels for deterministic entities.

[0125] Furthermore, the first matching module includes: a first construction submodule for constructing an index tree based on a domain knowledge base; a first identification submodule for performing positive maximum matching on the NOTAM data based on the index tree to identify term entities; and a first scanning submodule for scanning the NOTAM data using a predefined regular expression rule base to identify fixed-format sentence entities.

[0126] Furthermore, the first matching module also includes: a first labeling submodule, used to label words that match multiple named entities simultaneously as ambiguous candidate words; a first extraction submodule, used to extract context data of ambiguous candidate words and construct an input sequence based on the ambiguous candidate words and context data; a first output submodule, used to input the input sequence into a pre-trained bidirectional encoding model and output the named entity probability distribution of the ambiguous candidate words through the bidirectional encoding model; and a first summarization submodule, used to select the named entity with the highest probability in the named entity probability distribution as the deterministic entity of the ambiguous candidate word, and combine it with the type label of the deterministic entity to summarize the entity recognition result and type label.

[0127] Furthermore, the output unit includes: a first generation module, used to generate a preliminary label sequence based on the type labels corresponding to the named entities in the entity recognition results; a first compression module, used to perform redundant compression on the consecutively occurring labels of the same type in the preliminary label sequence to obtain a standardized label sequence; and a first mapping module, used to input the standardized label sequence into a pre-trained embedding model, and through the feature learning capability of the embedding model, map the standardized label sequence into a high-dimensional dense vector to obtain the embedding vector corresponding to the navigation notification data.

[0128] Further, the construction unit includes: a first calculation module, used in step one, to set the candidate range of the number of clusters, and to obtain the number of candidate clusters by sequentially increasing the number of candidate clusters with a preset step size, and to execute a clustering algorithm for each candidate cluster number to calculate the silhouette coefficient and the sum of squared errors within the cluster; a first determination module, used in step two, to determine the target number of clusters K based on the silhouette coefficient and the sum of squared errors within the cluster, and to select K cluster centroids using a preset initialization strategy, where K is a positive integer; a first update module, used in step three, to calculate the Euclidean distance from each embedding vector to each cluster centroid, to assign the embedding vector to the cluster corresponding to the cluster centroid with the smallest Euclidean distance, and to update each cluster centroid; and a first repetition module, used to repeat step three until the maximum number of iterations is reached to obtain the converged K clusters.

[0129] Furthermore, the construction unit includes: a second calculation module, used to calculate the Euclidean distance between all embedding vectors within the cluster and the centroid of the cluster for each cluster; and a second construction module, used to select the NOTAM message corresponding to the embedding vector with the smallest Euclidean distance as the NOTAM sample for that cluster, and to construct a NOTAM sample set based on the NOTAM samples of all clusters.

[0130] The above-mentioned apparatus for generating navigation notice samples may also include a processor and a memory. The above-mentioned acquisition unit 41, identification unit 42, output unit 43, construction unit 44, etc. are all stored in the memory as program units, and the processor executes the above-mentioned program units stored in the memory to realize the corresponding functions.

[0131] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and flight notification samples can be generated by adjusting kernel parameters.

[0132] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0133] According to another aspect of the present invention, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute any of the above-described methods for generating a NOTAM sample.

[0134] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any of the above-described methods for generating a NOTAM sample.

[0135] According to another aspect of the present invention, a computer program product is also provided, the computer program product including a computer program, wherein the computer program, when executed by a processor, implements the method for generating any of the above-described NOTAM samples.

[0136] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program having the following method steps: acquiring a set of NOTAM messages, and preprocessing the NOTAM messages in the NOTAM message set to obtain a NOTAM data set; performing multi-level named entity recognition on the NOTAM data in the NOTAM data set to obtain entity recognition results and type labels; constructing a label sequence based on the entity recognition results and type labels, and inputting the label sequence into an embedding model to output the embedding vector corresponding to the NOTAM data; performing cluster analysis on all NOTAM messages based on the embedding vector corresponding to the NOTAM data to obtain multiple clusters, and constructing NOTAM samples based on each cluster to obtain a NOTAM sample set.

[0137] This application also provides a computer program product that, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: loading a geographic location library, a facility and equipment library, an action state library, a time expression library, a common phrase library, and a unit and value library to construct a domain knowledge base; based on the domain knowledge base, using a forward maximum matching algorithm and regular expressions to perform entity matching on the NOTAM data, identifying deterministic entities, and obtaining type labels for the deterministic entities.

[0138] This application also provides a computer program product that, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: constructing an index tree based on a domain knowledge base; performing a forward maximum match on the NOTAM data based on the index tree to identify term entities; and applying a predefined regular expression rule base to scan the NOTAM data to identify fixed-format sentence entities.

[0139] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: marking words that simultaneously match multiple named entities as ambiguous candidate words; extracting context data of the ambiguous candidate words, constructing an input sequence based on the ambiguous candidate words and context data; inputting the input sequence into a pre-trained bidirectional encoding model, outputting the named entity probability distribution of the ambiguous candidate words through the bidirectional encoding model; selecting the named entity with the highest probability in the named entity probability distribution as the deterministic entity of the ambiguous candidate word, and combining the type label of the deterministic entity to obtain the entity recognition result and type label.

[0140] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: generating a preliminary label sequence based on the type labels corresponding to named entities in the entity recognition results; performing redundant compression on consecutively occurring labels of the same type in the preliminary label sequence to obtain a standardized label sequence; inputting the standardized label sequence into a pre-trained embedding model, and using the feature learning capability of the embedding model, mapping the standardized label sequence into a high-dimensional dense vector to obtain the embedding vector corresponding to the NOTAM data.

[0141] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following steps: Step 1, setting a candidate range for the number of clusters, and sequentially increasing the number of candidate clusters with a preset step size, executing a clustering algorithm for each candidate cluster number, and calculating the silhouette coefficient and the sum of squared errors within the cluster; Step 2, determining the target number of clusters K based on the silhouette coefficient and the sum of squared errors within the cluster, and selecting K cluster centroids using a preset initialization strategy, where K is a positive integer; Step 3, calculating the Euclidean distance from each embedding vector to each cluster centroid, assigning the embedding vector to the cluster corresponding to the cluster centroid with the smallest Euclidean distance, and updating each cluster centroid; repeating Step 3 until the maximum number of iterations is reached, and obtaining the converged K clusters.

[0142] This application also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following method steps: for each cluster, calculating the Euclidean distance between all embedding vectors within the cluster and the centroid of the cluster; selecting the NOTAM message corresponding to the embedding vector with the smallest Euclidean distance as the NOTAM sample for that cluster, and constructing a NOTAM sample set based on the NOTAM samples of all clusters.

[0143] Figure 5 This is a hardware structure block diagram of an electronic device (or mobile device) that performs a method for generating NOTAM samples according to an embodiment of the present invention. Figure 5 As shown, an electronic device may include one or more processors ( Figure 5 The processor, denoted by 502a, 502b, ..., 502n, can include, but is not limited to, a processing device such as a microprocessor (MCU) or a programmable logic device (FPGA), and a memory 505 for storing data. In addition, it may include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports in the I / O interface), a network interface, a keyboard, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 5The structure shown is for illustrative purposes only and does not limit the structure of the electronic device described above. For example, the electronic device may also include... Figure 5 The more or fewer components shown, or having the same Figure 5 The different configurations shown.

[0144] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0145] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0146] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0147] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0148] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0149] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0150] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating a sample of a NOTAM (Notification of Navigation) notice, characterized in that, include: Obtain a set of NOTAM messages and preprocess the NOTAM messages in the set to obtain a set of NOTAM data. Multi-level named entity recognition is performed on the NOTAM data in the NOTAM dataset to obtain entity recognition results and type labels; A tag sequence is constructed based on the entity recognition result and the type label, and the tag sequence is input into the embedding model to output the embedding vector corresponding to the flight notification data; Cluster analysis is performed on all the NOTAM messages based on the embedding vectors corresponding to the NOTAM data to obtain multiple clusters, and NOTAM samples are constructed based on each cluster to obtain a NOTAM sample set.

2. The method according to claim 1, characterized in that, The steps for performing multi-level named entity recognition on the NOTAM data in the NOTAM dataset to obtain entity recognition results and type labels include: Load geolocation database, facility and equipment database, action status database, time expression database, common phrase database, and unit and value database to build a domain knowledge base; Based on the domain knowledge base, the forward maximum matching algorithm and regular expressions are used to perform entity matching on the NOTAM data, identify deterministic entities, and obtain the type labels of the deterministic entities.

3. The method according to claim 2, characterized in that, Based on the aforementioned domain knowledge base, the steps for entity matching of the NOTAM data using the forward maximum matching algorithm and regular expressions include: Construct an index tree based on the aforementioned domain knowledge base; Based on the index tree, perform a positive maximum match on the NOTAM data to identify term entities; The navigation notification data is scanned using a predefined regular expression rule base to identify fixed-format sentence entities.

4. The method according to claim 2, characterized in that, After performing entity matching on the NOTAM data using the forward maximum matching algorithm and regular expressions, the process includes: Words that match multiple named entities simultaneously are marked as ambiguous candidate words; Extract the context data of the ambiguous candidate words, and construct an input sequence based on the ambiguous candidate words and the context data; The input sequence is fed into a pre-trained bidirectional encoding model, and the bidirectional encoding model outputs the named entity probability distribution of the ambiguous candidate words. The named entity with the highest probability in the named entity probability distribution is selected as the deterministic entity of the ambiguous candidate word. The entity recognition result and type label are then obtained by combining the type label of the deterministic entity.

5. The method according to claim 1, characterized in that, The steps of constructing a label sequence based on the entity recognition result and the type label, inputting the label sequence into the embedding model, and outputting the embedding vector corresponding to the NOTAM data include: Based on the type labels corresponding to the named entities in the entity recognition results, a preliminary label sequence is generated; Redundant compression is performed on consecutively occurring tags of the same type in the preliminary tag sequence to obtain a standardized tag sequence; The standardized label sequence is input into a pre-trained embedding model. Through the feature learning capability of the embedding model, the standardized label sequence is mapped into a high-dimensional dense vector to obtain the embedding vector corresponding to the navigation notification data.

6. The method according to claim 1, characterized in that, The step of performing cluster analysis on all the NOTAM messages based on the embedding vectors corresponding to the NOTAM data to obtain multiple clusters includes: Step 1: Set the candidate range for the number of clusters, and obtain the candidate clusters by increasing the step size sequentially. For each candidate cluster, execute the clustering algorithm to calculate the silhouette coefficient and the sum of squared errors within the cluster. Step 2: Determine the target number of clusters K based on the contour coefficient and the sum of squared errors within the cluster, and select K cluster centroids using a preset initialization strategy, where K is a positive integer; Step 3: Calculate the Euclidean distance from each embedding vector to the centroid of each cluster, assign the embedding vector to the cluster corresponding to the centroid of the cluster with the smallest Euclidean distance, and update the centroid of each cluster. Repeat step three until the maximum number of iterations is reached, and obtain the K converged clusters.

7. The method according to claim 6, characterized in that, The steps for constructing a NOTAM sample set based on each of the aforementioned clusters include: For each cluster, calculate the Euclidean distance between all embedding vectors within the cluster and the centroid of the cluster; The NOTAM message corresponding to the embedding vector with the smallest Euclidean distance is selected as the NOTAM sample of the cluster, and the NOTAM sample set is constructed based on the NOTAM samples of all the clusters.

8. An apparatus for generating a sample of a NOTAM (Notification to Airmen), characterized in that, include: The acquisition unit is used to acquire a set of NOTAM messages and preprocess the NOTAM messages in the set of NOTAM messages to obtain a set of NOTAM data. The identification unit is used to perform multi-level named entity recognition on the NOTAM data in the NOTAM data set to obtain entity recognition results and type labels. The output unit is used to construct a tag sequence based on the entity recognition result and the type label, input the tag sequence into the embedding model, and output the embedding vector corresponding to the flight notification data; The construction unit is used to perform cluster analysis on all the NOTAM messages based on the embedding vectors corresponding to the NOTAM data, obtain multiple clusters, and construct NOTAM samples based on each cluster to obtain a NOTAM sample set.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method for generating a NOTAM sample according to any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for generating a NOTAM sample as described in any one of claims 1 to 7.

Images