Classification method, device and electronic equipment for NOTAM messages

Abstract

This application discloses a method, device, and electronic device for classifying NOTAM messages, relating to the field of artificial intelligence. The method comprises: extracting raw feature data from a target NOTAM message; inputting the raw feature data into a first model; using the first model to determine, based on the raw feature data, the probability distribution information of the categories into which the target NOTAM message belongs; using the category probability distribution information as target feature data for the target NOTAM message; and determining the target type of the target NOTAM message based on the raw feature data and the target feature data. This application addresses the technical problem of low classification efficiency in prior art NOTAM message classification methods based on expert rules.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and more specifically, to a classification method, device, and electronic equipment for NOTAM messages. Background Art

[0002] NOTAMs are crucial documents used in aviation to convey critical data such as flight safety, airport conditions, and route information. With the continuous growth of global air traffic, the number of NOTAMs has increased dramatically, and their content has become increasingly complex and diverse. Effectively parsing and classifying NOTAMs is crucial to ensuring the safety and timeliness of aviation operations. However, traditional expert rule-based approaches face significant technical challenges and limitations when handling this task.

[0003] For example, in expert-rule-based classification systems, rule development relies on the experience and expertise of domain experts. This is not only a time-consuming process but also requires a deep understanding of aviation NOTAMs. Experts must manually identify and define key words and phrases associated with specific capability items, such as "refueling," "weather," and "runway." This type of rule development can easily miss uncommon or context-specific keywords, compromising the comprehensiveness and accuracy of the classification.

[0004] Furthermore, once a rule set is defined, it is typically static and difficult to adapt quickly to dynamic changes in NOTAM content or emerging vocabulary. Aviation is a highly regulated and constantly evolving industry, with new terms, abbreviations, or standards frequently introduced. Expert rule-based systems struggle to keep up to date with these rules, causing classification results to lag behind real-world needs, leading to low classification efficiency.

[0005] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0006] The embodiments of the present application provide a method, device, and electronic device for classifying NOTAM messages, so as to at least solve the technical problem of low classification efficiency in the prior art when classifying NOTAM messages based on expert rules.

[0007] According to one aspect of an embodiment of the present application, a method for classifying NOTAM messages is provided, comprising: extracting original feature data of a target NOTAM message; inputting the original feature data of the target NOTAM message into a first model, and determining, by the first model, category probability distribution information when the target NOTAM message belongs to various categories of messages based on the original feature data; using the category probability distribution information as target feature data of the target NOTAM message; and determining a target type of the target NOTAM message based on the original feature data and the target feature data.

[0008] Optionally, the target type of the target navigation notice message is determined based on the original feature data and the target feature data, including: splicing the original feature data and the target feature data into a target feature matrix; inputting the target feature matrix into a second model, analyzing the correlation and data patterns between different data in the target feature matrix through the second model, and determining the target type of the target navigation notice message based on the analysis results.

[0009] Optionally, the output layer of the first model is connected to the input layer of the second model, and the analysis accuracy of the first model for multidimensional linear data is higher than the analysis accuracy of the second model for multidimensional linear data; the analysis accuracy of the second model for the correlation relationship of different feature data is higher than the analysis accuracy of the first model for the correlation relationship of different feature data.

[0010] Optionally, the training step of the first model includes: obtaining N historical NOTAM messages that have been classified, wherein N is an integer greater than 1; converting the N historical NOTAM messages into a vector set, wherein the vector set includes N high-dimensional vectors, and the N high-dimensional vectors correspond one-to-one to the N historical NOTAM messages, and each high-dimensional vector is used to represent feature data of multiple dimensions included in a historical NOTAM message; performing a target deletion operation on the vector set to obtain a target vector set, wherein the target deletion operation is used to delete abnormal vectors in the vector set, and the abnormal vectors include: vectors converted from historical NOTAM messages with a text length less than a preset length, and vectors converted from historical NOTAM messages including preset characters; and training the neural network according to the target vector set to obtain the first model.

[0011] Optionally, a neural network is trained according to a target vector set to obtain a first model, including: calculating a mean vector of the target vector set; calculating a difference between each vector in the target vector set and the mean vector to obtain a difference vector; calculating a transposed vector of the difference vector; determining a covariance matrix of the target vector set according to the difference vector and the transposed vector; and training a neural network according to the covariance matrix of the target vector set to obtain the first model.

[0012] Optionally, a neural network is trained according to the covariance matrix of the target vector set to obtain a first model, including: determining multiple eigenvalues ​​included in the covariance matrix and eigenvectors corresponding to each eigenvalue, wherein each eigenvalue corresponds to a historical navigation notice message; selecting eigenvectors corresponding to at least one largest eigenvalue from the multiple eigenvalues ​​included in the covariance matrix to form a projection vector matrix; projecting the target vector set to a low-dimensional space according to the projection vector matrix to obtain a low-dimensional eigenvector set; and training the neural network according to the low-dimensional eigenvector set to obtain the first model.

[0013] Optionally, a neural network is trained based on a low-dimensional feature vector set to obtain a first model, including: determining the type of the historical navigation notice message corresponding to each vector in the low-dimensional feature vector set as the training label corresponding to the vector; splitting the low-dimensional feature vector set into K subsets, where K is an integer greater than 1; predicting the type of the historical navigation notice message corresponding to the vector in each subset through a neural network based on the vector in the subset to obtain a prediction result for each vector in each subset; determining the prediction accuracy of the neural network for each subset based on the prediction result of each vector in each subset and the training label corresponding to each vector; and training the neural network based on the prediction accuracy of the neural network for each subset to obtain the first model.

[0014] Optionally, the neural network is trained according to the prediction accuracy of the neural network for each subset to obtain a first model, including: calculating the average prediction accuracy of the K subsets; performing multiple iterative operations on the neural network according to the average prediction accuracy of the K subsets until the average prediction accuracy of the neural network for the K subsets is higher than a preset threshold, ending the iterative operation, and using the neural network after the last iterative operation as the first model, wherein the iterative operation is used to adjust the first parameter and the second parameter of the neural network, the first parameter is used to control the degree of error term penalty of the neural network, and the second parameter is used to control the neural network's similarity measurement method between different data.

[0015] Optionally, determining the category probability distribution information of the target Notify Navigation message when it belongs to each category of messages based on the original feature data through the first model includes: determining the decision value when the target Notify Navigation message belongs to each category of messages based on the original feature data through the first model, wherein the decision value is used to characterize the degree of proximity between the original feature data and each category of messages; determining the probability that the target Notify Navigation message belongs to each category of messages based on the decision value when the target Notify Navigation message belongs to each category of messages and a constant bias item set for each category of messages, wherein the constant bias item is used to adjust the method of converting the decision value of each category of messages into a probability; determining the category probability distribution information of the target Notify Navigation message belonging to each category of messages based on the probability that the target Notify Navigation message belongs to each category of messages.

[0016] Optionally, the training step of the second model includes: splicing the category probability distribution information determined by the first model for the historical NOTAM messages and the original feature data of the historical NOTAM messages into a training feature data set; inputting the training feature data set into the initial deep neural network, and updating the weight matrices of the hidden layer and output layer of the initial deep neural network according to the message category predicted by the initial deep neural network based on the training feature data set and the actual category of the historical NOTAM messages, until the message category predicted by the initial deep neural network based on the training feature data set and the actual category of the historical NOTAM messages are the same, thereby determining to obtain the second model.

[0017] Optionally, the second model includes an input layer, at least two hidden layers, a target network layer and an output layer, wherein the input layer is used to receive a training feature data set, and the dimension of the input layer is consistent with the dimension of the training feature data set; the at least two hidden layers are used to extract data features in the training feature data set step by step and analyze the relationship between different data features; the target network layer is used to randomly control some neurons in the initial deep learning network to be in an inactivated state; and the output layer is used to output the message category predicted by the initial deep neural network based on the training feature data set.

[0018] According to another aspect of an embodiment of the present application, a classification device for NOTAM messages is provided, which includes: a feature extraction unit for extracting original feature data of a target NOTAM message; a first processing unit for inputting the original feature data of the target NOTAM message into a first model, and determining, through the first model, category probability distribution information when the target NOTAM message belongs to various types of messages based on the original feature data; a second processing unit for using the category probability distribution information as target feature data of the target NOTAM message; and a determination unit for determining a target type of the target NOTAM message based on the original feature data and the target feature data.

[0019] According to another aspect of an embodiment of the present application, a computer-readable storage medium is further provided, wherein a computer program is stored in the computer-readable storage medium. When the computer program is executed, the device where the computer-readable storage medium is located executes the above-mentioned classification method for navigation notice messages.

[0020] According to another aspect of an embodiment of the present application, an electronic device is also provided, wherein the electronic device 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 one or more processors, the one or more processors execute the above-mentioned classification method of navigation notice messages.

[0021] Traditionally, classification methods based on expert rules require manual identification and definition of keywords and rule sets. This approach is not only time-consuming, but the classification efficiency is directly affected by the size of the rule set. In contrast, the present application uses the first model to enable the classification system to further deepen the learning of the original feature data of the target NOTAM message, especially to output the category probability distribution information when the target NOTAM message belongs to each category of messages. By combining the category probability distribution information with the original feature data to generate target feature data, the scope of feature utilization is broadened, so that classification decisions are no longer limited to simple keyword matching, but are based on an in-depth understanding of the message content and multi-dimensional analysis. This comprehensive feature utilization method greatly enhances the intelligence and adaptability of the classification system, reduces excessive reliance on expert rules, and improves the overall efficiency of the classification task. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0023] Figure 1 is a flowchart of an optional classification method for NOTAM messages according to an embodiment of the present application;

[0024] Figure 2 This is an architectural diagram of an optional multi-feature fusion NOTAM message classification method according to an embodiment of the present application;

[0025] Figure 3 This is a business flow chart of an optional multi-feature fusion NOTAM message classification method according to an embodiment of the present application;

[0026] Figure 4 This is a schematic diagram of an optional classification device for NOTAM messages according to an embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] It should also be noted that the 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, displayed data, etc.) collected by this application are information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with the relevant laws, regulations and standards of the relevant regions, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation portals for users to choose to authorize or refuse. For example, an interface is set up between this system and the relevant users or institutions. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or institution through the interface, and obtain the relevant information after receiving the consent information fed back by the aforementioned user or institution.

[0030] In an optional embodiment, a classification system for NOTAM messages (hereinafter referred to as the classification system) may serve as the execution entity of the NOTAM message classification method according to the embodiments of the present application. The classification system may be a software system or an embedded system combining software and hardware. Of course, those skilled in the art will appreciate that, in addition to using the classification system as the execution entity, other forms, such as the execution entity of an apparatus or device, may also be used to execute the NOTAM message classification method. The embodiments of the present application do not particularly limit the specific form of the method execution entity.

[0031] For the sake of convenience, the following describes the solution using the classification system as the execution entity.

[0032] According to an embodiment of the present application, an embodiment of a method for classifying navigation notice messages is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0033] Figure 1 is a flow chart of an optional classification method of NOTAM messages according to an embodiment of the present application, such as Figure 1 As shown, the method includes the following steps:

[0034] Step S101: extracting original feature data of the target NOTAM message.

[0035] Optionally, NOTAM messages are essentially text data. After receiving the target NOTAM message, the classification system can first perform text vectorization on the target NOTAM message. After text vectorization, the data can be cleaned to remove null or invalid values ​​generated during the vectorization process. This is because if null values ​​or meaningless vectors exist, they will not only not provide any valuable information for classification, but may also interfere with the model's learning process. After cleaning the data, the data can also be standardized, which includes converting the message category name into a numerical label, such as marking "information notification capability" as the number "0", "refueling capability" as "1", and so on.

[0036] Optionally, after vectorization and normalization, the message is represented as a high-dimensional numerical vector. However, high-dimensional feature data often introduces issues with computational efficiency and model generalization. To address this, dimensionality reduction can be performed on the feature data to identify key features in the original data while removing redundant information, thereby reducing data dimensionality and improving computational efficiency.

[0037] In step S102, the original feature data of the target NOTAM message is input into the first model, and the first model determines the category probability distribution information of the target NOTAM message when it belongs to each category of messages based on the original feature data.

[0038] Optionally, the classification system may input raw feature data into the first model, which will attempt to find an optimal classification boundary based on the feature data to distinguish different capability categories of the message.

[0039] It should be noted that, in the present application, the first model may be a neural network model, including but not limited to: an SVM model, a random forest model or a gradient boosting tree model. The present application does not specifically limit the specific form of the first model.

[0040] Step S103: Using the category probability distribution information as target feature data of the target NOTAM message.

[0041] Optionally, when the first model classifies a target NOTAM message, it not only outputs a hard classification label (i.e., the message's most likely category) but also, through probability calibration techniques, outputs a probability distribution describing all possible categories to which the message belongs. This set of probabilities reflects the confidence level of the message's belonging to each category, providing richer and more detailed information than a single classification label.

[0042] In traditional feature engineering, feature data typically refers to numerical or categorical information extracted from raw data for model training and prediction. In this application, in addition to the raw feature data, the category probability distribution information output by the first model is also used as new target feature data, thereby enriching the available feature data.

[0043] For example, if a target NOTAM message can be classified into eight different capability categories, the first model will output a probability vector of length 8 for that message, where each element represents the probability of the message belonging to the corresponding category. This probability vector is then concatenated with the original feature data to form a new feature vector that contains both the original information and the category probability distribution information.

[0044] Step S104: determining the target type of the target NOTAM message based on the original characteristic data and the target characteristic data.

[0045] Optionally, Table 1 is an optional classification label and related explanation of the navigation notice message according to an embodiment of the present application.

[0046] Table 1

[0047]

[0048] Optionally, by integrating the original feature data with the target feature data, the classification system can utilize more comprehensive information, including direct text content features and confidence information for the preliminary classification of these features. This helps improve the accuracy and reliability of classification.

[0049] In an optional embodiment, the target type of the target NOTAM message is determined based on the original feature data and the target feature data, including: splicing the original feature data and the target feature data into a target feature matrix; inputting the target feature matrix into a second model, analyzing the correlation and data patterns between different data in the target feature matrix through the second model, and determining the target type of the target NOTAM message based on the analysis results.

[0050] Optionally, the classification system can obtain two key feature sets from the target NOTAM message: the original feature data, which is the text feature data after vectorization and dimensionality reduction; and the target feature data, which is the class probability distribution information output by the first model. The classification system can combine these two feature sets to form a richer and more comprehensive feature matrix, called the target feature matrix. All information contained in this matrix will be provided as input to the second model for deep feature learning and classification prediction.

[0051] In an optional embodiment, Figure 2 This is an architecture diagram of an optional multi-feature fusion NOTAM message classification method according to an embodiment of the present application, such as Figure 2 As shown in the figure, after obtaining training data (historical NOTAM messages), the training data is first vectorized and labeled to obtain feature data (represented in vector form) and the corresponding training labels (i.e., the actual classification of the historical NOTAM messages). Based on the feature data and training labels, a first model is trained (primarily by adjusting the first and second parameters of the neural network and the neural network weights). After parameter optimization of the first model's output data, the type probability distribution information for the historical NOTAM messages is obtained. This type probability distribution information is then fused with the original feature data, and the fusion result is fed into the initial deep neural network for training the second model.

[0052] It should be noted that the second model is a neural network model. For example, the second model can be a CNN (Convolutional Neural Network), an RNN (Recurrent Neural Network), an LSTM (Long Short-Term Memory Network) model, or a DNN (Deep Neural Network) model.

[0053] It should also be noted that the second model can handle highly complex data patterns and relationships between features. In this application, the second model is used as a tool to analyze and determine the target message type. The second model will receive the target feature matrix as input, and analyze the mutual influence and data patterns between different features in the matrix through its internal multi-layer structure. The second model can learn the deep associations between features and identify the potential connections between the original feature data and the target feature data. In addition, the second model can also handle complex nonlinear relationships between features through multi-layer nonlinear transformations, which is especially important in classification tasks because the semantic information of the message content often presents a nonlinear distribution.

[0054] Optionally, after analyzing the data in the target feature matrix, the second model will output a prediction to determine the target type of the target NOTAM message. This prediction is based on the model's understanding of the relationships between features and the classification patterns learned from the training data. It is a comprehensive judgment of the second model's message type.

[0055] By splicing the original feature data with the target feature data output by the first model and performing in-depth analysis using the second model, the type of target NOTAM messages can be accurately and efficiently classified.

[0056] In an optional embodiment, the output layer of the first model is connected to the input layer of the second model, and the analysis accuracy of the first model for multidimensional linear data is higher than the analysis accuracy of the second model for multidimensional linear data; the analysis accuracy of the second model for the correlation between different feature data is higher than the analysis accuracy of the first model for the correlation between different feature data.

[0057] Optionally, the first model acts as a preliminary classifier in the technical solution of this application, primarily for processing and analyzing the raw feature data, particularly those portions that can be considered multidimensional linear data. Because the first model is more adept at solving high-dimensional linear classification problems, it can effectively extract key linear boundaries from the raw feature data for distinguishing different message categories.

[0058] The first model is particularly suitable for processing linear data because it can find a better classification hyperplane and maintain good classification performance even when the feature space is very high-dimensional. In addition, the relevant kernel function of the first model can handle nonlinear classification problems. However, in this application, the first model first uses linear data as the main processing object because the main goal at this stage is to obtain category probability distribution information rather than to perform deep feature pattern learning.

[0059] Secondly, the second model plays the role of in-depth analysis and final classifier in the technical solution of this application. The second model receives the category probability distribution information output by the first model and the original feature data. Through multiple layers of nonlinear transformation, it can identify and analyze the complex correlations between different feature data.

[0060] Compared to the first model, the second model is better at handling nonlinear relationships and learning deep feature patterns. Its multi-layered structure allows it to abstract features layer by layer, starting with simple features at the bottom level and gradually building higher-level, more complex feature representations. This hierarchical feature learning capability enables the second model to capture nonlinear interactions between features, resulting in higher analytical accuracy when processing correlations between different types of feature data.

[0061] Optionally, in the present application, the output layer of the first model is directly connected to the input layer of the second model, so that the classification results of the first model (especially the category probability distribution information) will be input into the second model as additional features, and deep learning and classification will be performed together with the original feature data. By connecting the first model and the second model in series, the present application can achieve two-stage classification: preliminary linear classification and deep nonlinear classification. Among them, the first model provides a preliminary classification based on linear features, and the second model further explores the nonlinear features and cross-feature associations on this basis, and finally determines the target type of the message. This model series design not only takes advantage of the first model's ability to process linear data, but also utilizes the second model's ability to deeply explore nonlinear relationships, thereby improving the accuracy and robustness of classification as a whole.

[0062] In an optional embodiment, the training step of the first model includes the following steps:

[0063] N historical NOTAM messages that have been classified are obtained, where N is an integer greater than 1; the N historical NOTAM messages are converted into a vector set, where the vector set includes N high-dimensional vectors, the N high-dimensional vectors correspond one-to-one to the N historical NOTAM messages, and each high-dimensional vector is used to represent feature data of multiple dimensions included in a historical NOTAM message; a target deletion operation is performed on the vector set to obtain a target vector set, where the target deletion operation is used to delete abnormal vectors in the vector set, and the abnormal vectors include: vectors converted from historical NOTAM messages with a text length less than a preset length, and vectors converted from historical NOTAM messages including preset characters; and a neural network is trained according to the target vector set to obtain a first model.

[0064] Optionally, to train the first model, the classification system can first collect a certain number of historical NOTAM messages to ensure sufficient data support for model training. Next, the classification system converts each message into a high-dimensional vector. The construction of the high-dimensional vector is primarily based on the textual content of the message. For example, the classification system can utilize models such as the BGEM3 model and the TF-IDF model to capture semantic features in the message and represent them in numerical form for subsequent processing by machine learning algorithms.

[0065] Optionally, each vector in the vector set corresponds one-to-one to a historical NOTAM message, meaning that the features of each message are converted into a data point, and the entire message set is converted into a vector space.

[0066] After vector conversion is complete, the classification system performs a target removal operation to eliminate anomalous vectors that may interfere with subsequent analysis. These include vectors generated from messages with text lengths below a preset length and those containing predefined characters (such as irrelevant or malformed characters). This data cleaning operation is designed to improve model training efficiency and prediction accuracy. By removing these anomalous vectors, every element in the vector set is ensured to be valid and contain useful information, thereby constructing a purer target vector set for model training.

[0067] Optionally, after the target vector set is prepared, the next step is to train a neural network. The neural network will be used to process and analyze the target vector set and perform preliminary classification tasks. During this process, the neural network will learn the relationship between different features in the vector set and the message categories, attempting to find an optimal classification hyperplane to optimally classify messages into different categories.

[0068] From the above content, we can see that in order to improve the quality of data, data preprocessing is required before training the neural network through the algorithm. It mainly includes:

[0069] Message vectorization: Converting the text content of NOTAMs into high-dimensional vector representations so that machine learning algorithms can process this unstructured data. This process captures the semantic information in the text and enhances the performance of subsequent classification tasks.

[0070] Removing empty vectorized values: During the vectorization process, some messages may not generate valid vectors (for example, if the text is too short or contains special characters). These invalid samples need to be identified and removed to ensure the quality and consistency of the training dataset.

[0071] Dimensionality reduction: Reduce the dimensionality of high-dimensional vectors to retain the main features while reducing redundant information, speeding up calculations, avoiding overfitting problems, and improving model generalization capabilities.

[0072] Message category labeling: The categories of NOTAMs are converted into numerical labels, as shown in Table 1. Integer codes (0, 1, 2, etc.) are used to represent different message types, making them easier for supervised learning algorithms to understand and process.

[0073] In an optional embodiment, training a neural network based on a target vector set to obtain a first model includes: first calculating a mean vector of the target vector set, then calculating the difference between each vector in the target vector set and the mean vector to obtain a difference vector, and then calculating a transposed vector of the difference vector. Then, determining a covariance matrix of the target vector set based on the difference vector and the transposed vector, and training the neural network based on the covariance matrix of the target vector set to obtain the first model.

[0074] Optionally, the calculation formula of the covariance matrix M can refer to formula (1):

[0075]

[0076] In formula (1), represents the target vector set, yes The mean vector of , n represents the number of vectors in the target vector set, A transposed vector representing the difference vector.

[0077] It's important to note that by calculating the mean vector of the target vector set and finding the difference between each vector and the mean vector, the mean of the target vector set is adjusted to the origin, which helps optimize model performance and make the model more stable. Next, the transposed vector of the difference vector is calculated, which provides the necessary data preparation for the subsequent covariance matrix calculation. Using the covariance matrix and dimensionality reduction techniques, we can further reduce data dimensionality, remove redundant features, improve model computational efficiency, and reduce the risk of overfitting.

[0078] In addition, the covariance matrix contains statistical information about the target vector set, reflecting the interdependence between features. For neural networks, the statistical information in the covariance matrix helps the model find the optimal classification hyperplane, that is, the one that can best distinguish different types of NOTAM messages.

[0079] In an optional embodiment, a neural network is trained based on the covariance matrix of the target vector set to obtain a first model, including: determining multiple eigenvalues ​​included in the covariance matrix and eigenvectors corresponding to each eigenvalue, wherein each eigenvalue corresponds to a historical navigation notice message; selecting eigenvectors corresponding to at least one largest eigenvalue from the multiple eigenvalues ​​included in the covariance matrix to form a projection vector matrix; projecting the target vector set to a low-dimensional space based on the projection vector matrix to obtain a low-dimensional eigenvector set; and training the neural network based on the low-dimensional eigenvector set to obtain the first model.

[0080] Optionally, after obtaining the covariance matrix M, the classification system can select the eigenvectors corresponding to the first k largest eigenvalues ​​from the covariance matrix to form a projection vector matrix , and finally based on the projection vector matrix Project the target vector set V into a low-dimensional space ,in, is the feature of the input model after dimensionality reduction (i.e. the low-dimensional feature vector set mentioned above).

[0081] Alternatively, in a high-dimensional feature space, the covariance matrix reflects the relationships between features and their contributions to the overall variance. By extracting the eigenvalues ​​and eigenvectors of the covariance matrix, we can identify which features explain the most variance in the data—that is, which features carry the most information. Selecting the eigenvector corresponding to the largest eigenvalue preserves the primary direction of variation in the data, allowing the original dataset to be represented in a smaller dimension while preserving the most information.

[0082] Projecting the target vector set into a low-dimensional space composed of selected feature vectors can significantly reduce the time and computational resources required for model training. In high-dimensional space, data points are often sparsely distributed, which can lead to difficulties in model training and overfitting. Dimensionality reduction not only simplifies data representation but also makes it easier for neural networks to find classification boundaries, improving the model's generalization ability. Furthermore, projecting into a low-dimensional space helps remove noise and redundant features, allowing the model to focus on key features that have a decisive impact on the classification task, thereby improving classification accuracy.

[0083] Furthermore, in low-dimensional space, the task of finding the optimal classification boundary becomes clearer for the neural network. Based on the reduced low-dimensional feature vector set, the neural network can more easily identify the differences between different categories of NOTAM messages, thereby finding a hyperplane that maximizes the distance between the categories, improving the model's classification accuracy and robustness.

[0084] In an optional embodiment, a neural network is trained based on a low-dimensional feature vector set to obtain a first model, including: determining the type of the historical Notification of Navigation message corresponding to each vector in the low-dimensional feature vector set as a training label corresponding to the vector; splitting the low-dimensional feature vector set into K subsets, where K is an integer greater than 1; predicting the type of the historical Notification of Navigation message corresponding to the vector in each subset through a neural network based on the vector in the subset to obtain a prediction result for each vector in each subset; determining the prediction accuracy of the neural network for each subset based on the prediction result of each vector in each subset and the training label corresponding to each vector; and training the neural network based on the prediction accuracy of the neural network for each subset to obtain the first model.

[0085] Optionally, before training the low-dimensional feature vector set, each vector needs to be assigned a training label. The training label is derived from the true categories of the corresponding historical NOTAM messages, with the message category serving as the corresponding label for the feature vector. This ensures that the model learns the correct classification associations during training, enabling accurate prediction of unknown messages.

[0086] Optionally, to more comprehensively evaluate and optimize the performance of the neural network, the classification system can first randomly split the low-dimensional feature vector set into K subsets, where K is an integer greater than 1. This creates multiple training and validation sets, with each subset having a chance to become the validation set, while the remaining subsets serve as the training set. Through multiple training and validation cycles, the general performance of the neural network on different datasets can be determined, avoiding the random bias that may be introduced by a single validation set.

[0087] Optionally, the neural network uses each subset as a validation set and then predicts the vectors of the remaining subsets, generating predictions about the packet type corresponding to each vector. The predictions are then compared with the actual training labels to calculate the neural network's prediction accuracy for each subset. Accuracy refers to the proportion of packet categories correctly predicted by the model and is an important evaluation metric reflecting the model's classification performance.

[0088] Based on the performance of the neural network in each cross-validation step, the overall accuracy of the model can be gradually improved by optimizing the model parameters. This optimization process may include using optimization algorithms (including but not limited to Bayesian algorithms, annealing algorithms, ant colony algorithms, and others) to find the optimal parameter combination. Through K iterations, each using a different subset as the validation set, a neural network model is ultimately obtained that performs well on all subsets. This model can accurately classify NOTAM messages, maintaining high classification accuracy even for unseen messages.

[0089] Through dimensionality reduction and cross-validation, the model's learning process became more robust. Dimensionality reduction removed redundant features, while cross-validation tested the model's stability through different data splits, effectively preventing the model from overfitting, a phenomenon known as excessive learning, on the training set. The resulting first model, after optimization and validation, demonstrated excellent classification results on low-dimensional feature vector sets. This demonstrates that the model accurately captures and understands key information in the messages, maintaining high classification accuracy even when the feature dimensionality is reduced.

[0090] In an optional embodiment, a neural network is trained based on the prediction accuracy of the neural network for each subset to obtain a first model, including: calculating the average prediction accuracy of K subsets; performing multiple iterative operations on the neural network based on the average prediction accuracy of the K subsets until the average prediction accuracy of the neural network for the K subsets is higher than a preset threshold, ending the iterative operation, and using the neural network after the last iterative operation as the first model, wherein the iterative operation is used to adjust a first parameter and a second parameter of the neural network, the first parameter is used to control the degree of error term penalty of the neural network, and the second parameter is used to control the neural network's similarity measurement method between different data.

[0091] Optionally, during each round of training, the neural network uses one subset as the validation set and the remaining subsets as the training set. Classification predictions are then made on the validation set. After each prediction round, the prediction accuracy for that validation set is calculated, which is the proportion of correctly predicted categories that match the actual categories. After completing all K rounds of validation, the average of these K prediction accuracies is calculated, reflecting the overall performance of the neural network on the different data subsets.

[0092] The process of optimizing model parameters based on the average prediction accuracy of K subsets. The goal of this iterative process is to improve the model's prediction accuracy above a preset threshold. In each iteration, the neural network's first parameter (i.e., the degree of error penalty) and second parameter (which controls the similarity measure between different data) are adjusted. Adjusting these parameters influences the model's tolerance for training errors and its ability to capture nonlinear relationships, thereby impacting its generalization performance.

[0093] Optionally, the iteration process continues until the average prediction accuracy of the neural network for the K subsets exceeds a preset threshold. This threshold is an important performance metric that sets a minimum standard of accuracy that the model must achieve. Once the model's average accuracy exceeds this threshold, the iteration process stops, and the model is considered sufficiently optimized to be used as the first model for preliminary classification.

[0094] The first model, after sufficient iterative training, not only performs well when processing the current dataset, but also maintains stable classification performance when facing possible changes in data distribution in the future. This robustness and adaptability are particularly important in the rapidly evolving aviation field, ensuring the long-term effectiveness and safety of the classification system.

[0095] Alternatively, assuming that the first parameter is defined as C and the second parameter is defined as g, where , construct a Gaussian process proxy model to maximize the accuracy of cross-validation. The classification target of the classification function is set at the average accuracy of cross-validation, where the classification function can be seen in formula (2):

[0096]

[0097] in, It is the average of the K prediction accuracies obtained after K rounds of validation.

[0098] In an optional embodiment, the category probability distribution information of the target Notify Navigation message when it belongs to each category of messages is determined based on the original feature data through the first model, including: determining the decision value when the target Notify Navigation message belongs to each category of messages based on the original feature data through the first model, wherein the decision value is used to characterize the degree of proximity between the original feature data and each category of messages; determining the probability that the target Notify Navigation message belongs to each category of messages based on the decision value when the target Notify Navigation message belongs to each category of messages and a constant bias item set for each category of messages, wherein the constant bias item is used to adjust the method of converting the decision value of each category of messages into a probability; determining the category probability distribution information of the target Notify Navigation message belonging to each category of messages based on the probability that the target Notify Navigation message belongs to each category of messages.

[0099] Optionally, the raw feature data is first input into a trained and optimized first model. Using its learned classification boundaries and weights, the first model calculates the decision value for each category of the target NOTAM message. The decision value essentially reflects the distance or similarity between the message data and each category's classification boundary. For the first model, messages with larger decision values ​​are further from a category's classification boundary, while messages with smaller decision values ​​are closer to that boundary, potentially even within it.

[0100] Optionally, after obtaining the decision value, a constant bias term is introduced to perform probability conversion. In the first model, the decision value cannot be directly interpreted as a probability. Therefore, by adding a specific constant bias term and adjusting the decision value, it can be made to conform to the properties of probability, that is, the probability distribution is in the interval [0, 1], and the sum of all class probabilities is equal to 1. This conversion typically involves mapping the decision value to a sigmoid function or other probability conversion function to obtain a probability estimate for each class. For example, the softmax function can be used to convert the decision value to a probability distribution. The bias term acts like an offset in the adjustment function, affecting the mapping curve from decision value to probability, making the conversion more flexible and accurate.

[0101] Based on the above probability conversion, the target NOTAM message's probability for each category is ultimately determined. These probability values ​​form a category probability distribution, which intuitively expresses the likelihood that the message belongs to each category. For example, if a message has a high probability for "runway capability" but lower probabilities for categories like "refueling capability" and "weather / RVR capability," it can be preliminarily inferred that the message contains more runway-related information. This category probability distribution not only provides multiple possibilities for the message's possible classification but also provides a quantitative estimate for each possibility, providing a more comprehensive feature representation for subsequent deep learning models.

[0102] Optionally, the probability conversion process can refer to formula (3):

[0103]

[0104] In formula (3), B is a constant bias term, It represents the decision value when the first model determines that the target NOTAM message belongs to the i-th category message based on the original feature data.

[0105] It's important to note that introducing the class probability distribution between the first and second models can be considered an intermediate representation, helping to mitigate potential information loss between the two models and optimize the overall classification process. The first model provides preliminary classification results, while the second model refines them, forming a hierarchical and complementary classification system.

[0106] In an optional embodiment, the training step of the second model includes: splicing the category probability distribution information determined by the first model for the historical NOTAM messages and the original feature data of the historical NOTAM messages into a training feature data set; inputting the training feature data set into an initial deep neural network, and updating the weight matrices of the hidden layer and output layer of the initial deep neural network according to the message category predicted by the initial deep neural network based on the training feature data set and the actual category of the historical NOTAM messages, until the message category predicted by the initial deep neural network based on the training feature data set and the actual category of the historical NOTAM messages are the same, thereby determining to obtain the second model.

[0107] Optionally, to train the second model, the first model's category probability distribution information for historical NOTAM messages is concatenated with the original feature data of the messages to form a new training feature dataset. This step aims to use the classification output of the first model as additional features, providing richer information for training the second model. In this way, the initial deep neural network not only learns representations of the original feature data but also uses the predictions of the first model to understand which features are strongly associated with specific categories, thereby better utilizing this information in the deep learning process.

[0108] Next, the constructed training feature dataset is used to train the initial deep neural network. In each round of training, the original feature data and category probability distribution information are input into the network, which then predicts the message category based on the current weight matrix. The predicted category is then compared with the actual category of the message, and the prediction error is calculated. This error is fed back into the network, and the backpropagation algorithm is used to update the weight matrices of the hidden and output layers of the network to reduce the prediction error.

[0109] The training process continues until the initial deep learning network's predictions of packet categories based on the training feature dataset are highly consistent with the actual packet categories, or until a preset convergence criterion is reached. This process involves repeated adjustments to the network weights until the model achieves optimal performance, minimizing error in predicting packet categories. In practice, the convergence criterion typically sets an error threshold or a certain number of training rounds to ensure that the model neither overfits nor underperforms due to insufficient training.

[0110] In an optional embodiment, Figure 3 This is a business flow chart of an optional multi-feature fusion NOTAM message classification method according to an embodiment of the present application, such as Figure 3 As shown, historical NOTAM messages are first collected, and then the messages are vectorized to obtain a vector set. Subsequently, data preprocessing is performed on the vector set to obtain a target vector set. Subsequently, the target vector set is used to train a first model. By optimizing the parameters of the first model, type probability distribution information about the historical NOTAM messages is generated. The original feature data and type probability distribution information of the historical NOTAM messages are used to train a second model until the second model can accurately output the target category of the historical NOTAM messages.

[0111] As can be seen from the above, through initial deep neural network training, the model is able to automatically learn and extract deeper feature representations, which may contain complex relationships and patterns implicit in the original data. Compared with the first model, the second model is more adept at processing high-dimensional nonlinear data, capable of discovering more subtle feature correlations, thereby improving classification accuracy and robustness. By combining the category probability distribution information from the first model with the original feature data, the second model can integrate multi-level features for learning, thus overcoming the limitations of a single method and enhancing the model's overall understanding of message characteristics and classification capabilities.

[0112] In an optional embodiment, the second model includes an input layer, at least two hidden layers, a target network layer and an output layer, wherein the input layer is used to receive a training feature data set, and the dimension of the input layer is consistent with the dimension of the training feature data set; the at least two hidden layers are used to extract data features in the training feature data set step by step and analyze the relationship between different data features; the target network layer is used to randomly control some neurons in the initial deep learning network to be in an inactivated state; and the output layer is used to output the message category predicted by the initial deep neural network based on the training feature data set.

[0113] Alternatively, in the second model, the input layer is the first layer of the deep neural network, whose primary function is to receive the model's input data, namely the training feature dataset. In this application scenario, the dimensionality of the input layer matches the dimensionality of the training feature dataset, meaning that each set of input data is directly mapped to a neuron in the input layer, with each neuron representing a feature in the dataset.

[0114] Additionally, the second model includes at least two hidden layers. Hidden layers are the core component of deep neural networks, used to progressively extract features from input data and analyze the complex relationships between different features. Each additional hidden layer allows the network to extract higher-level features and capture nonlinear patterns in the data. In this application, at least two hidden layers are used to ensure that the information in the feature dataset can be fully mined and analyzed. Each hidden layer consists of multiple neurons, which recognize and extract features through learned weights (connection strengths), providing more abstract feature representations for the next layer.

[0115] The second model also features a target network layer. During training, this layer randomly "deactivates" some neurons in the network, removing them from participating in the current forward and backward propagation. This random deactivation mechanism creates distinct "sub-networks," forcing the network to learn more robust feature representations and reduce its reliance on specific neurons, thereby improving the model's generalization capabilities. During the prediction phase, the target network layer does not apply deactivation; instead, it scales down the outputs of all neurons to simulate the deactivation effect during training.

[0116] The output layer is the final layer of a deep neural network, responsible for converting the network's final feature representation into a prediction result, namely the message category. The output layer typically contains neurons equal to the number of categories in the classification task, and the output of each neuron represents the probability of a category. In multi-classification tasks, the output layer typically uses a softmax function to convert the neuron output into a probability distribution, such that the sum of the probabilities of each category is 1. This way, the prediction result not only includes a category label but also the probability of belonging to each category, providing a basis for more detailed classification and decision-making.

[0117] For example, select a batch of data sets Input the first model and use the test set probability distribution output by the first model as the new feature matrix Splice to dataset , forming a new data set . The dataset is divided into training set and test set in a ratio of 7:3.

[0118] The structure of the second model is as follows:

[0119] Input layer: The dimension is , corresponding to the data set dimension.

[0120] The hidden layer is set to two layers: , ,in, and Represent the weight matrices of the two hidden layers, and They represent the bias matrices of the two hidden layers respectively, and ReLU represents the activation function.

[0121] Set the target network layer to avoid overfitting: ,in, Represents the probability value of random neuron inactivation.

[0122] Output layer: ,in, represents the weight matrix of the output layer; Represents the bias matrix of the output layer.

[0123] Loss function: .

[0124] In the above loss function, n represents the number of samples in the data set, i represents the i-th sample in the data set, K is the number of categories, and j represents the j-th category. Represents the true probability that the i-th sample belongs to the j-th category, Represents the probability of the neural network predicting that the i-th sample belongs to the j-th category; represents the weight matrix of the output layer; and Represent the weight matrices of the two hidden layers respectively; is a preset constant.

[0125] According to another aspect of the embodiment of the present application, a classification device for NOTAM messages is provided, wherein: Figure 4 is a schematic diagram of an optional classification device for NOTAM messages according to an embodiment of the present application, such as Figure 4 As shown, the device includes: a feature extraction unit 401, used to extract original feature data of the target NOTAM message; a first processing unit 402, used to input the original feature data of the target NOTAM message into a first model, and determine the category probability distribution information of the target NOTAM message when it belongs to various types of messages based on the original feature data through the first model; a second processing unit 403, used to use the category probability distribution information as target feature data of the target NOTAM message; and a determination unit 404, used to determine the target type of the target NOTAM message based on the original feature data and the target feature data.

[0126] Optionally, the determination unit 404 includes: a first processing subunit, used to splice the original feature data and the target feature data into a target feature matrix; a second processing subunit, used to input the target feature matrix into a second model, analyze the correlation and data pattern between different data in the target feature matrix through the second model, and determine the target type of the target navigation notice message based on the analysis results.

[0127] Optionally, the output layer of the first model is connected to the input layer of the second model, and the analysis accuracy of the first model for multidimensional linear data is higher than the analysis accuracy of the second model for multidimensional linear data; the analysis accuracy of the second model for the correlation relationship of different feature data is higher than the analysis accuracy of the first model for the correlation relationship of different feature data.

[0128] Optionally, the classification device for NOTAM messages includes: an acquisition unit for acquiring N historical NOTAM messages that have been classified, wherein N is an integer greater than 1; a conversion unit for converting the N historical NOTAM messages into a vector set, wherein the vector set includes N high-dimensional vectors, and the N high-dimensional vectors correspond one-to-one to the N historical NOTAM messages, and each high-dimensional vector is used to represent feature data of multiple dimensions included in a historical NOTAM message; a preprocessing unit for performing a target deletion operation on the vector set to obtain a target vector set, wherein the target deletion operation is used to delete abnormal vectors in the vector set, and the abnormal vectors include: vectors converted from historical NOTAM messages with a text length less than a preset length, and vectors converted from historical NOTAM messages including preset characters; a training unit for training a neural network according to the target vector set to obtain a first model.

[0129] Optionally, the training unit includes: a first calculation subunit, used to calculate the mean vector of the target vector set; a second calculation subunit, used to calculate the difference between each vector in the target vector set and the mean vector to obtain a difference vector; a third calculation subunit, used to calculate the transposed vector of the difference vector; a first determination subunit, used to determine the covariance matrix of the target vector set based on the difference vector and the transposed vector; and a first training subunit, used to train the neural network based on the covariance matrix of the target vector set to obtain a first model.

[0130] Optionally, the first training subunit includes: a first determination module, used to determine multiple eigenvalues ​​included in the covariance matrix and the eigenvectors corresponding to each eigenvalue, wherein each eigenvalue corresponds to a historical navigation notice message; a first processing module, used to select the eigenvectors corresponding to at least one largest eigenvalue from the multiple eigenvalues ​​included in the covariance matrix to form a projection vector matrix; a second processing module, used to project the target vector set to a low-dimensional space according to the projection vector matrix to obtain a low-dimensional feature vector set; a training module, used to train the neural network according to the low-dimensional feature vector set to obtain a first model.

[0131] Optionally, the training module includes: a first determination submodule, used to determine the type of historical navigation notice message corresponding to each vector in the low-dimensional feature vector set as the training label corresponding to the vector; a first processing submodule, used to split the low-dimensional feature vector set into K subsets, where K is an integer greater than 1; a second processing submodule, used to predict the type of historical navigation notice message corresponding to the vector in each subset through a neural network based on the vector in the subset, and obtain a prediction result for each vector in each subset; a second determination submodule, used to determine the prediction accuracy of the neural network for each subset based on the prediction result of each vector in each subset and the training label corresponding to each vector; a training submodule, used to train the neural network according to the prediction accuracy of the neural network for each subset to obtain a first model.

[0132] Optionally, the training submodule includes: a first calculation submodule, used to calculate the average value of the prediction accuracy of K subsets; an iterative training submodule, used to perform multiple iterative operations on the neural network based on the average value of the prediction accuracy of the K subsets, until the average value of the prediction accuracy of the neural network for the K subsets is higher than a preset threshold, and then terminate the iterative operation, and use the neural network after the last iterative operation as the first model, wherein the iterative operation is used to adjust the first parameter and the second parameter of the neural network, the first parameter is used to control the degree of error term penalty of the neural network, and the second parameter is used to control the similarity measurement method of the neural network for different data.

[0133] Optionally, the first processing unit 402 includes: a decision value determination subunit, used to determine the decision value when the target navigation notice message belongs to each type of message based on the original characteristic data through the first model, wherein the decision value is used to characterize the degree of proximity between the original characteristic data and each type of message; a probability determination subunit, used to determine the probability that the target navigation notice message belongs to each type of message based on the decision value when the target navigation notice message belongs to each type of message and the constant bias item set for each type of message, wherein the constant bias item is used to adjust the method of converting the decision value of each type of message into a probability; a probability distribution determination subunit, used to determine the category probability distribution information of the target navigation notice message belonging to each type of message based on the probability that the target navigation notice message belongs to each type of message.

[0134] Optionally, the classification device for the navigation notice message includes: a data set splicing unit, used to splice the category probability distribution information determined by the first model for the historical navigation notice message and the original feature data of the historical navigation notice message into a training feature data set; a second model training unit, used to input the training feature data set into an initial deep neural network, and update the weight matrices of the hidden layer and the output layer of the initial deep neural network according to the message category predicted by the initial deep neural network based on the training feature data set and the actual category of the historical navigation notice message, until the message category predicted by the initial deep neural network based on the training feature data set and the actual category of the historical navigation notice message are the same, thereby determining to obtain the second model.

[0135] Optionally, the second model includes an input layer, at least two hidden layers, a target network layer and an output layer, wherein the input layer is used to receive a training feature data set, and the dimension of the input layer is consistent with the dimension of the training feature data set; the at least two hidden layers are used to extract data features in the training feature data set step by step and analyze the relationship between different data features; the target network layer is used to randomly control some neurons in the initial deep learning network to be in an inactivated state; and the output layer is used to output the message category predicted by the initial deep neural network based on the training feature data set.

[0136] According to another aspect of an embodiment of the present application, a computer-readable storage medium is further provided, wherein a computer program is stored in the computer-readable storage medium. When the computer program is executed, the device where the computer-readable storage medium is located executes the above-mentioned classification method for navigation notice messages.

[0137] According to another aspect of an embodiment of the present application, an electronic device is also provided, wherein the electronic device 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 one or more processors, the one or more processors execute the above-mentioned classification method of navigation notice messages.

[0138] The serial numbers of the above embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0139] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0140] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of the units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0141] The units described as separate components may or may not be physically separate, and 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 may be selected according to actual needs to achieve the purpose of the present embodiment.

[0142] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0143] If the integrated unit is implemented in the form of 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 application, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc. Various media that can store program code.

[0144] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A classification method for NOTAM messages, characterized in that: include: Extracting original feature data of the target NOTAM message, wherein the original feature data is text feature data after vectorization and dimensionality reduction processing; Inputting original feature data of the target NOTAM message into a first model, and determining, by the first model, category probability distribution information when the target NOTAM message belongs to each category of messages based on the original feature data; Using the category probability distribution information as target feature data of the target NOTAM message; determining a target type of the target NOTAM message based on the original characteristic data and the target characteristic data; wherein determining the target type of the target NOTAM message based on the original feature data and the target feature data includes: concatenating the original feature data and the target feature data into a target feature matrix; inputting the target feature matrix into a second model, analyzing the association relationship and data pattern between different data in the target feature matrix by the second model, and determining the target type of the target NOTAM message based on the analysis result; Among them, the output layer of the first model is connected to the input layer of the second model, the analysis accuracy of the first model for multidimensional linear data is higher than the analysis accuracy of the second model for multidimensional linear data; the analysis accuracy of the second model for the correlation relationship of different feature data is higher than the analysis accuracy of the first model for the correlation relationship of different feature data.

2. The method according to claim 1, characterized in that The training step of the first model includes: Obtain N historical NOTAM messages that have been classified, where N is an integer greater than 1; Converting the N historical NOTAM messages into a vector set, wherein the vector set includes N high-dimensional vectors, the N high-dimensional vectors correspond one-to-one to the N historical NOTAM messages, and each high-dimensional vector is used to represent feature data of multiple dimensions included in a historical NOTAM message; performing a target deletion operation on the vector set to obtain a target vector set, wherein the target deletion operation is used to delete abnormal vectors in the vector set, the abnormal vectors including: vectors converted from historical NOTAM messages having a text length less than a preset length, and vectors converted from historical NOTAM messages including preset characters; The neural network is trained according to the target vector set to obtain the first model.

3. The method according to claim 2, characterized in that Training a neural network according to the target vector set to obtain the first model includes: Calculating a mean vector of the target vector set; Calculating the difference between each vector in the target vector set and the mean vector to obtain a difference vector; Calculating a transposed vector of the difference vector; Determine a covariance matrix of the target vector set according to the difference vector and the transposed vector; The neural network is trained according to the covariance matrix of the target vector set to obtain the first model.

4. The method according to claim 3, characterized in that The neural network is trained according to the covariance matrix of the target vector set to obtain the first model, comprising: Determining a plurality of eigenvalues ​​included in the covariance matrix and an eigenvector corresponding to each eigenvalue, wherein each eigenvalue corresponds to a historical NOTAM message; Selecting an eigenvector corresponding to at least one largest eigenvalue from a plurality of eigenvalues ​​included in the covariance matrix to form a projection vector matrix; Projecting the target vector set into a low-dimensional space according to the projection vector matrix to obtain a low-dimensional feature vector set; The neural network is trained according to the low-dimensional feature vector set to obtain the first model.

5. The method according to claim 4, characterized in that Training the neural network according to the low-dimensional feature vector set to obtain the first model includes: Determining the type of the historical NOTAM message corresponding to each vector in the low-dimensional feature vector set as a training label corresponding to the vector; Splitting the low-dimensional feature vector set into K subsets, where K is an integer greater than 1; Predicting the type of the historical NOTAM message corresponding to each vector in each subset using the neural network, based on the vector in each subset, to obtain a prediction result for each vector in each subset; Determining the prediction accuracy of the neural network for each subset based on the prediction result of each vector in each subset and the training label corresponding to each vector; The neural network is trained according to the prediction accuracy of the neural network for each subset to obtain the first model.

6. The method according to claim 5, characterized in that The neural network is trained according to the prediction accuracy of the neural network for each subset to obtain the first model, including: Calculating the average of the prediction accuracies of the K subsets; According to the average value of the prediction accuracy of the K subsets, the neural network is iterated multiple times until the average value of the prediction accuracy of the neural network for the K subsets is higher than a preset threshold, the iterative operation is terminated, and the neural network after the last iterative operation is used as the first model, wherein the iterative operation is used to adjust the first parameter and the second parameter of the neural network, the first parameter is used to control the degree of error term penalty of the neural network, and the second parameter is used to control the similarity measurement method of the neural network between different data.

7. The method according to claim 1, characterized in that Determining, by the first model and based on the original feature data, category probability distribution information of the target NOTAM message belonging to various categories of messages, includes: Determining, by the first model and based on the original characteristic data, a decision value when the target NOTAM message belongs to each message category, wherein the decision value is used to represent the degree of closeness between the original characteristic data and each message category; determining a probability that the target NOTAM message belongs to each message category based on a decision value when the target NOTAM message belongs to each message category and a constant bias term set for each message category, wherein the constant bias term is used to adjust a method for converting the decision value of each message category into a probability; According to the probability that the target NOTAM message belongs to each message category, category probability distribution information of the target NOTAM message belonging to each message category is determined.

8. The method according to claim 1, characterized in that The training step of the second model includes: splicing the category probability distribution information determined by the first model for the historical NOTAM messages and the original feature data of the historical NOTAM messages into a training feature data set; The training feature data set is input into an initial deep neural network, and according to the message category predicted by the initial deep neural network based on the training feature data set and the actual category of the historical NOTAM messages, the weight matrices of the hidden layer and the output layer of the initial deep neural network are updated until the message category predicted by the initial deep neural network based on the training feature data set is the same as the actual category of the historical NOTAM messages, thereby determining to obtain the second model.

9. The method according to claim 8, characterized in that The second model includes an input layer, at least two hidden layers, a target network layer and an output layer, wherein the input layer is used to receive the training feature data set, and the dimension of the input layer is consistent with the dimension of the training feature data set; the at least two hidden layers are used to extract the data features in the training feature data set step by step and analyze the relationship between different data features; the target network layer is used to randomly control some neurons in the initial deep learning network to be in an inactivated state; and the output layer is used to output the message category predicted by the initial deep neural network based on the training feature data set.

10. A classification device for NOTAM messages, characterized in that: include: A feature extraction unit, configured to extract original feature data of the target NOTAM message, wherein the original feature data is text feature data after vectorization and dimensionality reduction processing; a first processing unit, configured to input original feature data of the target NOTAM message into a first model, and determine, using the first model and based on the original feature data, category probability distribution information of each category of messages to which the target NOTAM message belongs; a second processing unit, configured to use the category probability distribution information as target feature data of the target NOTAM message; a determining unit, configured to determine a target type of the target NOTAM message based on the original characteristic data and the target characteristic data; The determination unit includes: a first processing subunit, configured to combine the original feature data and the target feature data into a target feature matrix; a second processing subunit, configured to input the target feature matrix into a second model, analyze the association relationship and data pattern between different data in the target feature matrix through the second model, and determine the target type of the target NOTAM message based on the analysis results; Among them, the output layer of the first model is connected to the input layer of the second model, the analysis accuracy of the first model for multidimensional linear data is higher than the analysis accuracy of the second model for multidimensional linear data; the analysis accuracy of the second model for the correlation relationship of different feature data is higher than the analysis accuracy of the first model for the correlation relationship of different feature data.

11. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located executes the classification method for NOTAM messages according to any one of claims 1 to 9.

12. An electronic device, characterized in that: The invention comprises one or more processors and a memory, 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 execute the classification method of the navigation notice message according to any one of claims 1 to 9.

Images