Method for classifying aviation data based on whale optimization and time sequential hypergraph neural network

Abstract

The application discloses an aviation data classification method based on whale optimization and a time sequence hypergraph neural network, relates to the technical field of aviation field data classification, and comprises the following steps: firstly, standardizing aviation field heterogeneous data, decomposing the aviation field heterogeneous data into two-dimensional time and space data, and constructing the space and time correlation between the aviation field heterogeneous data by searching for adjacent data in the Euclidean space; then, the whale optimization algorithm with global optimization capability is used to optimize the time sequence hypergraph neural network parameters, so as to establish the stable correlation between the aviation field heterogeneous data, and realize accurate classification of the aviation field data.

Description

Technical Field

[0001] This invention belongs to the field of aviation data classification technology, specifically involving an aviation data classification method based on whale optimization and temporal hypergraph neural network. Background Technology

[0002] In recent years, airports have become increasingly automated, with a significant increase in flight volume. Airport data has also become massive and complex, leading to a substantial increase in risks related to flight delays and aircraft health conditions over time. This, in turn, raises civil aviation operating costs, reduces operational safety, and lowers passenger satisfaction. One key approach to addressing these issues is information fusion. Airport operations generate a wealth of observational data. Information fusion can extract this underlying data from multiple platforms, including airlines, airports, and air traffic control, to better optimize functionality in practical engineering areas such as flight delay prediction and aircraft health diagnosis and analysis. Furthermore, it enhances information exchange among multiple platforms and improves the actual travel experience for customers.

[0003] Current efforts in information fusion have demonstrated their immense value in various real-world applications. For example, Google's leadership team integrated Edge-GNN into a RL policy network, using it as the encoder for chip layout derivation; L. Ma et al. used CNN-LSTM to fuse aircraft flight information and weather information for flight trajectory prediction; Zhou F et al. used hierarchical graph fusion of social media messages for user geolocation; and Q. Li et al. used fusion of human skeletal and joint information for motion prediction. Driven by the development of computer chips, deep neural networks have developed rapidly. Graph neural networks, as an emerging deep neural network model inspired by graph theory, have attracted widespread attention for their ability to construct tightly connected structural frameworks for complex, highly correlated, heterogeneous data. Various methods for constructing graph fusion network architectures have been developed in various fields. Some previous graph fusion research started with image data, focusing on spatial features but neglecting the implicit temporal relationships between data points, leading to incomplete information collection. Later research began to focus on the temporal dimension of the data. Zhilong Lu et al. used directed graph networks embedded in LSTM to predict urban traffic flow, while Z.Li et al. used a hierarchical graph network framework to simulate the interaction of heterogeneous traffic participants and combined it with LSTM to predict their trajectories. These techniques combined graph structures with time-related models, but did not consider the impact of the number of connected nodes in the graph structure and the number of LSTM layers on the model balance. When the interconnection relationship between nodes is unclear, the construction of the number of edges in the graph neural network becomes an uncertain factor. Summary of the Invention

[0004] To address the shortcomings of existing technologies that fail to consider the implicit temporal relationships between data and the number of nodes in the graph structure, thus affecting model balance, this invention provides a whale optimization-temporal hypergraph neural network fusion framework. This framework constructs spatial and temporal relationships between heterogeneous data in the aviation field by searching for neighboring data in Euclidean space. It then utilizes the whale optimization algorithm, with its global optimization capabilities, to optimize the parameters of the temporal hypergraph neural network, thereby establishing robust relationships within the heterogeneous data in the aviation field and achieving accurate classification of aviation data. This solves the problem of existing technologies failing to consider the implicit temporal relationships between data and the number of nodes in the graph structure, thus affecting model balance.

[0005] An aeronautical data classification method based on whale optimization and temporal hypergraph neural networks includes the following steps:

[0006] Acquiring heterogeneous data in the aviation field;

[0007] Heterogeneous data is standardized to obtain data with a standard normal distribution;

[0008] The spatial nearest hyperedges are constructed using the KNN algorithm on the data of the standard normal distribution, and the temporal nearest hyperedges are constructed using the TN relationship. The spatial nearest hyperedges and the temporal nearest hyperedges are then concatenated to obtain a set of hyperedges without edge weights.

[0009] Based on the Euclidean distance between data node pairs, the weight of each hyperedge after splicing is calculated, and the hyperedge weights are combined with the hyperedge set to construct an association matrix;

[0010] The degree matrices of nodes and edges are constructed using the incidence matrix, and then the Laplacian operator of the hypergraph is constructed.

[0011] The hypergraph Laplacian operator and data features are input into the HGNN hypergraph neural network, and the optimal THGNN model is output.

[0012] The number of spatial and temporal neighboring hyperedges between data points and the number of hidden layer neurons in the neural network are used as inputs to the Whale Optimization Algorithm (WOA). The output accuracy of each iteration of THGNN is used as the output of the WOA optimization algorithm for continuous optimization to obtain the optimal output parameters.

[0013] The aeronautical data is classified based on the optimal output parameters.

[0014] Furthermore, the standardization process for heterogeneous data involves calculating the mean and variance of the data to transform the data characteristics into data with a standard normal distribution. The heterogeneous data includes data from airlines, airports, weather, and aircraft.

[0015] Furthermore, the weight of each hyperedge is calculated based on the Euclidean distance between data node pairs, and the calculation formula is as follows:

[0016]

[0017] in, yes The internal parameter of represents the weight of the node i to which the hyperedge j connects. It is the average pairwise Euclidean distance between nodes. This represents the difference in Euclidean distance between node i and node j.

[0018] Furthermore, the correlation matrix H is represented as:

[0019]

[0020] Among them, the hyperedge set of spatial and temporal connections between data ε i Represents the set of nodes connected to the i-th spatial hyperedge; ∈ j Let h represent the set of nodes connected to the j-th temporal superedge. n represents the total number of nodes, and h represents the set of nodes connected to the j-th temporal superedge. ij is a parameter in H, representing the value of the weight of the j-th hyperedge inside node i.

[0021] Furthermore, the degree matrix of the node is represented as:

[0022]

[0023] Where * denotes positional multiplication, and diag indicates conversion to a diagonal matrix. This represents the set of node weights, initialized as an identity matrix. This represents the weight of node p. This indicates that the column summation is performed on the k-th row of the correlation matrix H.

[0024] Furthermore, the degree matrix of the edge is represented as:

[0025]

[0026] Here, diag denotes transformation into a diagonal matrix. This indicates that the row summation is performed on the m-th column of the correlation matrix H.

[0027] Furthermore, the hypergraph Laplacian operator is expressed as:

[0028]

[0029] Where T represents transpose. This represents the weight of the hyperedge set, initialized as the identity matrix.

[0030] Furthermore, the number of hidden layer neurons in the neural network is provided by the WOA optimization algorithm.

[0031] Furthermore, the steps of continuously optimizing the WOA optimization algorithm by using the number of spatially and temporally adjacent hyperedges between data points and the number of hidden layer neurons in the neural network as inputs, and using the current optimal output accuracy of the THGNN as the output of the WOA optimization algorithm, include:

[0032] WOA uses the number of spatial connections x, the number of temporal connections y, and the number of hidden layer neurons z in the THGNN as independent variables of the function.

[0033] The optimal accuracy of the THGNN output is used as the function output;

[0034] By adjusting the constraint ranges of x, y, and z, the output precision of each iteration of THGNN is used as the solution of the WOA function.

[0035] The solution is optimized by seeking the best accuracy.

[0036] Furthermore, the WOA has two self-defined fixed parameters, including the search propagation parameter 'a' and the spiral update parameter. By determining the search propagation parameter a and the spiral update parameter The overall parameter search strategy for the WOA is obtained, and its expression is:

[0037]

[0038] Wherein, parameter vector b is the logarithmic spiral constant, and l, r1, r2 are random numbers ranging from -1 to 1. For the current population, the best accuracy denoted as the precision of the current individual, and p is an arbitrary random number uniformly distributed between 0 and 1, used to control the WOA algorithm in selecting the hunting strategy.

[0039] This invention provides an aeronautical data classification method based on whale optimization and temporal hypergraph neural networks, which has the following beneficial effects:

[0040] This invention proposes a whale optimization-temporal hypergraph neural network fusion framework for aerospace data classification. By standardizing heterogeneous aerospace data, it decomposes the data into temporal and spatial dimensions. Spatial and temporal relationships between these heterogeneous data are constructed by searching neighboring data in Euclidean space. The whale optimization algorithm, with its global optimization capabilities, optimizes the parameters of the temporal hypergraph neural network to establish robust relationships within the heterogeneous data, achieving accurate classification of aerospace data. Furthermore, WOA-THGNN is capable of handling various data fusion problems with complex correlations, such as visual recognition, citation classification, flight delays, and aircraft engine parameters. Attached Figure Description

[0041] Figure 1 This is a diagram illustrating the model framework structure of WOA-THGNN in this embodiment of the invention.

[0042] Figure 2 This is a schematic diagram of the construction of time-nearest hyperedge in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the hyperedge generation hypergraph correlation matrix in an embodiment of the present invention;

[0044] Figure 4 This is a flowchart illustrating the WOA optimization process in an embodiment of the present invention;

[0045] Figure 5 This is a flowchart of the WOA-THGNN framework in an embodiment of the present invention;

[0046] Figure 6 This is a bar chart comparing the results of a general dataset in an embodiment of the present invention;

[0047] Figure 7 This is a bar chart comparing the results of the aviation data set in this embodiment of the invention;

[0048] Figure 8 This is a stability analysis curve of the THGNN internal framework on aerospace data in an embodiment of the present invention;

[0049] Figure 9 This is a schematic diagram of the confusion matrix for the classification of takeoff delays in the European Airlines dataset in an embodiment of the present invention;

[0050] Figure 10 This is a schematic diagram of the confusion matrix for arrival delay classification of the European Airlines dataset in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0052] This invention treats data from airlines, airports, weather, and aircraft as a large amount of heterogeneous information. By constructing two internally related topological structures from the spatial and temporal characteristics of this heterogeneous data, it creates multi-source information groups representing the relationships between data points. Essentially, this connects the close relationships between different dimensions of the data. Then, an optimization algorithm is used to optimize the model parameters to maintain a balanced level of connectivity between the data. By unraveling the internal connections between different data sources, the underlying hidden information is analyzed.

[0053] like Figure 1-10 This invention proposes a hypergraph network fusion framework for heterogeneous data with two different dimensions of information, namely time and space. It obtains the desired fusion result by preserving and extracting more realistic and complete information from the data.

[0054] A three-parameter whale optimization algorithm was used to optimize the temporal proximity, spatial proximity, and number of hidden neurons in the graph neural network to improve the generalization performance and model accuracy of the graph network fusion framework.

[0055] The practicality of the WOA-THGNN model framework was demonstrated through comparative experiments using public and domain datasets, and its stability was proven through multiple experiments. Experimental results also show that the fusion framework has good performance.

[0056] The Hypergraph Neural Network (HGNN) architecture was first proposed in 2019 by Yifan Feng, Haoxuan You, and others. Classical graph neural network architectures connect graph edges based on the relationships between nodes, constructing weighted adjacency matrices based on the nodes. However, the strength of these connections is not clearly defined, leading to significant generalization issues and overfitting during propagation. Hypergraph Neural Networks, on the other hand, focus on edge construction, using the nodes connected to each edge to build edge-based weighted adjacency matrices. This results in more robust connections between nodes, allowing for more diverse information transmission from a single node connected to multiple edges.

[0057] The K-Nearest Neighbor (KNN) algorithm, originally proposed by Cover and Hart in 1968, is an algorithm for finding the K nearest neighbors in a dataset. Its core idea is to calculate the Euclidean distance between all pairs of points and then select the K points with the smallest distances to the point to be classified as belonging to the same class. Its classification method is simple, easy to implement, and requires no parameter estimation, ensuring the algorithm's stability and reliability. It also performs well when dealing with spatially structured data points.

[0058] TN (Elapsed-Time Nearest Neighbor) is a method for calculating the number of T adjacent nodes within the nearest elapsed time interval to a given data point. Figure 2 As shown, the core idea is to first calculate the time difference between the point to be measured and any point within the time interval that occurred. Then, from these time differences, select T points that are closest to the time of the point to be measured and whose time occurred before the point to be measured, and classify them as data of the same category. This classification method is used to construct data links in the time dimension, ensuring the logicality and continuity of the data, enabling a better understanding and analysis of the characteristics and evolution trends of time series data, and providing reliability and stability for spatiotemporal graph networks.

[0059] The Whale Optimization Algorithm (WOA) is a novel swarm optimization algorithm proposed in 2016 by Mirjalili et al. from Griffith University, Australia. Inspired by the hunting behavior of humpback whales, it simulates their prey-surrounding, hunting, and prey-searching behaviors to optimize the solution set of a function. The WOA solution process mainly involves three steps: first, limiting the parameter range; second, finding the optimal solution within that range; and third, using the optimal solution to limit the parameter range for a new round. Its advantages include simple operation, minimal parameter adjustments, and a strong ability to escape local optima.

[0060] The WOA-THGNN model faces challenges due to the diverse sources, highly heterogeneous data formats, and significant time-sensitivity of datasets in the aviation field. Conventional convolutional neural network frameworks struggle to fully construct the correlations between heterogeneous features in such data, leading to incomplete feature fusion. Graph neural networks, however, effectively overcome this limitation. Unlike conventional neural networks, they can construct deep correlations within heterogeneous data through edges. Therefore, this invention proposes a graph fusion framework with a hyperparameter optimization algorithm. This framework constructs edges of different dimensions between data nodes and uses optimization algorithms to refine these connections, thereby uncovering deep, implicit information between data nodes. This invention presents a time-series hypergraph neural network structure (WOA-THGNN) based on the whale optimization algorithm. The paper first briefly introduces the overall model architecture, then explains how time series data are fused into the hypergraph neural network, followed by the whale optimization algorithm's parameter optimization portion. A flowchart of the entire model architecture is then provided. Finally, the feasibility of the proposed model framework is verified on a general dataset.

[0061] The proposed WOA-THGNN model framework consists of three main parts: a data feature hyperedge relation constructor, a graph feature fusionist, and an optimizer. Figure 1 As shown, the hyperedge relation builder comprises two builders: one for temporal features and one for spatial features. First, the raw data is preprocessed to transform it into features more suitable for machine processing. These features are then fed into the hyperedge relation builder to generate an association matrix H. After generating the hypergraph Laplacian operator matrix Δ from the association matrix H, it is fused with the data features into a graph feature fusion unit, outputting categorizable labels. The accuracy function, generated by comparing the predicted and actual labels, is then fed into the whale optimization algorithm for multiple parameter optimizations, ultimately obtaining the optimal model parameters.

[0062] THGNN Architecture Based on Spatial and Temporal Series: A Dataset Network Dataset Node Hyperedge sets that provide spatial and temporal connectivity between data {∈1,∈2,...,∈ n Each node corresponding to a hyperedge has a corresponding weight set. Each hyperedge has a corresponding weight set. Each node has a corresponding weight set. The neural network internally contains learnable parameter settings Θ. The dataset network g can be represented by an association matrix. The specific definitions of the parameters within its correlation matrix H are as follows:

[0063]

[0064] For all nodes Its node degree matrix is ​​defined as a diagonal matrix. For all edges Its degree matrix is ​​defined as a diagonal matrix. The overall hypergraph Laplace operator is represented as: Based on the hypergraph operators above, the update representation of the model can be constructed. The ReLU function is used as the activation function in the neural network, and the Dropout function is added to prevent overfitting, resulting in a two-layer neural network.

[0065] Graph neural networks that utilize the temporal dimension of data for information diffusion and propagation can enhance the richness of node information while reducing the overgeneralization of node information. Since spatially correlated data possesses strong connectivity and information diffusion, the temporal characteristics of the data are crucial for limiting data diffusion and enriching information. For spatial connections, this invention employs the KNN algorithm, using the Euclidean distance between node data features to cluster the x nearest nodes to generate a hyperedge cluster with the same number of nodes for edge construction. For temporal connections, the y nearest (TN) nodes before and after the data are used to construct edges. In data without explicit temporal correlation, the construction of temporally correlated hyperedges becomes a way to strengthen connectivity between data of the same category, while also further diffusing this relationship to data with extremely similar feature structures but different classifications. Considering the weighting of hyperedges, since the greater the spatial distance, the smaller the amount of information propagation, a negative correlation function is needed to assign weights to the hyperedge set. The weight matrix for the node connection hyperedge set in both time and space is specifically divided into two parts. The specific representation of the node connection hyperedge set is as follows:

[0066]

[0067] in, It is the average pairwise Euclidean distance between nodes. After concatenating the temporal and spatial edge relationships to form a general node-edge relationship, this is fused with the weight matrix of the edges connecting the nodes to generate the association matrix H of the dataset network, such as... Figure 3 As shown. Its specific representation is as follows:

[0068]

[0069] The degree matrix of nodes and edges is specifically represented as follows:

[0070]

[0071]

[0072] Therefore, the hypergraph Laplace operator can be specifically expressed as:

[0073]

[0074] The Laplacian operator in a hypergraph can be used to update and iterate data features, and the properties of graph neural networks can be used to propagate these features. By enriching the original data features with feature hyperedges in the spatial dimension of the data, while using feature hyperedges in the temporal dimension of the data to maintain the stability and diffusion constraints of the existing data, a balance between the quality and quantity of feature propagation between data points can be achieved.

[0075] THGNN Parameter Combination Optimization Based on WOA: The number of connected nodes in an edge affects the information propagation strength of the entire graph neural network. More connections mean wider information propagation, but excessive connections can lead to network assimilation, causing each node to lose its original features. Simultaneously, the number of neurons in the hidden layer also significantly impacts the fusion effect. Increasing the number of neurons strengthens the network's robustness, but too many neurons can also cause significant learning bias, often resulting in the learning of only the most prevalent category. Therefore, this invention utilizes the WOA optimization algorithm to optimize the number of spatial connections (x), the number of temporal connections (y), and the number of hidden layer neurons (z) to ensure the model is unaffected by network assimilation and learning bias, thereby obtaining the optimal balance parameters for the model.

[0076] The WOA function takes the number of spatial connections x, the number of temporal connections y, and the number of hidden layer neurons z in the THGNN as independent variables and the optimal accuracy of the THGNN output as the function output. By adjusting the constraint range of x, y, and z, the output accuracy of the THGNN in each iteration is taken as the solution of the WOA function. The solution is optimized by seeking the optimal accuracy.

[0077] In each optimization process, WOA randomly returns corresponding hypergraph parameters to THGNN within a given range to generate multiple populations containing multiple individuals. Within each population, the best individual is selected through comparison. Then, through multiple iterations, convergence operations are performed on the best and temporary individuals within each population to generate new best individuals. Finally, the optimal individual from multiple populations is selected as the final output.

[0078] WOA itself has two self-defined fixed parameters: the search propagation parameter 'a' and the spiral update parameter 'l'. These parameters reveal the overall parameter search strategy of WOA, specifically expressed as follows:

[0079]

[0080] Where the parameter vector b is the logarithmic spiral constant, and l, r1, r2 are random numbers ranging from -1 to 1. For the current population, the best accuracy denoted as the precision of the current individual, and p is an arbitrary random number uniformly distributed between 0 and 1, used to control the WOA algorithm in selecting the hunting strategy.

[0081] The specific process of the WOA-THGNN model:

[0082] Step 1 first standardizes the data by calculating the mean and variance of the data, transforming the data features into standard normal distribution data, and then initializes the network parameters.

[0083] Step 2 involves constructing spatial and temporal proximity hyperedges between the processed data features using the KNN algorithm and TN relationship. The number of neighboring nodes in the KNN algorithm and TN relationship is determined by the WOA optimization algorithm, thus constructing the association matrix H in THGNN. The degree matrix of the corresponding node can then be constructed using the association matrix H. Degree matrix of edges The hypergraph Laplacian operator Δ is further constructed.

[0084] Step 3 involves feeding the hypergraph Laplacian operator Δ and data features into the neural network for training, updating and propagating the data features, and finally outputting the current optimal THGNN model. This model is then validated on a validation set, and the optimal model and accuracy are saved. The number of hidden layer neurons in the neural network is determined by the WOA optimization algorithm.

[0085] Step 4 takes the KNN algorithm, the number of neighboring points of the TN relationship, and the number of hidden layer neurons of the neural network as function inputs, and the current best output accuracy of THGNN as function outputs and puts them into the WOA optimization algorithm for optimization. After optimization, the new number of neighboring points of the KNN algorithm, the number of neighboring points of the TN relationship, the number of hidden layer neurons, and the current best accuracy of the population in WOA are obtained. Steps 2 and 3 are repeated.

[0086] Step 5 involves multiple rounds of training and optimization using WOA to finally obtain the model and parameters with the best fusion effect.

[0087] Numerical simulation validation: The effectiveness of the model framework was evaluated and compared using three public datasets, confirming its validity. The baseline models compared included decision trees, random forests, graph attention mechanisms (GAT), special graph attention mechanisms (SpGAT), graph convolutional neural networks (GCN), and hypergraph neural networks (HGNN). In these datasets, 80% of the data was used for training, and the remaining 20% ​​was used for testing.

[0088] Data Sources: The datasets include point cloud image features from the ModelNet40 and NTU2012 datasets for visual object recognition, and citation network classification from the Cora dataset. The ModelNet40 dataset, as a 3D point cloud image feature dataset, consists of 12,311 objects from 40 popular categories. Its introductory training / test split was applied, with 9,843 objects used for training and 2,468 objects used for testing. The NTU dataset consists of 2,012 3D shapes from 67 categories. Group View Convolutional Neural Network (GVCNN) was used for shape representation due to its good performance in representing 3D objects. The Cora dataset was used for citation data classification, containing 2,708 data points, 5,429 edges, and 1,433 features.

[0089] Experimental Environment and Parameter Settings: This experiment used a two-layer THGNN with a dropout of 0.5 to avoid overfitting. ReLU was used as the non-linear activation function. The Adam optimizer was used during training to minimize the cross-entropy loss function, with a learning rate of 0.001. Each iteration of the THGNN framework ran for 300 epochs. The WOA optimization algorithm used 20 populations and 3 training epochs to optimize the internal structure of the THGNN, with a value of 2 for a, a value of 0.5 for b, and initial constraints for x, y, and z set to 1–20, 1–20, and 100–200, respectively. The comparison algorithm used an empirically determined number of 128 hidden layer neurons, with 1 attention head, while keeping other parameters consistent.

[0090] Experimental Results and Analysis: The comparative experimental results on a general dataset are shown in Table 1 and... Figure 6 As shown.

[0091] Table 1. Comparison of model accuracy (percentage) on a general dataset.

[0092]

[0093] Within the same round, this invention used average classification accuracy as the evaluation metric. The results are shown in Table 1. Compared to some state-of-the-art methods in basic machine learning and deep learning, the method employed in this invention achieves an accuracy improvement of 11.8% to 14.7% on three general datasets. Figure 6 As can be seen, on general datasets, compared to several state-of-the-art methods, the WOA-THGNN model achieves the best performance. Furthermore, continuous experiments have shown that THGNN remains stable within approximately 550 training iterations, while WOA can find highly accurate parameters in three optimization cycles across 20 populations. If, considering time factors, the training process ends after exceeding a certain appropriate accuracy, suitable constraints can be added to the WOA optimization process to reduce time complexity.

[0094] Example Analysis and Validation: In this section, experiments were conducted using an aviation-related dataset within the proposed WOA-THGNN model framework for comparison. The aviation dataset employed QAR-based aircraft engine parameters to classify aircraft blade vibrations and to predict takeoff and landing delays for European flights. Furthermore, this invention preprocessed all datasets, calculating the mean and variance of the original data to transform the data features into a standard normal distribution.

[0095] Experimental environment and parameter settings for aviation domain datasets: Three aviation domain datasets were used, including the QAR engine parameter dataset, the European aviation takeoff delay dataset, and the European aviation landing delay dataset. Missing data was removed. Due to the extreme imbalance of the aviation delay dataset, the SMOTE algorithm was used to balance the aviation delay data. The QAR engine parameter dataset contains 10,372 data points and 25 features, including engine exhaust temperature, low-pressure turbine vibration coefficient, and engine speed, using the vibration coefficient of the left engine blade as the label. The European aviation takeoff delay dataset contains 15,729 data points and 49 features, including estimated departure date, estimated arrival date, same departure airport (LIS airport), preceding flight departure date, arrival airport (using label_encoding), weather, and aircraft type (using one-hot encoding). The European aviation landing delay dataset contains 16,584 data points and 55 features, including estimated departure date, estimated arrival date, same arrival airport (LIS airport), actual departure date, preceding flight arrival date, departure airport, weather, and aircraft type. The labels are based on delay time levels, and the standard used references the flight delay level classification in the "Regulations on Flight Regularity Management," as shown in Table 2. In the above datasets, 80% of the data in the QAR dataset and the flight departure delay dataset was used for training, and 20% for testing. In the flight landing delay dataset, 75% of the data was used for training, and 25% for testing. This experiment used the same experimental environment and parameter configuration as the numerical verification simulation experiment.

[0096] Table 2 Delay Level Determination

[0097]

[0098] Experimental Results and Comparison: The framework proposed in this invention was compared with some existing frameworks, and the experimental results proved that the theory of the proposed framework is valid.

[0099] Experimental results:

[0100] Table 3 Classification results of aviation-related datasets

[0101]

[0102]

[0103] Table 3 shows the experimental and comparative results on aviation-related datasets. As can be seen from the table, the WOA-THGNN framework significantly outperforms several other baseline models in terms of accuracy on flight delay datasets, achieving improvements of 21.3% and 17.5%, respectively. Simultaneously, it shows a 3.2% improvement over the baseline deep neural network model in QAR-based aircraft engine parameters. The experimental results demonstrate the effectiveness of the method on time-dependent domain datasets, proving the theoretical validity of the proposed framework.

[0104] Comparative analysis: through Figure 7 As can be seen, although the accuracy of the WOA-THGNN model on the QAR dataset is not as high as that of decision trees in machine learning, it is higher than the other frameworks overall. The reason for this is that the QAR dataset has a small amount of data and fewer features, while the linear relationship of the features is relatively clear. Traditional machine learning is better than deep neural networks when learning features of a small number of samples and low-dimensional linear data. This is because deep neural networks do not learn features of a small number of data incompletely.

[0105] Iterative experiments using the THGNN internal framework on aerospace datasets revealed that the overall validation accuracy stabilized after more than 550 iterations. Therefore, to ensure experimental accuracy, the number of iterations was set to 600. These internal iterative experiments demonstrate the excellent model stability of the THGNN framework across different datasets and provide reliable experimental support for selecting appropriate iteration counts to reduce the time required for optimization.

[0106] An experimental visualization comparison of the European air delay dataset was conducted using a confusion matrix, such as... Figure 9 Confusion matrix for takeoff delay classification in the European Airlines dataset (a: unnormalized HGNN, b: normalized HGNN, c: normalized WOA-THGNN) and Figure 10 The confusion matrix for arrival delay classification on the European Airlines dataset is shown below (d: unstandardized HGNN, e: standardized HGNN, f: standardized WOA-THGNN). The horizontal axis represents the predicted classification label, and the vertical axis represents the actual classification label. Compared to the HGNN method without data preprocessing and the WOA-THGNN method after data standardization, it can be seen that the WOA-THGNN model can classify flight delay data with higher accuracy.

[0107] Through numerical simulation verification and experimental comparison with aerospace data, it can be seen that the model framework proposed in this invention has good classification performance for different types of data, and also has better generalization performance compared with the basic HGNN framework. When dealing with data and features of relatively high magnitude, the model framework proposed in this invention can uncover deeper information compared to some newer deep network baseline models and machine learning models. Numerical simulation experiments revealed that although some similar data have large Euclidean distances, there are implicit data relationships within the same classification data. Spatial relationships constructed solely through Euclidean space cannot effectively uncover these implicit relationships; however, random relationships constructed using data of the same classification effectively extract and construct these implicit relationships. Furthermore, since the HGNN model only constructs spatial dimensions of relationships, the comparison with the HGNN model demonstrates the significant importance of the temporal dimension in uncovering deeper information in the data.

[0108] This invention proposes a fusion framework based on the Whale Optimization Algorithm and a Time Series Hypergraph Neural Network (WOA-THGNN). This framework extends the time series fusion process and the optimization process of the network's internal hyperparameters on the basis of the original Hypergraph Neural Network (HGNN). Compared with traditional graph networks and machine learning, WOA-THGNN is a more general and high-precision fusion framework that can solve complex and closely related data feature problems, while being less prone to getting trapped in local optima during the fusion process. Experiments were conducted to evaluate the model using WOA-THGNN on general visual classification and citation network classification datasets, as well as on aviation-related datasets such as flight delay classification and aircraft blade vibration classification. Compared with other listed models, the WOA-THGNN model showed significant accuracy improvements in several experiments. WOA-THGNN is capable of handling various data fusion processing problems with complex correlations, such as visual recognition, citation classification, flight delays, and aircraft engine parameters.

[0109] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for classifying aeronautical data based on whale optimization and temporal hypergraph neural networks, characterized in that, Includes the following steps: Acquiring heterogeneous data in the aviation field; Heterogeneous data is standardized to obtain data with a standard normal distribution; The spatial nearest hyperedges are constructed using the KNN algorithm on the data of the standard normal distribution, and the temporal nearest hyperedges are constructed using the TN relationship. The spatial nearest hyperedges and the temporal nearest hyperedges are then concatenated to obtain a set of hyperedges without edge weights. Based on the Euclidean distance between data node pairs, the weight of each hyperedge after splicing is calculated, and the hyperedge weights are combined with the hyperedge set to construct an association matrix; The degree matrices of nodes and edges are constructed using the incidence matrix, and then the Laplacian operator of the hypergraph is constructed. The hypergraph Laplacian operator and data features are input into the HGNN hypergraph neural network, and the optimal THGNN model is output. The number of spatial and temporal neighboring hyperedges between data points and the number of hidden layer neurons in the neural network are used as inputs to the Whale Optimization Algorithm (WOA). The output accuracy of each iteration of THGNN is used as the output of the WOA optimization algorithm for continuous optimization to obtain the optimal output parameters. The aeronautical data is classified based on the optimal output parameters.

2. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 1, characterized in that, The standardization process for heterogeneous data involves calculating the mean and variance of the data to transform the data characteristics into data with a standard normal distribution. The heterogeneous data includes data from airlines, airports, weather, and aircraft.

3. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 1, characterized in that, The weight of each hyperedge is calculated based on the Euclidean distance between data node pairs, and the calculation formula is as follows: in, , yes The internal parameter of represents the weight of the node i to which the hyperedge j connects. It is the average pairwise Euclidean distance between nodes. This represents the difference in Euclidean distance between node i and node j.

4. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 1, characterized in that, The correlation matrix H is represented as follows: Among them, the hyperedge set of spatial and temporal connections between data , Represents the set of nodes connected to the i-th spatial hyperedge; , Let n represent the set of nodes connected to the j-th temporal superedge; n represents the total number of nodes. is a parameter in H, representing the value of the weight of the j-th hyperedge inside node i.

5. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 1, characterized in that, The degree matrix of the node is represented as: in, This indicates positional multiplication, and `diag` indicates conversion to a diagonal matrix. This represents the set of node weights, initialized as an identity matrix. This represents the weight of node p. This indicates that the column summation is performed on the k-th row of the correlation matrix H.

6. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 4, characterized in that, The degree matrix of the edge is represented as: Here, diag denotes transformation into a diagonal matrix. This indicates that the row summation is performed on the m-th column of the correlation matrix H.

7. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 4, characterized in that, The hypergraph Laplacian operator is represented as follows: Where T represents transpose. This represents the weight of the hyperedge set, initialized as the identity matrix.

8. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 1, characterized in that, The number of hidden layer neurons in the neural network is provided by the WOA optimization algorithm.

9. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 8, characterized in that, The method involves continuously optimizing the WOA optimization algorithm by using the number of spatially and temporally adjacent hyperedges between data points and the number of hidden layer neurons in the neural network as inputs, and using the current optimal output accuracy of the THGNN as the output of the WOA optimization algorithm. The steps include: WOA counts the number of edges connecting nodes in the THGNN node space. Number of time-connected edges and the number of neurons in the hidden layer of a neural network As the independent variable of a function; The optimal accuracy of the THGNN output is used as the function output; By adjusting The constraint range, the output accuracy of THGNN at each iteration is used as the solution of the WOA function; The solution is optimized by seeking the best accuracy.

10. The aeronautical data classification method based on whale optimization and temporal hypergraph neural network according to claim 8, characterized in that, The WOA has two self-defined fixed parameters, including the search propagation parameter. and spiral update parameters By determining the search propagation parameters and spiral update parameters The overall parameter search strategy for the WOA is obtained, and its expression is: Wherein, parameter vector , , It is the logarithmic spiral constant. , It is a random number in the range of -1 to 1. For the current population, the best accuracy For the current individual's precision, It is an arbitrary random number uniformly distributed between 0 and 1, used to control the WOA algorithm to select the hunting strategy.

Images