GEO spacecraft maneuver detection method considering sample imbalance

Abstract

The application relates to a GEO spacecraft maneuver detection method considering sample imbalance, which comprises the following steps: forming timing data according to an SGP4 model; screening feature vectors from the timing data based on a maneuver control equation of a GEO satellite; performing PCA dimension reduction processing on the feature vectors, then performing clustering on the dimension-reduced data, sampling and copying according to a preset rule according to a clustering result, and generating a balanced sample training set; constructing a CCBGA deep learning framework, adaptively optimizing network parameters of the CCBGA deep learning framework by adopting a Bayesian optimization algorithm, training the optimized CCBGA framework by using the balanced sample training set; and inputting orbit parameter timing data of a to-be-detected GEO spacecraft into the trained CCBGA framework, and outputting a maneuver detection result. The method can improve the precision and stability of GEO spacecraft maneuver detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of spacecraft orbit monitoring, in particular to a GEO spacecraft maneuver detection method considering sample imbalance. BACKGROUND

[0002] As the core infrastructure in the fields of communication, navigation, meteorology, etc., the stability of the orbit position of a GEO satellite directly determines the in-orbit service capability. Due to the limited orbit resources, the GEO satellite needs to maintain the operating state through frequent orbit maneuvers (including orbit maintenance, temporary patrol, etc.). In the process of in-orbit monitoring, the orbit data of the satellite presents a significant sample imbalance feature: the proportion of samples in the non-maneuver state (majority class) is extremely high, while the proportion of samples in the maneuver state (minority class) is extremely low.

[0003] Traditional spacecraft maneuver detection algorithms are mostly designed based on balanced sample data sets. When facing unbalanced samples, the model is prone to bias towards the majority class, resulting in a significant decrease in the detection accuracy of maneuver points, which cannot meet the actual needs of space situation awareness (SSA). In the prior art, the methods for solving sample imbalance mainly include two categories: oversampling and undersampling. The oversampling method increases the proportion of minority class samples by duplicating them, but it is prone to model overfitting. The undersampling method achieves sample balance by removing majority class samples, but it is prone to losing key feature information. At the same time, traditional detection models mostly use a single neural network structure, which is difficult to fully extract the spatial correlation and temporal dependence features of orbit time series data, further limiting the detection accuracy. SUMMARY

[0004] Therefore, it is necessary to provide a GEO spacecraft maneuver detection method considering sample imbalance, which can improve the accuracy and stability of GEO spacecraft maneuver detection.

[0005] A GEO spacecraft maneuver detection method considering sample imbalance, the method comprising:

[0006] According to the SGP4 model, the catalog TLE data is predicted forward at a preset time interval, and the prediction result is converted into orbit parameters to form time series data;

[0007] Based on the maneuver control equation of the GEO satellite, the feature vectors are selected from the time series data;

[0008] After PCA dimensionality reduction processing of the feature vectors, the reduced data is clustered, and according to the clustering result, sampling and duplication are performed according to a preset rule to generate a balanced sample training set;

[0009] A CCBGA deep learning framework is constructed, a Bayesian optimization algorithm is used to adaptively optimize the network parameters of the CCBGA deep learning framework, and a balanced sample training set is used to train the optimized CCBGA framework.

[0010] The orbit parameter time series data of the GEO spacecraft to be detected are input into the trained CCBGA framework, and a maneuver detection result is output.

[0011] The above GEO spacecraft maneuver detection method considering sample imbalance uses a PCA dimension reduction combined with clustering resampling strategy, compared with traditional oversampling / undersampling methods, which not only retains the local key features of the orbit data, but also avoids data distortion and model overfitting, effectively solving the sample imbalance problem. Then the constructed CCBGA deep learning framework combines double-layer CNN, BiLSTM, GRU and multi-head self-attention mechanism, which can capture local spatial correlation features and long-time dependence features of orbit time series data at the same time, and improve the recognition ability of maneuver features. Finally, the Bayesian optimization algorithm is used to adaptively adjust the model hyperparameters, which optimizes the learning ability and generalization ability of the model, so that the method can maintain stable detection performance in different GEO satellite orbit data sets. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 FIG. 1 is a flowchart of one embodiment of a GEO spacecraft maneuver detection method considering sample imbalance;

[0013] Figure 2 FIG. 2 is a schematic diagram of a CCBGA framework in one embodiment. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0015] In one embodiment, as shown in FIG. 1, a GEO spacecraft maneuver detection method considering sample imbalance is provided, including the following steps: Figure 1

[0016] Step 102, according to the SGP4 model, the catalog TLE data is predicted forward according to the preset time interval, the prediction result is converted into orbit parameters, and time series data is formed.

[0017] ​Cataloged TLE data (two-line orbital element data) is a standard data format for describing satellite orbits, containing key basic information about satellite orbits. The SGP4 model (Simplified General Perturbations No. 4) is an internationally recognized satellite orbit prediction model, applicable to orbit calculations for satellites in low Earth orbit to geostationary orbit. In this step, the preset time interval is preferably 6 hours. Predicting TLE data forward at this interval ensures the smoothness and continuity of the time series data. Orbital parameters include longitude, semi-major axis, orbital eccentricity, orbital plane inclination, right ascension of the ascending node, argument of perigee, and mean perigee angle. These parameters directly reflect the dynamic changes in satellite orbits and are the core analytical data for subsequent maneuver detection. This step transforms discrete TLE data into continuous, standardized time series data, laying the foundation for feature extraction.

[0018] Step 104: Based on the maneuver control equations of the GEO satellite, feature vectors are selected from the time series data; PCA is performed on the feature vectors for dimensionality reduction, and then the dimensionality-reduced data is clustered. Based on the clustering results, sampling and replication are performed according to preset rules to generate a balanced sample training set.

[0019] The maneuvering control equations of the GEO satellite are the core equations describing the changes in orbital parameters during satellite maneuvers. Selecting feature vectors based on these equations ensures that the selected features are directly related to the maneuvering behavior, improving feature effectiveness. Specifically, the feature vectors are five-dimensional time-series data covering key parameters such as the orbital semi-major axis, orbital eccentricity, orbital plane inclination, right ascension of the ascending node, and argument of perigee, comprehensively reflecting the orbital change characteristics during satellite maneuvers. PCA (Principal Component Analysis) dimensionality reduction effectively reduces redundant information in high-dimensional data, lowering data processing costs while preserving core data features. Clustering grouped samples according to feature similarity, avoiding the loss of key features caused by random sampling in traditional resampling. The preset rules are a targeted resampling strategy based on clustering results. Through sampling and replication, the ratio of majority class (non-maneuvering samples) to minority class (maneuvering samples) is balanced, providing a balanced dataset for model training and preventing the model from favoring the majority class.

[0020] Step 106: Construct the CCBGA deep learning framework, use the Bayesian optimization algorithm to adaptively optimize the network parameters of the CCBGA deep learning framework, and use a balanced sample training set to train the optimized CCBGA framework.

[0021] The CCBGA deep learning framework is a hierarchical hybrid framework that integrates CNN (Convolutional Neural Network), BiLSTM (Bidirectional Long Short-Term Memory Network), GRU (Gated Recurrent Unit), and multi-head self-attention mechanism. CNN is used to extract local spatial correlation features of temporal data, BiLSTM can capture the forward and backward temporal dependencies of data, GRU can suppress gradient explosion and optimize information flow, and the multi-head self-attention mechanism can focus on key features. The modules work together to achieve high-precision feature extraction. The Bayesian optimization algorithm iteratively optimizes hyperparameters by constructing a surrogate model. Without extensive grid searches, it can efficiently find the optimal parameter combination, enabling adaptive adjustment of network parameters and improving the model's learning ability and generalization performance. Using a balanced training set to train the optimized framework ensures that the model's ability to recognize both moving and non-moving samples is balanced, avoiding detection bias caused by imbalanced samples.

[0022] Step 108: Input the orbital parameter timing data of the GEO spacecraft to be tested into the trained CCBGA framework and output the maneuver detection results.

[0023] The data to be detected needs to undergo the same preprocessing procedure as the training data, namely, prediction based on the SGP4 model and conversion into unified orbital parameters to ensure that the data format and feature dimensions match, enabling the model to effectively extract features from the data to be detected. The trained CCBGA framework extracts and classifies features from the input data, and outputs detection results indicating whether a maneuver is occurring and the specific type of maneuver (radial, tangential, normal), providing direct data support for space situational awareness, spacecraft collision warning, and other applications.

[0024] The aforementioned methods for detecting GEO spacecraft maneuvers considering imbalanced samples employ a PCA dimensionality reduction combined with clustering resampling strategy. Compared to traditional oversampling / undersampling methods, this approach preserves key local features of the orbital data while avoiding data distortion and model overfitting, effectively addressing the imbalanced sample problem. The constructed CCBGA deep learning framework integrates two-layer CNN, BiLSTM, GRU, and multi-head self-attention mechanisms, simultaneously capturing local spatial correlation features and long-term dependency features of the orbital time-series data, enhancing the recognition capability of maneuver features. Finally, Bayesian optimization is used to adaptively adjust the model's hyperparameters, optimizing its learning and generalization abilities, ensuring stable detection performance across different GEO satellite orbit datasets.

[0025] In one embodiment, the maneuver control equation is:

[0026]

[0027] in, For the semi-major axis of the track, For orbital eccentricity, The inclination angle of the track surface. Right ascension of the ascending node, The perigee argument, For the near-point angle, For true near point angle, For the near-point angle, The orbital angular velocity, For the track semi-drilled diameter, Argument of latitude These are the radial, tangential, and normal velocity maneuvers, respectively. This indicates the amount of change in each orbital element.

[0028] Specifically, this equation can clarify the variation patterns of various orbital parameters (such as semi-major axis, eccentricity, and mean anomalous angle) during satellite maneuvers, providing a clear physical basis for feature vector selection and ensuring that the selected features can accurately reflect the correlation between maneuvering behavior and changes in orbital parameters, thereby improving the recognition accuracy of subsequent detection models.

[0029] In one embodiment, PCA dimensionality reduction is performed on the feature vector, including:

[0030] The eigenvectors are standardized to obtain a standardized data matrix; the covariance matrix of the standardized data matrix is ​​calculated, and the covariance matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalues ​​and corresponding eigenvectors in descending order.

[0031] Select the eigenvectors corresponding to the top k largest eigenvalues, project the standardized data matrix onto the low-dimensional space formed by the eigenvectors, and obtain the dimensionality-reduced principal component matrix.

[0032] Specifically, the purpose of standardization is to eliminate dimensional differences between different orbital parameters, ensuring that all parameters participate in the analysis on the same order of magnitude, and avoiding feature weight bias caused by dimensional influence. The covariance matrix is ​​used to quantify the correlation between parameters. After eigenvalue decomposition, the magnitude of the eigenvalues ​​reflects the importance of the corresponding eigenvectors. The selection of the first... k The eigenvectors corresponding to the largest eigenvalues ​​can reduce high-dimensional data to low-dimensional data while preserving core information, significantly reducing the computational cost of subsequent clustering and model training, while avoiding the risk of model overfitting caused by data redundancy.

[0033] In one embodiment, clustering the dimensionality-reduced data includes:

[0034] Step 1: Randomly select from the dimensionality-reduced data k An initial centroid ;

[0035] Step 2: For each sample in the dimensionality-reduced data Calculate its distance to all centroids and assign it to the cluster corresponding to the nearest centroid. , t Indicates the number of iterations;

[0036] Step 3: Recalculate the centroid of each cluster; repeat steps 2 and 3 until the change in centroid is less than the threshold or the maximum number of iterations is reached.

[0037] Specifically, Euclidean distance is used for distance calculation, which can effectively quantify the feature similarity between samples and centroids. During the iteration process, the centroids are continuously updated and optimized until convergence, ensuring the stability and accuracy of the clustering results. Clustering can divide majority and minority class samples into feature-similar sub-clusters, providing a basis for subsequent targeted resampling and avoiding the loss of key features caused by random operations in traditional resampling.

[0038] In one embodiment, the process of recalculating the centroid of each cluster is as follows:

[0039]

[0040] in, Indicates the first t The cluster corresponding to the nearest centroid in the next iteration Indicates the first t During the nth iteration, the 1st i A single sample in a cluster, Indicates the first t +1 iterations i The centroid of a cluster.

[0041] Specifically, a new centroid is obtained by calculating the mean of all samples within a cluster. This ensures that the centroid accurately represents the core features of the samples within the cluster, making the clustering results more representative. Compared to other centroid calculation methods, the mean centroid effectively reflects the central tendency of samples within the cluster, avoids the influence of extreme values, and ensures that subsequent resampling based on the clustering results can accurately select samples with similar features, further improving the sample equalization effect and providing a high-quality balanced dataset for model training.

[0042] In one embodiment, clustering is used to obtain a four-cluster classification result for the maneuvering and non-maneuvering point data, and sampling weights are assigned to each sub-cluster. :

[0043]

[0044] In the formula, The number of samples in each sub-cluster;

[0045] Furthermore, the number of samples for each sub-cluster is:

[0046]

[0047] in, This represents the number of sample points in the new sample point set.

[0048] The sampling weights are allocated based on the proportion of samples in each sub-cluster, ensuring that each sub-cluster retains its original proportion after resampling, avoiding over-amplification or suppression of features in any one sub-cluster. The ratio of majority to minority class samples after resampling is set according to actual needs. Through this sampling strategy, the resampling process is no longer a random operation, but a targeted selection based on clustering results. This preserves the core features of each sub-cluster, ensuring a balance between the diversity and representativeness of the training set, and providing a guarantee for the model to learn comprehensive maneuvering and non-maneuvering features.

[0049] In one embodiment, the CCBGA deep learning framework comprises a two-layer CNN, a BiLSTM, a GRU, a multi-head self-attention mechanism module, and a fully connected layer connected sequentially; the two-layer CNN is used to extract local features from high-dimensional temporal data, and the operation formula of the convolutional layer is as follows: , express t Feature vectors extracted at each time step, For activation function, The convolution kernel weight matrix is... Given the input sequence, For bias terms, This represents the convolution operator.

[0050] In one embodiment, the implementation process of the multi-head self-attention mechanism module includes:

[0051] The input data is transformed using linear transformations to generate a query matrix, a key matrix, and a value matrix, and the input vector is then assigned to different subspaces.

[0052] The attention score is calculated for each subspace, and the weight coefficients are obtained after normalization by the softmax function. The weight coefficients are used to generate a weighted sum as the output of each head.

[0053] The outputs of all the heads are concatenated and then linearly transformed through a fully connected layer to obtain the final output.

[0054] In one embodiment, an attention score is calculated for each subspace, normalized using a softmax function to obtain weight coefficients, and these weight coefficients are used in a weighted matrix to generate a weighted sum as the output of each head, including:

[0055] Attention scores are calculated for each subspace, and after normalization using the softmax function, weight coefficients are obtained. These weight coefficients are then used to generate a weighted sum, which serves as the output for each head.

[0056]

[0057]

[0058]

[0059] in, Q It is a query matrix. K It is the key matrix and V It is a value matrix. W It is the weight matrix to be learned. Here is the weight matrix for each head. The dimension of the key vector. For input data, superscript T This indicates the transpose operation.

[0060] Specifically, CNNs can learn local features from high-dimensional time-series input data to identify spatial relationships between different features. A typical CNN structure consists of an input layer, convolutional layers, pooling layers, and fully connected layers. The input layer receives the raw time-series data. The convolutional layers capture local features of the input data through sliding window operations. Through multi-channel convolution operations, the model can capture local correlation patterns in multi-dimensional time series. The pooling layers reduce the spatial dimension of the data through downsampling operations, thereby effectively reducing the complexity of subsequent computations. The fully connected layers integrate and map the extracted features to generate feature representations related to the output category; these are usually located at the end of the network. The output layer then outputs the task-related final result data according to the specific task requirements.

[0061] Long Short-Term Memory (LSTM) networks address the vanishing gradient problem of traditional RNNs through a gating mechanism, making them suitable for processing and predicting time-series data. Their neuron structure comprises three gating units, specifically:

[0062] 1) Gate of Oblivion:

[0063] Control the degree of retention of historical information, output

[0064]

[0065] 2) Input Gate:

[0066] Adjust the update intensity of the current information and output. and candidate memory cell states And updated to :

[0067]

[0068] 3) Output gate:

[0069] Calculate the output of the output gate The hidden state of the output is determined. .

[0070]

[0071] In the formula: Represents the Hadamard product; For the weight parameters of each gate; These are the bias vectors for each gate; This is a hidden state.

[0072] In simple terms, LSTM is like an information conveyor belt. The gating mechanism selectively uploads, retains, or discards information, effectively learning long-term dependencies in time-series data. However, motion detection for GEO satellites requires considering the implicit correlation between historical motion characteristics and future motion trends.

[0073] BiLSTM consists of a forward LSTM and a backward LSTM. At each time step t, the output of BiLSTM is determined by both the forward and backward hidden states. Its bidirectional architecture, combined with the maneuver buffer setting, allows for better capture of the forward and backward variation patterns of feature vector temporal data.

[0074] GRU is a simplified architecture of LSTM, its core being the optimization of information flow. It integrates the forget gate and input gate of LSTM into an update gate, and merges the cell state and hidden state. The update gate is a key component of GRU; it calculates update coefficients to determine the degree to which the hidden state information from the previous time step is retained and the degree to which the current input information is accepted, thus dynamically balancing the ratio of old to new information. The reset gate controls the degree to which the hidden state information from the previous time step is reset. It generates reset coefficients based on the current input and the information from the hidden state of the previous time step, determining whether and to what extent the hidden state of the previous time step is reset, allowing the network to better adapt to new input patterns.

[0075] The introduction of multi-head self-attention mechanism aims to reduce the model's attention to irrelevant regions at a given time step, focusing more attention on key regions and alleviating the limitations of traditional LSTM, CNN, and other network models in handling long-term dependencies and complex sequence relationships. The specific operation of this mechanism is as follows: First, the input dataset is input into the model in the form of feature vectors. Then, the sample data is trained through linear transformation, generating three key representation matrices for each sample: a query matrix (…). Q ), key matrix ( K ) and value matrix ( V ), and assign the input vector to different subspaces.

[0076]

[0077] In the formula, represents the input data. W This is the weight matrix to be learned. A similarity function is used to score each subspace and obtain an attention score. Then, a softmax function is used to normalize the attention scores to obtain the weight coefficients for each part. These weights are used in the weighted matrix, generating a weighted sum as the output of each head. .

[0078]

[0079]

[0080] In the formula, This represents the dimension of the key vector, used to scale the dot product to maintain the softmax function. Here are the weight matrices for each head. A fully connected layer concatenates the outputs of each head into a long vector, followed by a final linear transformation to obtain the final multi-head attention output.

[0081]

[0082] Specifically, the Attention function calculates through dot products. Q and K The similarity, after softmax normalization, yields weight coefficients that can accurately allocate attention resources, allowing the model to focus on the abrupt changes in trajectory parameters during maneuvers; the weighting matrix... V Each generated head output corresponds to key features in a different subspace. After being concatenated, they are fused through a fully connected layer to form a more comprehensive and discriminative feature representation, which significantly improves the model's adaptability to complex maneuvering scenarios.

[0083] The dual-layer CNN employs a stacked convolutional structure, capturing local features at different scales through convolutional kernels of varying sizes. The ReLU activation function introduces non-linearity, enabling the network to learn complex feature mappings and avoiding the vanishing gradient problem. The BiLSTM module includes forward and backward LSTMs, simultaneously learning historical dependencies and future trends in time-series data, adapting to the correlation analysis of orbital parameters before and after GEO satellite maneuvers. The GRU module dynamically adjusts information flow through update and reset gates, simplifying the LSTM structure while suppressing gradient explosion and improving training efficiency. A multi-head self-attention mechanism captures key features from different subspaces in parallel, enhancing the model's ability to identify key maneuvering nodes. The output layer of the fully connected layer uses a softmax activation function to classify maneuvering / non-maneuvering and specific maneuvering types. These modules are sequentially connected, forming a complete framework for hierarchical feature extraction and classification, ensuring comprehensive feature coverage from local to global and from temporal to spatial dimensions.

[0084] In one embodiment, a Bayesian optimization algorithm is used to adaptively optimize the network parameters of the CCBGA deep learning framework, including:

[0085] The classification error rate of the CCBGA deep learning framework on the validation set is used as the objective function; the adaptability of the hyperparameter combination is evaluated based on the performance of the CCBGA deep learning framework on the training data.

[0086] The initial learning rate, L2 regularization factor, and the number of network units of BiLSTM and GRU are used as the hyperparameters to be optimized. The objective function value is calculated and the hyperparameter performance is evaluated.

[0087] Based on the evaluated hyperparameter configurations and objective function values, Gaussian process regression is used for iterative optimization.

[0088] Extract the optimal combination of hyperparameters from the Bayesian optimization results, and apply the optimized hyperparameters to network training.

[0089] Specifically, selecting the validation set classification error rate as the objective function directly reflects the model's generalization performance, avoiding optimization bias caused by overfitting the training set. Gaussian process regression, as a surrogate model, efficiently fits the mapping relationship between hyperparameters and the objective function based on the evaluated hyperparameter combinations and objective function values. It then selects the next optimal hyperparameter combination for evaluation using a sampling function, controlling computational costs while ensuring optimization effectiveness. The final extracted optimal hyperparameter combination achieves the best training performance on the balanced training set and the best generalization performance on the validation set, avoiding the subjectivity and inefficiency of manual parameter tuning and enabling adaptive improvement of model performance.

[0090] It should be understood that, although Figure 1The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0091] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting GEO spacecraft maneuvers considering sample imbalance, characterized in that, The method includes: Based on the SGP4 model, the cataloged TLE data is predicted forward at preset time intervals, and the prediction results are converted into orbital parameters to form time series data. Based on the maneuver control equations of GEO satellites, feature vectors are selected from the time-series data; PCA is performed on the feature vectors for dimensionality reduction, and then the dimensionality-reduced data is clustered. Based on the clustering results, sampling and replication are performed according to preset rules to generate a balanced sample training set. A CCBGA deep learning framework is constructed, and the network parameters of the CCBGA deep learning framework are adaptively optimized using the Bayesian optimization algorithm. The optimized CCBGA framework is then trained using the balanced sample training set. Input the orbital parameter timing data of the GEO spacecraft to be tested into the trained CCBGA framework, and output the maneuver detection results. The CCBGA deep learning framework comprises a two-layer CNN, a BiLSTM, a GRU, a multi-head self-attention mechanism module, and a fully connected layer, all connected sequentially. The two-layer CNN is used to extract local features from high-dimensional temporal data, and the convolutional layer operation formula is as follows: , express t Feature vectors extracted at each time step, For activation function, The convolution kernel weight matrix is... Given the input sequence, For bias terms, This represents the convolution operator.

2. The method according to claim 1, characterized in that, The motor control equation is: in, For the semi-major axis of the track, For orbital eccentricity, The inclination angle of the track surface. Right ascension of the ascending node, The perigee argument, For the near-point angle, For true near point angle, For the near-point angle, The orbital angular velocity, For the track semi-drilled diameter, Argument of latitude These are the radial, tangential, and normal velocity maneuvers, respectively. This indicates the amount of change in each orbital element.

3. The method according to claim 1, characterized in that, The feature vector is subjected to PCA dimensionality reduction processing, including: The eigenvectors are standardized to obtain a standardized data matrix; the covariance matrix of the standardized data matrix is ​​calculated, and the covariance matrix is ​​decomposed into eigenvalues ​​to obtain the eigenvalues ​​and corresponding eigenvectors in descending order. Select the eigenvectors corresponding to the top k largest eigenvalues, project the standardized data matrix onto the low-dimensional space formed by the eigenvectors, and obtain the dimensionality-reduced principal component matrix.

4. The method according to claim 1, characterized in that, Clustering of the dimensionality-reduced data includes: Step 1: Randomly select from the dimensionality-reduced data k An initial centroid ; Step 2: For each sample in the dimensionality-reduced data Calculate its distance to all centroids and assign it to the cluster corresponding to the nearest centroid. , t Indicates the number of iterations; Step 3: Recalculate the centroid of each cluster; repeat steps 2 and 3 until the change in centroid is less than the threshold or the maximum number of iterations is reached.

5. The method according to claim 4, characterized in that, The process of recalculating the centroid of each cluster is as follows: in, Indicates the first t The cluster corresponding to the nearest centroid in the next iteration Indicates the first t During the nth iteration, the 1st i A single sample in a cluster Indicates the first t +1 iterations i The centroid of a cluster.

6. The method according to claim 1, characterized in that, The method further includes: Clustering yielded a four-cluster classification of the maneuvering and non-maneuvering point data, and sampling weights were assigned to each sub-cluster. : In the formula, The number of samples in each sub-cluster; Furthermore, the number of samples for each sub-cluster is: in, This represents the number of sample points in the new sample point set.

7. The method according to claim 1, characterized in that, The implementation process of the multi-head self-attention mechanism module includes: The input data is transformed using linear transformations to generate a query matrix, a key matrix, and a value matrix, and the input vector is then assigned to different subspaces. For each subspace, an attention score is calculated, and after normalization by the softmax function, a weight coefficient is obtained. The weight coefficient is then used to generate a weighted sum as the output of each head. The outputs of all the heads are concatenated and then linearly transformed through a fully connected layer to obtain the final output.

8. The method according to claim 7, characterized in that, Attention scores are calculated for each subspace, and weight coefficients are obtained after normalization using the softmax function. These weight coefficients are then used to generate a weighted sum as the output of each head, including: Attention scores are calculated for each subspace, and weight coefficients are obtained after normalization using the softmax function. These weight coefficients are then used to generate a weighted sum, which serves as the output for each head. in, Q It is a query matrix. K It is the key matrix and V It is a value matrix. W It is the weight matrix to be learned. Here is the weight matrix for each head. The dimension of the key vector. For input data, superscript T This indicates the transpose operation.

9. The method according to claim 1, characterized in that, The network parameters of the CCBGA deep learning framework are adaptively optimized using a Bayesian optimization algorithm, including: The classification error rate of the CCBGA deep learning framework on the validation set is used as the objective function; the adaptability of the hyperparameter combination is evaluated based on the performance of the CCBGA deep learning framework on the training data. The initial learning rate, L2 regularization factor, and the number of network units of BiLSTM and GRU are used as the hyperparameters to be optimized. The objective function value is calculated and the hyperparameter performance is evaluated. Based on the evaluated hyperparameter configurations and objective function values, Gaussian process regression is used for iterative optimization. Extract the optimal combination of hyperparameters from the Bayesian optimization results, and apply the optimized hyperparameters to network training.

Images