Aircraft track description method based on double-flow feature fusion

Abstract

The embodiment of the invention discloses an aircraft track description method based on double-flow feature fusion, and relates to the technical field of aviation monitoring, and the method comprises the steps: obtaining the track data of a to-be-recognized aircraft, inputting the track data into a preset feature extractor, and obtaining the track feature vector of the to-be-recognized aircraft; enabling the track feature vector to pass through a learnable projection matrix to obtain a word vector; calculating DTW alignment distances between the track data and different preset template track data, and selecting the minimum DTW alignment distance to obtain track type embedding; splicing the track feature vector, the track type embedding and preset soft prompt data to obtain prompt word data; encoding the prompt word data, and determining a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data; and splicing the text vector and the cue word data to obtain fusion data, and inputting the fusion data into a preset large language model to obtain track description data.

Description

Technical Field

[0001] This invention relates to the field of aviation monitoring technology, and in particular to a method for describing aircraft tracks based on dual-stream feature fusion. Background Technology

[0002] As a crucial platform for modern air traffic management, flight safety assessment, and transportation efficiency analysis, aviation monitoring systems are key to processing highly dynamic and complex real-world flight environment data. In this environment, the flight path data generated by various aircraft not only records their trajectories but also carries core information reflecting flight intentions, performance characteristics, and safety status. Typical flight path data consists of a series of spatiotemporally continuous trajectory points, each containing six-dimensional status information: longitude, latitude, altitude, speed, heading angle, and a timestamp accurate to milliseconds. Some high-precision data may also include extended parameters such as attitude angles and acceleration. To build a precise flight monitoring system, the system needs to analyze massive amounts of flight path data and simultaneously output comprehensive data analysis to support efficient aviation management decision-making. The focus is on extracting key information (aircraft type, flight intentions, behavior patterns, and safety status) in real time from the raw spatiotemporal data stream and effectively transforming these in-depth analysis results into direct evidence supporting aviation scheduling, safety assessment, performance research, and flight training.

[0003] The kinematic features (such as acceleration changes and turning radius) and spatial relationship features (such as formation spacing and relative position changes) implicit in flight path data are important bases for aircraft type identification. For example, the typical stable flight of commercial airliners differs significantly from the agile maneuverability of small aircraft; the hovering trajectory of helicopters also exhibits completely different motion patterns from the cruise flight of fixed-wing aircraft. By constructing a complete technical chain of feature extraction, pattern recognition, and decision support, specific functions such as aircraft type identification, flight mission judgment, and safety assessment can be achieved. This capability can provide key decision-making references for dispatchers in aviation management.

[0004] Traditionally, data processing and evaluation have largely relied on manual analysis, resulting in low accuracy and efficiency. With the development of intelligent technologies, artificial intelligence has offered many effective solutions. Existing methods construct machine learning models based on deep convolutional neural networks, forming target classifiers after sample training and testing. Deep learning algorithms are then used for aircraft type identification, and intelligent models obtain classification results to determine the specific aircraft type. Large language models are then used to semantically describe the input data. These intelligent algorithms provide reliable technical means for real-time flight path analysis, but their generalization ability in data-scarce scenarios and robustness to outlier data still require continuous optimization. Summary of the Invention

[0005] In view of this, the present invention provides an aircraft trajectory description method based on dual-stream feature fusion to solve the problems of insufficient generalization ability and robustness to abnormal data in existing technologies under data-scarce scenarios.

[0006] In a first aspect, embodiments of the present invention provide a method for describing aircraft tracks based on dual-stream feature fusion, the method comprising: The flight path data of the aircraft to be identified is acquired, and the flight path data is input into a preset feature extractor to obtain the flight path feature vector of the aircraft to be identified; The trajectory feature vectors are mapped to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors; Calculate the DTW alignment distance between the track data and different preset template track data, and select the minimum DTW alignment distance to determine the track type corresponding to the track data, thereby obtaining the track type embedding. The preset template track data is the track data of the preset track type template. The trajectory feature vector, the embedded trajectory type, and the preset soft cue data are concatenated to obtain cue word data; The prompt word data is encoded using a preset encoder to obtain encoded data, and a text vector corresponding to the encoded data is determined in a preset knowledge base based on the encoded data. The preset encoder is an encoder corresponding to the preset knowledge base. The text vector and the prompt word data are concatenated to obtain fused data, and the fused data is input into the preset large language model to obtain track description data for the track data.

[0007] Optionally, before the step of inputting the trajectory data into a preset feature extractor to obtain the trajectory feature vector of the aircraft to be identified, the method further includes: The hierarchical classification network is trained using pre-set training data to obtain the target classification network; The one-dimensional convolutional layer and max pooling layer in the target classification network are solidified, and a preset feature extractor is constructed based on the solidified one-dimensional convolutional layer and the solidified max pooling layer.

[0008] Optionally, the step of training the hierarchical classification network using preset training data to obtain the target classification network includes: The length of each preset track data in the preset training data is processed according to the preset input sequence length to obtain an initial training set composed of initial track data of uniform length. The attribute values ​​of each initial track data in the initial training set are standardized based on the mean and standard deviation to obtain the target training set. The hierarchical classification network was trained using the target training set to obtain the target classification network.

[0009] Optionally, before the step of mapping the trajectory feature vector to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors, the method further includes: While keeping the parameters of the preset large language model and the preset feature extractor unchanged, the parameters, preset bias data and preset soft cue data in the learnable projection matrix are determined by gradient descent using the cross-entropy loss function. The preset soft cue data includes a first soft cue for associating features and task context and a second soft cue for enhancing track type embedding.

[0010] Optionally, the step of encoding the prompt word data using a preset encoder to obtain encoded data, and determining the text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, includes: The prompt word data is encoded using a preset encoder to obtain encoded data; The encoded data is aligned with the text vectors in the preset knowledge base through a preset projection layer to determine the text vectors corresponding to the encoded data. The preset projection layer includes a preset projection matrix and alignment bias data.

[0011] Optionally, before the step of aligning the encoded data with text vectors in a preset knowledge base through a preset projection layer to determine the text vector corresponding to the encoded data, the method further includes: While keeping the parameters of the preset encoder unchanged, the preset projection matrix and the alignment bias data are determined by calculating the similarity between the encoded data and each sample of the preset positive and negative sample pairs according to the preset positive and negative sample pairs, using a contrastive learning method.

[0012] On the other hand, this application provides an aircraft trajectory description system based on dual-stream feature fusion, the system comprising: The data acquisition module is used to acquire the trajectory data of the aircraft to be identified and input the trajectory data into a preset feature extractor to obtain the trajectory feature vector of the aircraft to be identified. The mapping module is used to map the trajectory feature vector to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors; The alignment module is used to calculate the DTW alignment distance between the track data and different preset template track data, and select the minimum DTW alignment distance to determine the track type corresponding to the track data, thereby obtaining the track type embedding. The preset template track data is the track data of the preset track type template. The splicing module is used to splice the trajectory feature vector, the trajectory type embedding, and the preset soft prompt data to obtain prompt word data; The encoding module is used to encode the prompt word data using a preset encoder to obtain encoded data, and to determine the text vector corresponding to the encoded data in a preset knowledge base based on the encoded data. The preset encoder is an encoder corresponding to the preset knowledge base. The description module is used to concatenate the text vector and the prompt word data to obtain fused data, and input the fused data into the preset large language model to obtain track description data for the track data.

[0013] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for describing aircraft tracks based on dual-stream feature fusion.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for describing aircraft tracks based on dual-stream feature fusion.

[0015] The technical solution of this invention involves acquiring the trajectory data of an aircraft to be identified and inputting the trajectory data into a preset feature extractor to obtain a trajectory feature vector of the aircraft to be identified; mapping the trajectory feature vector to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain a word vector; calculating the DTW alignment distance between the trajectory data and different preset template trajectory data, and selecting the minimum DTW alignment distance to determine the trajectory type corresponding to the trajectory data to obtain a trajectory type embedding, wherein the preset template trajectory data is the trajectory data of a preset trajectory type template; concatenating the trajectory feature vector, the trajectory type embedding, and preset soft cue data to obtain cue word data; encoding the cue word data using a preset encoder to obtain encoded data, and determining the text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, wherein the preset encoder is the encoder corresponding to the preset knowledge base; concatenating the text vector and the cue word data to obtain fused data, and inputting the fused data into the preset large language model to obtain trajectory description data for the trajectory data. This approach effectively overcomes the fragmentation problem caused by relying on multiple independent models for physical attribute analysis and behavioral pattern recognition in traditional solutions. It achieves synchronous and textual parsing of multi-dimensional track information (track physical state and flight mode) using a single model, significantly improving the efficiency of situational understanding and the comprehensiveness of decision support. By using contrastive learning to precisely align the dynamic cue word embedding vectors with the text semantic space of an external knowledge base, users can instantly identify and parse new tracks without retraining a new model, simply by dynamically adding descriptive knowledge of new aircraft types or behaviors to the knowledge base. Furthermore, by employing a unique three-stage training strategy, it optimizes feature extraction, feature-text conversion alignment, and the large language model cue inference component, ultimately requiring only lightweight integration during inference. This ensures that the core large model parameters remain frozen, eliminating the need for full model fine-tuning, thereby greatly reducing the computational resource consumption and time cost of model training and updates. This allows for efficient deployment on resource-constrained edge devices, meeting the stringent requirements for timeliness and deployment flexibility in real-world aviation scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] in: Figure 1This is a flowchart illustrating an aircraft trajectory description method based on dual-stream feature fusion in an embodiment of the present invention. Figure 2 This is an execution architecture diagram of another aircraft trajectory description method based on dual-stream feature fusion in an embodiment of the present invention; Figure 3 This is a training flowchart for another aircraft trajectory description method based on dual-stream feature fusion in an embodiment of the present invention. Figure 4 This is a structural diagram of the classification network in another aircraft trajectory description method based on dual-stream feature fusion in an embodiment of the present invention; Figure 5 This is a structural diagram of a preset feature extractor in another aircraft trajectory description method based on dual-stream feature fusion in an embodiment of the present invention; Figure 6 This is a schematic diagram of the contrastive learning process in another aircraft trajectory description method based on dual-stream feature fusion in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figures 1-2 As shown in the figure, an aircraft trajectory description method based on dual-stream feature fusion according to an embodiment of the present invention specifically includes the following steps: S110. Obtain the trajectory data of the aircraft to be identified, and input the trajectory data into a preset feature extractor to obtain the trajectory feature vector of the aircraft to be identified; In one possible implementation, before the step of inputting the trajectory data into a preset feature extractor to obtain the trajectory feature vector of the aircraft to be identified, the method further includes: The hierarchical classification network is trained using pre-set training data to obtain the target classification network; The one-dimensional convolutional layer and max pooling layer in the target classification network are solidified, and a preset feature extractor is constructed based on the solidified one-dimensional convolutional layer and the solidified max pooling layer.

[0020] For example, such as Figure 4 As shown, the classification network adopts a hierarchical structure design, and follows the... Figure 3The process shown completes the training. One-dimensional convolution captures the local correlation features of trajectory points in the spatial dimension. Max pooling is used to reduce the feature dimensionality, enhance the translation invariance of features, and suppress noise. Finally, a fully connected layer integrates the convolutional features, outputting a fixed-dimensional feature vector to obtain the classification result. The convolution and pooling operations for feature extraction can be represented as follows: The classification process can be represented as follows:

[0021] in, Indicates a fully connected operation. This represents the softmax classifier. This represents the classification probability.

[0022] like Figure 5 As shown, the feature extraction network directly reuses and solidifies the front-end of the pre-trained classification network, discarding the fully connected layers dedicated to classification and retaining the one-dimensional convolutional layers and max-pooling layers with feature extraction capabilities to form the feature extraction network. The feature extraction process is represented as follows:

[0023] in This is the extracted fixed-dimensional, low-dimensional, dense track feature vector. The track feature vector effectively encodes the core spatiotemporal and dynamic characteristics of the original track, providing robust and generalizable representational input for subsequent intelligent track analysis (such as intent recognition and behavior classification).

[0024] S120. The trajectory feature vector is mapped to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors. For example, a learnable projection matrix and preset bias data are mapped to the word embedding space of a preset large language model, specifically:

[0025] in Represents the learnable projection matrix. For the bias term (preset bias data), output It converts numerical trajectory features into word vectors suitable for processing by large language models, making them compatible with the original lexical semantic space of large models.

[0026] S130. Calculate the DTW alignment distance between the track data and different preset template track data, and select the minimum DTW alignment distance to determine the track type corresponding to the track data, thereby obtaining the track type embedding. The preset template track data is the track data of the preset track type template. For example, based on historical manually interpreted flight track data, predefined... Special flight path type templates (such as "figure-eight flight path", "high-GG turn", etc.), that is, preset template flight path data, for any template They all contain a set of standard trajectory point sequences ,in For template trajectory points, This represents the number of trajectory points in the template.

[0027] For a flight path Calculate its minimum DTW alignment distance with each template:

[0028] in Indicates the flight path With template track The DTW distance is obtained through dynamic programming in the DTW algorithm. Indicates the first One template. Set the matching threshold to... ,when If the condition is met, the track is determined to match the most similar template, and the corresponding track type label text is returned; otherwise, the text "No special track" is returned. Finally, the category text is converted into word embeddings. This enables automatic matching of flight paths with predefined typical flight patterns, providing the model with key high-level semantic label information and enhancing the semantic interpretability of the analysis.

[0029] S140. The trajectory feature vector, the trajectory type embedding, and the preset soft prompt data are concatenated to obtain prompt word data; For example, two types of learnable soft cue vectors (preset soft cue data) are defined, wherein the learnable vectors (First soft cue) is used to associate features with task context, as well as learnable vectors. (Second soft cue) is used to reinforce track type information. Then, the track feature embedding, soft cue, and track type embedding are concatenated to construct the overall cue word data:

[0030] To form the final prompt word data This serves as input to a large language model to generate trajectory analysis results. By fusing numerical representations of original trajectory features, adaptable task context information, and identified semantic label categories, the large language model is guided to generate analytical text that better meets task requirements, is more information-rich, and has higher accuracy.

[0031] S150. The prompt word data is encoded using a preset encoder to obtain encoded data, and the text vector corresponding to the encoded data is determined in a preset knowledge base based on the encoded data. The preset encoder is an encoder corresponding to the preset knowledge base. In one possible implementation, the step of encoding the prompt word data using a preset encoder to obtain encoded data, and determining the text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, includes: The prompt word data is encoded using a preset encoder to obtain encoded data; The encoded data is aligned with the text vectors in the preset knowledge base through a preset projection layer to determine the text vectors corresponding to the encoded data. The preset projection layer includes a preset projection matrix and alignment bias data.

[0032] For example, the process of encoding using the same encoder as the knowledge base text vector (the preset encoder) can be represented as follows:

[0033] in This is the encoded vector. After encoding, it passes through a projection layer to align the encoded vector containing the prompt words with the text vector in the knowledge base. A preset projection layer is then used to... Mapping to the knowledge base vector space:

[0034] in Indicates the preset projection matrix. To align the offset data, output It is a query vector aligned with the knowledge base vector space and can be used for efficient retrieval. It solves the retrieval and matching problem in heterogeneous embedding spaces.

[0035] S160. The text vector and the prompt word data are concatenated to obtain fused data, and the fused data is input into the preset large language model to obtain track description data for the track data.

[0036] For example, based on the input track data, a prompt word embedding is generated. And calculate the retrieval vector. Using retrieval vectors, the top-k most relevant entries are retrieved from the knowledge base, and the retrieved text is... The results are concatenated with the prompt word embeddings and then input into a larger model to generate analysis results.

[0037] The large language model generates the final analysis result text based on input that integrates original track features, type labels, task context cues, and relevant external knowledge. This significantly enhances the large model's ability to analyze complex and unknown track patterns and the accuracy of its domain knowledge application, while also improving zero-shot recognition performance on new data categories.

[0038] In one possible implementation, the step of training a hierarchical classification network using preset training data to obtain a target classification network includes: The length of each preset track data in the preset training data is processed according to the preset input sequence length to obtain an initial training set composed of initial track data of uniform length. The attribute values ​​of each initial track data in the initial training set are standardized based on the mean and standard deviation to obtain the target training set. The hierarchical classification network was trained using the target training set to obtain the target classification network.

[0039] For example, to accommodate the fixed length requirement of the input sequence for the feature extraction network, each track is truncated to a fixed length. Let the maximum allowed number of track points be... If the number of waypoints Then, the track data is divided into segments of equal length, and each segment shares the original track's category label; if the number of track points... Then zero-padding is performed at the end of the sequence until... The truncated and padded track data is represented as follows: .

[0040] To eliminate differences in the units of measurement of different attributes and accelerate network convergence, the attribute values ​​of each track data are standardized after sequence length processing, and the mean of each attribute is calculated. and standard deviation For the original attribute value under each attribute Standardize it as:

[0041] The preprocessed track data is represented as follows .

[0042] For example, the input is a .csv format track data file, where each file corresponds to a single track. Each track consists of several track points arranged in chronological order. Each track point contains multi-dimensional attributes such as longitude, latitude, altitude, heading, speed, and timestamp. The track is represented as follows:

[0043] in, For the first The attribute vectors of each trajectory point. The training data should also include the corresponding labels. It is obtained by parsing the textual description of each known flight path, and it encodes in detail the key prior knowledge such as the spatiotemporal information, maneuverability, target identity, flight intention, behavior pattern, and security situation of the flight path.

[0044] In one possible implementation, before the step of mapping the trajectory feature vector to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors, the method further includes: While keeping the parameters of the preset large language model and the preset feature extractor unchanged, the parameters, preset bias data and preset soft cue data in the learnable projection matrix are determined by gradient descent using the cross-entropy loss function. The preset soft cue data includes a first soft cue for associating features and task context and a second soft cue for enhancing track type embedding.

[0045] For example, the parameters in the learnable projection matrix are trained by freezing the parameters of the large model and the feature extraction network. and preset bias data and soft cue vectors (First soft tip) and (Second soft hint) The difference between the model-generated text and the real labeled text is calculated using the cross-entropy loss function, and the trainable parameters are updated using gradient descent. While retaining the powerful semantic understanding and generation capabilities of the pre-trained large language model, it efficiently adapts it to the specific trajectory analysis task in the aviation field, significantly reducing the model fine-tuning cost, while improving the domain specialization and task adaptability of the output results.

[0046] In one possible implementation, before the step of aligning the encoded data with text vectors in a preset knowledge base through a preset projection layer to determine the text vector corresponding to the encoded data, the method further includes: While keeping the parameters of the preset encoder unchanged, the preset projection matrix and the alignment bias data are determined by calculating the similarity between the encoded data and each sample of the preset positive and negative sample pairs according to the preset positive and negative sample pairs, using a contrastive learning method.

[0047] For example, the first step is to construct positive and negative sample pairs, where the positive samples... The negative sample is the knowledge base entry most relevant to the description of the current trajectory. For random sampling Unrelated entries. For example... Figure 6 As shown, the retrieval vector is calculated. The similarity to positive / negative sample pairs is used, and the parameters of the text encoder are frozen to optimize the contrastive loss for training. and Parameter distribution:

[0048] in To compare the losses. This represents the cosine similarity. The temperature coefficient controls the sharpness of the distribution. This training process drives the query vector to move closer to positive samples and further away from negative samples, thereby learning an effective spatial mapping relationship.

[0049] This method extracts physical feature vectors from flight paths using a pre-trained convolutional neural network and employs Dynamic Time Warping (DTW) to identify specific flight mode types, thereby constructing dynamic cue words. By combining the fragmented nature of flight path physical attribute and behavioral pattern analysis, this approach avoids the drawbacks of traditional methods that require deploying multiple independent models, enabling a single model to simultaneously output multi-dimensional analytical results. Through contrastive learning, the cue word embedding vectors are aligned with the knowledge base text vector space, and external knowledge bases are dynamically linked. Addressing the issue of unpredictable new aircraft models / flight path types requiring model retraining, this method only needs to add new knowledge to the knowledge base, achieving zero-shot retrieval enhancement and expansion. Different components of the model are trained separately and then combined for inference, achieving comprehensive analysis of unknown flight paths without altering the parameters of the large oracle model. This method avoids the resource consumption of traditional full-model fine-tuning and supports efficient deployment on low-resource edge devices.

[0050] First, a feature extraction network is trained through supervised learning. Then, the extracted features, track types, and soft cues are used to construct cue word embedding vectors. The feature mapping layer and soft cue parameters are jointly optimized. Finally, the cue word embedding vectors are aligned with the knowledge base text vectors for retrieval. This enables comprehensive analysis of track data using a large model, providing data support for aviation situation analysis and operational decision-making.

[0051] This application provides an aircraft trajectory description system based on dual-stream feature fusion, the system comprising: The data acquisition module is used to acquire the trajectory data of the aircraft to be identified and input the trajectory data into a preset feature extractor to obtain the trajectory feature vector of the aircraft to be identified. The mapping module is used to map the trajectory feature vector to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors; The alignment module is used to calculate the DTW alignment distance between the track data and different preset template track data, and select the minimum DTW alignment distance to determine the track type corresponding to the track data, thereby obtaining the track type embedding. The preset template track data is the track data of the preset track type template. The splicing module is used to splice the trajectory feature vector, the embedded trajectory type, and preset soft prompt data to obtain prompt word data; The encoding module is used to encode the prompt word data using a preset encoder to obtain encoded data, and to determine the text vector corresponding to the encoded data in a preset knowledge base based on the encoded data. The preset encoder is an encoder corresponding to the preset knowledge base. The description module is used to concatenate the text vector and the prompt word data to obtain fused data, and input the fused data into the preset large language model to obtain track description data for the track data.

[0052] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the aircraft trajectory description method based on dual-stream feature fusion as described above.

[0053] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for describing aircraft tracks based on dual-stream feature fusion.

[0054] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for describing aircraft tracks based on dual-stream feature fusion, characterized in that, include: The flight path data of the aircraft to be identified is acquired, and the flight path data is input into a preset feature extractor to obtain the flight path feature vector of the aircraft to be identified; The trajectory feature vectors are mapped to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors; Calculate the DTW alignment distance between the track data and different preset template track data, and select the minimum DTW alignment distance to determine the track type corresponding to the track data, thereby obtaining the track type embedding. The preset template track data is the track data of the preset track type template. The trajectory feature vector, the embedded trajectory type, and the preset soft cue data are concatenated to obtain cue word data; The prompt word data is encoded using a preset encoder to obtain encoded data, and the text vector corresponding to the encoded data is determined in a preset knowledge base based on the encoded data. The preset encoder is an encoder corresponding to the preset knowledge base. The text vector and the prompt word data are concatenated to obtain fused data, and the fused data is input into the preset large language model to obtain track description data for the track data.

2. The aircraft trajectory description method based on dual-stream feature fusion according to claim 1, characterized in that, Before the step of inputting the trajectory data into a preset feature extractor to obtain the trajectory feature vector of the aircraft to be identified, the method further includes: The hierarchical classification network is trained using pre-set training data to obtain the target classification network; The one-dimensional convolutional layer and max pooling layer in the target classification network are solidified, and a preset feature extractor is constructed based on the solidified one-dimensional convolutional layer and the solidified max pooling layer.

3. The method for describing aircraft trajectory based on dual-stream feature fusion according to claim 2, characterized in that, The step of training a hierarchical classification network using pre-set training data to obtain a target classification network includes: The length of each preset track data in the preset training data is processed according to the preset input sequence length to obtain an initial training set composed of initial track data of uniform length. The attribute values ​​of each initial track data in the initial training set are standardized based on the mean and standard deviation to obtain the target training set. The hierarchical classification network was trained using the target training set to obtain the target classification network.

4. The aircraft trajectory description method based on dual-stream feature fusion according to claim 1, characterized in that, Before the step of mapping the trajectory feature vectors to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors, the method further includes: While keeping the parameters of the preset large language model and the preset feature extractor unchanged, the parameters, preset bias data and preset soft cue data in the learnable projection matrix are determined by gradient descent using the cross-entropy loss function. The preset soft cue data includes a first soft cue for associating features and task context and a second soft cue for enhancing track type embedding.

5. The method for describing aircraft trajectory based on dual-stream feature fusion according to claim 1, characterized in that, The step of encoding the prompt word data using a preset encoder to obtain encoded data, and determining the text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, includes: The prompt word data is encoded using a preset encoder to obtain encoded data; The encoded data is aligned with the text vectors in the preset knowledge base through a preset projection layer to determine the text vectors corresponding to the encoded data. The preset projection layer includes a preset projection matrix and alignment bias data.

6. The aircraft trajectory description method based on dual-stream feature fusion according to claim 5, characterized in that, Before the step of aligning the encoded data with text vectors in a preset knowledge base through a preset projection layer to determine the text vector corresponding to the encoded data, the method further includes: While keeping the parameters of the preset encoder unchanged, the preset projection matrix and the alignment bias data are determined by calculating the similarity between the encoded data and each sample of the preset positive and negative sample pairs according to the preset positive and negative sample pairs, using a contrastive learning method.

7. An aircraft trajectory description system based on dual-stream feature fusion, characterized in that, The system includes: The data acquisition module is used to acquire the trajectory data of the aircraft to be identified and input the trajectory data into a preset feature extractor to obtain the trajectory feature vector of the aircraft to be identified. The mapping module is used to map the trajectory feature vector to the word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain word vectors; The alignment module is used to calculate the DTW alignment distance between the track data and different preset template track data, and select the minimum DTW alignment distance to determine the track type corresponding to the track data, thereby obtaining the track type embedding. The preset template track data is the track data of the preset track type template. The splicing module is used to splice the trajectory feature vector, the trajectory type embedding, and the preset soft prompt data to obtain prompt word data; The encoding module is used to encode the prompt word data using a preset encoder to obtain encoded data, and to determine the text vector corresponding to the encoded data in a preset knowledge base based on the encoded data. The preset encoder is an encoder corresponding to the preset knowledge base. The description module is used to concatenate the text vector and the prompt word data to obtain fused data, and input the fused data into the preset large language model to obtain track description data for the track data.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a method for describing aircraft tracks based on dual-stream feature fusion as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the aircraft trajectory description method based on dual-stream feature fusion as described in any one of claims 1 to 6.

Images