Aircraft track description method based on double-flow feature fusion

Abstract

The embodiment of the application discloses a kind of aircraft track description methods based on double-flow feature fusion, it is related to aviation monitoring technical field, method includes: obtaining the track data of aircraft to be identified, and track data is input into preset feature extractor, obtain the track feature vector of aircraft to be identified;Track feature vector is obtained through learnable projection matrix, and word vector is obtained;The DTW alignment distance of track data and different preset template track data is calculated, and the minimum DTW alignment distance is selected to obtain track type embedding;Track feature vector, track type embedding and preset soft prompt data are spliced, and prompt word data is obtained;Prompt word data is encoded, and text vector corresponding to encoding data is determined in preset knowledge base based on encoding data;Text vector and prompt word data are spliced, and fusion data is obtained, and fusion data is input into preset large language model, and track description data is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of aviation monitoring, and in particular to a flight path description method based on double-flow feature fusion. BACKGROUND

[0002] As an important platform for modern air traffic management, flight safety evaluation and transportation efficiency analysis, the key of the aviation monitoring system lies in processing real flight environment data with high dynamics and complexity. In this environment, the flight path data generated by various aircraft not only records the movement trajectory, but also carries the core information reflecting the flight intention, performance characteristics and safety status. Typical flight path data is composed of a series of time and space continuous trajectory points, each trajectory point contains six-dimensional state information: longitude, latitude, height, speed, heading angle and time stamp accurate to milliseconds, and some high-precision data may also contain extended parameters such as attitude angle and acceleration. In order to build an accurate flight monitoring system, the system needs to analyze a large amount of flight path data and synchronously output comprehensive analysis of the data to support efficient operation of aviation management decision-making. Focusing on real-time extraction of key information (aircraft type, flight intention, behavior pattern, safety situation) from original time and space data stream, and effectively transforming these deep analysis results into direct basis for supporting aviation scheduling, safety evaluation, performance research and flight training.

[0003] The kinematic features (such as acceleration change, turning radius) and spatial relationship features (such as formation spacing, relative position change) implied in the flight path data are important basis for aircraft type identification. For example, the typical smooth flight of commercial passenger aircraft and the flexible maneuvering characteristics of small aircraft are significantly different; the hovering trajectory of a helicopter and the cruising flight of a fixed-wing aircraft also present completely different movement patterns. By constructing a complete technical chain of feature extraction-pattern recognition-decision support, specific functions such as aircraft type identification, flight task judgment and safety evaluation can be realized, which can provide key decision-making reference for dispatch personnel in aviation management.

[0004] In the past, data processing and evaluation were mostly carried out by manual analysis, which had the problems of low accuracy and low efficiency. With the development of intelligent technology, artificial intelligence has proposed many effective solutions. The existing method constructs a machine learning model based on deep convolutional neural network, forms a target classifier after sample training and testing, uses deep learning algorithm for aircraft type identification, obtains classification results through intelligent model, determines the specific type of aircraft, and uses large language model to perform semantic description on 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 abnormal data still need to be continuously optimized. SUMMARY

[0005] Therefore, the application provides an aircraft track description method based on double-flow feature fusion, which is used to solve the problem of insufficient generalization ability and robustness of abnormal data in the prior art in a data-scarce scenario.

[0006] In a first aspect, an aircraft track description method based on double-flow feature fusion is provided, and the method comprises the following steps:

[0007] Track data of an aircraft to be identified is obtained, and the track data is input into a preset feature extractor to obtain a track feature vector of the aircraft to be identified;

[0008] The track feature vector is mapped to a word embedding space of a preset large language model through a learnable projection matrix and a preset bias data to obtain a word vector;

[0009] The DTW alignment distance between the track data and different preset template track data is calculated, and the minimum DTW alignment distance is selected to determine the track type corresponding to the track data, thereby obtaining a track type embedding, wherein the preset template track data is track data of a preset track type template;

[0010] The track feature vector, the track type embedding, and preset soft prompt data are spliced to obtain prompt word data;

[0011] 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, wherein the preset encoder is an encoder corresponding to the preset knowledge base;

[0012] The text vector and the prompt word data are spliced to obtain fusion data, and the fusion data is input into the preset large language model to obtain track description data for the track data.

[0013] Optionally, before the step of inputting the track data into the preset feature extractor to obtain the track feature vector of the aircraft to be identified, the method further comprises the following steps:

[0014] A classification network with a hierarchical structure is trained using preset training data to obtain a target classification network;

[0015] A one-dimensional convolution layer and a maximum pooling layer in the target classification network are solidified, and a preset feature extractor is formed based on the solidified one-dimensional convolution layer and the solidified maximum pooling layer.

[0016] Optionally, the step of training the classification network with a hierarchical structure using preset training data to obtain a target classification network comprises the following steps:

[0017] performing length processing on each piece of preset track data in the preset training data according to a preset input sequence length, to obtain an initial training set composed of initial track data with uniform lengths;

[0018] performing standardization processing on attribute values of each piece of initial track data in the initial training set based on a mean value and a standard deviation, to obtain a target training set;

[0019] training a classification network designed in a hierarchical structure using the target training set, to obtain a target classification network.

[0020] Optionally, before the step of mapping the track feature vector to a word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain a word vector, the method further includes:

[0021] under the condition that parameters of the preset large language model and the preset feature extractor remain unchanged, determining parameters in the learnable projection matrix, preset bias data and preset soft prompt data through a cross-entropy loss function according to a gradient descent method, wherein the preset soft prompt data includes a first soft prompt for associating features and task context and a second soft prompt for reinforcing track type embedding.

[0022] Optionally, the step of encoding the prompt word data using a preset encoder to obtain encoded data and determining a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data includes:

[0023] encoding the prompt word data using a preset encoder to obtain encoded data;

[0024] aligning the encoded data with a text vector in a preset knowledge base through a preset projection layer to determine a text vector corresponding to the encoded data, wherein the preset projection layer includes a preset projection matrix and alignment bias data.

[0025] Optionally, before the step of aligning the encoded data with a text vector in a preset knowledge base through a preset projection layer to determine a text vector corresponding to the encoded data, the method further includes:

[0026] under the condition that parameters of the preset encoder remain unchanged, determining the preset projection matrix and the alignment bias data according to a preset positive-negative sample pair through a method of contrast learning by calculating similarities between the encoded data and each sample of the preset positive-negative sample pair.

[0027] In another aspect, the application provides a flight path description system based on double-flow feature fusion, the system comprising:

[0028] The data acquisition module is configured to acquire track data of an aircraft to be identified, and input the track data into a preset feature extractor to obtain a track feature vector of the aircraft to be identified.

[0029] The mapping module is configured to map the track feature vector to a word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain a word vector.

[0030] The alignment module is configured to calculate a DTW alignment distance between the track data and different preset template track data, and select a minimum DTW alignment distance to determine a track type corresponding to the track data to obtain a track type embedding, the preset template track data being track data of a preset track type template.

[0031] The splicing module is configured to splice the track feature vector, the track type embedding and preset soft prompt data to obtain prompt word data.

[0032] The encoding module is configured to encode the prompt word data using a preset encoder to obtain encoded data, and determine a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, the preset encoder being an encoder corresponding to the preset knowledge base.

[0033] The description module is configured to splice the text vector and the prompt word data to obtain fusion data, and input the fusion data into the preset large language model to obtain track description data for the track data.

[0034] In a third aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the aircraft track description method based on double-flow feature fusion when executing the computer program.

[0035] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the aircraft track description method based on double-flow feature fusion.

[0036] The technical scheme of the embodiment of the application comprises the following steps: obtaining track data of an aircraft to be identified, inputting the track data into a preset feature extractor to obtain a track feature vector of the aircraft to be identified, mapping the track feature vector to a word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain a word vector, calculating DTW alignment distances of the track data and different preset template track data, selecting a minimum DTW alignment distance to determine a track type corresponding to the track data, obtaining a track type embedding, the preset template track data being track data of a preset track type template, splicing the track feature vector, the track type embedding, and preset soft prompt data to obtain prompt word data, encoding the prompt word data using a preset encoder to obtain encoded data, determining a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, the preset encoder being an encoder corresponding to the preset knowledge base, splicing the text vector and the prompt word data to obtain fusion data, and inputting the fusion data into the preset large language model to obtain track description data for the track data. The problem of fragmentation caused by the dependence of physical property analysis and behavior pattern recognition on multiple independent models in the traditional scheme is effectively overcome, and a single model is used to simultaneously analyze and textually parse multi-dimensional track information (track physical state and flight mode), which greatly improves the efficiency of situation understanding and the comprehensiveness of decision support. Through contrastive learning, the dynamic prompt word embedding vector is accurately aligned with the text semantic space of the external knowledge base, so that the user can add descriptive knowledge of new models or new behaviors to the knowledge base without retraining a new model, and can realize instant recognition and analysis of new tracks. In addition, by using a unique three-stage training strategy, feature extraction, feature-text conversion alignment, and large language model prompt reasoning components are optimized, and finally only a lightweight combination is required during reasoning. The core model parameters remain frozen, and the full model does not need to be fine-tuned, thereby greatly reducing the computational resource consumption and time cost of model training and updating, enabling efficient deployment on resource-constrained edge devices, and meeting the stringent requirements of timeliness and deployment flexibility in actual aviation scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0038] Among them:

[0039] Figure 1A flowchart of a flight path description method based on double-flow feature fusion according to an embodiment of the present application is shown in FIG. 1.

[0040] Figure 2 An execution architecture diagram of another flight path description method based on double-flow feature fusion according to an embodiment of the present application is shown in FIG. 2.

[0041] Figure 3 A training flowchart of another flight path description method based on double-flow feature fusion according to an embodiment of the present application is shown in FIG. 3.

[0042] Figure 4 A structure diagram of a classification network in another flight path description method based on double-flow feature fusion according to an embodiment of the present application is shown in FIG. 4.

[0043] Figure 5 A structure diagram of a preset feature extractor in another flight path description method based on double-flow feature fusion according to an embodiment of the present application is shown in FIG. 5.

[0044] Figure 6 A flowchart of a contrast learning in another flight path description method based on double-flow feature fusion according to an embodiment of the present application is shown in FIG. 6. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] As shown in FIG. 1, a flight path description method based on double-flow feature fusion according to an embodiment of the present application specifically includes the following steps: Figures 1-2

[0047] S110, obtaining flight path data of a to-be-identified aircraft, and inputting the flight path data into a preset feature extractor to obtain a flight path feature vector of the to-be-identified aircraft;

[0048] In a possible implementation, before the step of inputting the flight path data into the preset feature extractor to obtain the flight path feature vector of the to-be-identified aircraft, the method further includes:

[0049] training the classification network with a hierarchical structure using preset training data to obtain a target classification network;

[0050] solidifying a one-dimensional convolution layer and a maximum pooling layer in the target classification network, and constructing the preset feature extractor based on the solidified one-dimensional convolution layer and the solidified maximum pooling layer.​

[0051] As shown in Figure 4 , the classification network adopts a hierarchical structure design, and is trained according to the process as shown in Figure 3 , one-dimensional convolution is used to capture the local correlation features of the trajectory points in the spatial dimension, maximum pooling is used to reduce the feature dimension, enhance the translation invariance of the features and suppress noise, and finally a fully connected layer is used to integrate the convolution features and output a fixed-dimensional feature vector to obtain the classification result. The convolution and pooling operations for feature extraction are represented as , and the classification process is represented as

[0052]

[0053] wherein, represents a full connection operation, represents a softmax classifier, represents a classification probability.

[0054] As shown in Figure 5 , the feature extraction network directly reuses and solidifies the front-end part of the trained classification network, discards the full connection layer dedicated to classification, and retains the one-dimensional convolution layer and the maximum pooling layer with feature extraction capability to form a feature extraction network , and the feature extraction process is represented as

[0055]

[0056] wherein is the fixed-dimensional, low-dimensional dense track feature vector extracted. The track feature vector effectively encodes the core spatio-temporal and dynamic characteristics of the original track, and provides a robust and generalizable representation input for subsequent track intelligent analysis (such as intent recognition and behavior classification).

[0057] S120, mapping the track feature vector to the word embedding space of the preset large language model through a learnable projection matrix and preset bias data to obtain a word vector;

[0058] As an example, the track feature vector is mapped to the word embedding space of the preset large language model through a learnable projection matrix and preset bias data, specifically:

[0059]

[0060] wherein represents a learnable projection matrix, is a bias term (preset bias data), and the output is to convert the numerical track feature into a word vector form suitable for processing by the large language model, so that it can be compatible with the original vocabulary semantic space of the large model.

[0061] 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.

[0062] 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.

[0063] For a flight path Calculate its minimum DTW alignment distance with each template:

[0064]

[0065] 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.

[0066] S140. The trajectory feature vector, the trajectory type embedding, and the preset soft prompt data are concatenated to obtain prompt word data;

[0067] 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:

[0068]

[0069] Forming final prompt word data As the input of the large language model to generate the track analysis result. By fusing the original track feature numerical representation, the adaptable task context information and the identified semantic label category, the large language model is guided to generate analysis text that is more in line with the task requirements, richer in information and more accurate.

[0070] S150, encode the prompt word data using a preset encoder to obtain encoded data, and determine a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, the preset encoder being an encoder corresponding to the preset knowledge base;

[0071] In one possible implementation, the step of encoding the prompt word data using a preset encoder to obtain encoded data, and determining a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, includes:

[0072] The prompt word data is encoded using a preset encoder to obtain encoded data.

[0073] The encoded data is aligned with the text vector in the preset knowledge base through a preset projection layer to determine the text vector corresponding to the encoded data, the preset projection layer including a preset projection matrix and alignment bias data.

[0074] For example, the encoded data is encoded by the same encoder (preset encoder) as the knowledge base text vector, and this process is represented as:

[0075]

[0076] wherein is the encoded vector. After encoding, the encoded vector of the prompt word embedding is aligned with the knowledge base text vector through a projection layer, and the is mapped to the knowledge base vector space using a preset projection layer:

[0077]

[0078] wherein represents a preset projection matrix, is alignment bias data, and the output is a query vector aligned with the knowledge base vector space and available for efficient retrieval. The retrieval and matching problem in the heterogeneous embedding space is solved.

[0079] S160, splice the text vector and the prompt word data to obtain fusion data, and input the fusion data into the preset large language model to obtain track description data for the track data.

[0080] Exemplarily, a prompt embedding is generated according to input track data , and a retrieval vector is calculated . The top-k most relevant entries in the knowledge base are retrieved using the retrieval vector, the retrieved text is spliced with the prompt embedding, and the spliced result is input into a large model to generate an analysis result

[0081]

[0082] The large language model generates a final analysis result text based on input that integrates original track features, type labels, task context prompts, and external related knowledge. The analysis capability of the large model for complex and unknown track patterns is significantly enhanced, and the accuracy of the use of domain knowledge is improved, thereby improving the zero-shot recognition performance on new category data.

[0083] In one possible implementation, the step of training the classification network with a hierarchical structure design using preset training data to obtain a target classification network includes:

[0084] The length of each piece of preset track data in the preset training data is processed according to a preset input sequence length to obtain an initial training set composed of initial track data with uniform length;

[0085] The attribute values of each piece of initial track data in the initial training set are standardized based on the mean and standard deviation to obtain a target training set;

[0086] The classification network with a hierarchical structure design is trained using the target training set to obtain a target classification network.

[0087] Exemplarily, to adapt to the fixed requirement of the feature extraction network on the input sequence length, each track is truncated to a fixed length. Let the maximum allowed number of track points be , if the number of track points is , the track data is segmented into equal lengths, and each segment shares the category label of the original track; if the number of track points is , zero padding is performed at the end of the sequence to . The track data after truncation and padding is represented as .

[0088] To eliminate the differences in the dimensions of different attributes and accelerate the convergence of the network, the attribute values of each piece of track data are standardized after sequence length processing. The mean and the standard deviation of each attribute are calculated, and the original attribute value under each attribute is standardized as

[0089]

[0090] The preprocessed track data is represented as follows .

[0091] 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:

[0092]

[0093] 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.

[0094] 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:

[0095] 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.

[0096] 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.

[0097] 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:

[0098] Under the condition of keeping the parameters of the preset encoder unchanged, the preset projection matrix and the alignment bias data are determined by calculating the similarity of the encoded data and each sample of the preset positive and negative sample pair according to the contrast learning method.

[0099] For example, first, a positive and negative sample pair needs to be constructed, wherein the positive sample is the knowledge base entry most relevant to the description of the current track, and the negative sample is a random sampling irrelevant entries. As Figure 6 shown, by calculating the similarity of the retrieval vector and the positive / negative sample pair, and freezing the parameters of the text encoder, the parameter distribution of and is optimized by optimizing the contrast loss:

[0100]

[0101] wherein is the contrast loss. represents the cosine similarity. is a temperature coefficient that controls the sharpness of the distribution. This training process drives the query vector to be close to the positive sample and far from the negative sample, thereby learning an effective spatial mapping relationship.

[0102] The physical feature vector of the track is extracted by pre-training the convolutional neural network, and the dynamic time warping algorithm (DTW) is used to identify the specific flight mode type of the track to construct a dynamic prompt word. Combining the characteristics of track physical properties and behavior pattern analysis, this method avoids the disadvantages of deploying multiple independent models in traditional schemes, enabling a single model to output multi-dimensional analysis results simultaneously. Through contrast learning, the prompt word embedding vector is aligned with the knowledge base text vector space, and an external knowledge base is dynamically connected. To address the problem of being unable to predict when a new aircraft model / new track type appears and needing to retrain the model, this method only needs to add new knowledge to the knowledge base, achieving zero-shot retrieval enhancement generation expansion. For different components of the model, separate training is performed, and finally combined for inference, without changing the parameters of the large prophetic model, the method achieves comprehensive analysis of unknown tracks, avoids the resource consumption of traditional full-model fine-tuning, and supports efficient deployment of low-resource edge devices.

[0103] First, the feature extraction network is trained through supervised learning; then the extracted features and track types and soft prompts are used to construct prompt word embedding vectors, the feature mapping layer and soft prompt parameters are jointly optimized, and finally the prompt word embedding vectors and knowledge base text vectors are aligned and retrieved to achieve comprehensive analysis of track data using a large model, providing data support for aviation situation analysis and operation decision-making.

[0104] The application provides a flight path description system based on double-flow feature fusion, the system comprising:

[0105] a data acquisition module, configured to acquire flight path data of a to-be-identified aircraft, and input the flight path data into a preset feature extractor to obtain a flight path feature vector of the to-be-identified aircraft;

[0106] a mapping module, configured to map the flight path feature vector to a word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain a word vector;

[0107] an alignment module, configured to calculate DTW alignment distances of the flight path data and different preset template flight path data, and select a minimum DTW alignment distance to determine a flight path type corresponding to the flight path data to obtain a flight path type embedding, the preset template flight path data being flight path data of a preset flight path type template;

[0108] a splicing module, configured to splice the flight path feature vector, the flight path type embedding and preset soft prompt data to obtain prompt word data;

[0109] an encoding module, configured to encode the prompt word data by using a preset encoder to obtain encoded data, and determine a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, the preset encoder being an encoder corresponding to the preset knowledge base;

[0110] a description module, configured to splice the text vector and the prompt word data to obtain fusion data, and input the fusion data into the preset large language model to obtain flight path description data for the flight path data.

[0111] In a third aspect, the application provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the flight path description method based on double-flow feature fusion as described above when executing the computer program.

[0112] In a fourth aspect, the application provides a computer-readable storage medium, which stores a computer program, wherein the computer program is executable on a processor to implement the steps of the flight path description method based on double-flow feature fusion as described above.

[0113] The above disclosure is only the preferred embodiments of the application, and of course cannot limit the scope of the application, therefore, equivalent changes made according to the claims of the application are still within the scope of the application.

Claims

1. A method for aircraft track description based on dual-flow feature fusion, characterized in that, The method comprises the following steps: acquiring track data of a to-be-identified aircraft, inputting the track data into a preset feature extractor, and obtaining a track feature vector of the to-be-identified aircraft; mapping the track feature vector to a word embedding space of a preset large language model through a learnable projection matrix and preset bias data, and obtaining a word vector; calculating DTW alignment distances of the track data and different preset template track data, and selecting a minimum DTW alignment distance to determine a track type corresponding to the track data, and obtaining a track type embedding, wherein the preset template track data is track data of a preset track type template; concatenating the track feature vector, the track type embedding and preset soft prompt data to obtain prompt word data; encoding the prompt word data using a preset encoder to obtain encoded data, and determining a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data, wherein the preset encoder is an encoder corresponding to the preset knowledge base; concatenating the text vector and the prompt word data to obtain fusion data, and inputting the fusion data into the preset large language model to obtain track description data for the track data.

2. The aircraft track description method based on dual-flow feature fusion according to claim 1, characterized in that, Before the step of inputting the track data into the preset feature extractor to obtain the track feature vector of the to-be-identified aircraft, the method further comprises the following steps: training a classification network with a hierarchical structure using preset training data to obtain a target classification network; solidifying a one-dimensional convolution layer and a maximum pooling layer in the target classification network, and constructing a preset feature extractor based on the solidified one-dimensional convolution layer and the solidified maximum pooling layer.

3. The aircraft track description method based on dual-flow feature fusion according to claim 2, characterized in that, The step of training the classification network with a hierarchical structure using preset training data to obtain a target classification network comprises the following steps: performing length processing on each preset track data in the preset training data according to a preset input sequence length to obtain an initial training set composed of initial track data with uniform length; performing standardization processing on attribute values of each initial track data in the initial training set based on mean and standard deviation to obtain a target training set; training the classification network with a hierarchical structure using the target training set to obtain a target classification network.

4. The aircraft track description method based on dual-flow feature fusion according to claim 1, characterized in that, Before the step of mapping the track feature vector to a word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain a word vector, the method further comprises the following steps: under the condition that parameters of the preset large language model and the preset feature extractor remain unchanged, determining parameters in the learnable projection matrix, preset bias data and preset soft prompt data through a cross-entropy loss function according to a gradient descent method, wherein the preset soft prompt data comprises a first soft prompt for associating features and task context and a second soft prompt for strengthening track type embedding.

5. The aircraft track description method based on dual-flow 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 a text vector corresponding to the encoded data in a preset knowledge base based on the encoded data comprises the following steps: encoding the prompt word data using a preset encoder to obtain encoded data; The encoding data is aligned with a text vector in a preset knowledge base through a preset projection layer to determine a text vector corresponding to the encoding data, the preset projection layer comprising a preset projection matrix and alignment bias data.

6. The aircraft track description method based on dual-flow feature fusion according to claim 5, characterized in that, Before the step of aligning the encoding data with a text vector in a preset knowledge base through a preset projection layer to determine a text vector corresponding to the encoding data, the method further comprises: Under the condition of keeping the parameters of the preset encoder unchanged, the preset projection matrix and the alignment bias data are determined by calculating the similarity of the encoding data and each sample of the preset positive and negative sample pair according to the preset positive and negative sample pair through a contrast learning method.

7. An aircraft track description system based on dual-flow feature fusion, characterized in that, The system comprises: The data acquisition module is configured to acquire track data of an aircraft to be identified and input the track data into a preset feature extractor to obtain a track feature vector of the aircraft to be identified. The mapping module is configured to map the track feature vector to a word embedding space of a preset large language model through a learnable projection matrix and preset bias data to obtain a word vector. The alignment module is configured to calculate a DTW alignment distance between the track data and different preset template track data, and select a minimum DTW alignment distance to determine a track type corresponding to the track data to obtain a track type embedding, the preset template track data being track data of a preset track type template. The concatenation module is configured to concatenate the track feature vector, the track type embedding, and preset soft prompt data to obtain prompt word data. The encoding module is configured to encode the prompt word data using a preset encoder to obtain encoding data, and determine a text vector corresponding to the encoding data in a preset knowledge base based on the encoding data, the preset encoder being an encoder corresponding to the preset knowledge base. The description module is configured to concatenate the text vector and the prompt word data to obtain fusion data, and input the fusion 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, The processor executes the computer program to implement a method for aircraft track description based on double-flow feature fusion according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement a method for aircraft track description based on double-flow feature fusion according to any one of claims 1 to 6.

Images