Track prediction method and device based on pre-trained large language model

Abstract

The invention provides a track prediction method and device based on a pre-trained large language model, and relates to the field of artificial intelligence. The method comprises the following steps: acquiring historical flight path data of a target aircraft and carrying out data preprocessing to obtain a flight path sequence; constructing a cue word, and inputting the cue word into a pre-trained large language model to obtain a word embedding representation; the cue word comprises track prediction task description and track data description; encoding the track sequence to obtain an encoded track sequence and a track position sequence, and performing word segmentation on the track position sequence; constructing a space-time diagram of the track sequence based on the encoded track sequence, and extracting space-time dependent feature representation of the track sequence; fusing the word embedding representation and the space-time dependent feature representation by using a pre-trained large language model to generate a track information feature representation; and inputting the track information feature representation into a trained prediction module to obtain predicted values of the longitude, the latitude and the height of the target aircraft at the future moment.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a track prediction method and device based on a pre-trained large language model. Background Art

[0002] The operation mode based on trajectory (TBO) can effectively reduce the uncertainty of flight trajectory, improve the utilization rate of airspace resources, and ensure the safety of aviation operations by accurately predicting the flight time and flight trajectory of the aircraft. Trajectory prediction is to predict the three-dimensional position (longitude, latitude, altitude) of the aircraft in the future based on the historical three-dimensional position, speed and other information of the aircraft to assist air traffic control.

[0003] At present, there are three main research methods for track prediction by scholars at home and abroad:

[0004] (1) Methods based on aerodynamic or kinematic models. This type of method builds a prediction model based on kinematic assumptions and requires many parameters such as atmospheric conditions, pilot actions, and flight intentions. Kinematic models are mostly implemented under some ideal assumptions and rarely consider the impact of actual constraints and human behavior. In addition, these methods require a large number of parameters, some of which are sensitive and difficult to obtain. Once the data source is limited, the prediction accuracy of the model will be greatly reduced or even unusable.

[0005] (2) Methods based on parameter-free estimation, such as various improved Kalman filter algorithms. These algorithms require a lot of modeling and analysis work, and their internal structure is relatively simple. They cannot fully explore the hidden dependencies in historical data, and the model is not robust enough.

[0006] (3) Methods based on deep learning. With the rapid development of deep learning, researchers have proposed track prediction models based on LSTM, bidirectional LSTM, CNN, Transformer and other models to mine the changing patterns of aircraft tracks from a large amount of historical track data. This type of method often models the track prediction problem as a multivariate time series problem, ignoring the learning modeling of the semantic correlation between track positions. The model prediction accuracy still needs to be improved. Summary of the invention

[0007] In order to introduce learning modeling of the semantic correlation between aircraft trajectory positions and further improve the accuracy of track prediction, the present invention proposes a track prediction method and device based on a pre-trained large language model, which utilizes the text understanding and reasoning capabilities of the existing pre-trained large language model to better model the semantic correlation of track point positions and assist track prediction.

[0008] In a first aspect, the present invention provides a track prediction method based on a pre-trained large language model, comprising:

[0009] Step 1: Acquire the historical track data of the target aircraft and perform data preprocessing to obtain a track sequence; wherein the information of each track point in the track sequence includes longitude, latitude, altitude and speed;

[0010] Step 2: Construct prompt words and input the prompt words into the pre-trained large language model to obtain word embedding representation; wherein the prompt words include the track prediction task description and the track data description;

[0011] Step 3: Generate geocoding for the longitude, latitude, and altitude of each track point in the track sequence to obtain the corresponding track position sequence, combine the speed as the attribute of each track point with the track position sequence as the encoded track sequence, and perform word segmentation on the track position sequence;

[0012] Step 4: constructing a spatiotemporal graph of the track sequence based on the encoded track sequence, and extracting a spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence;

[0013] Step 5: Using a pre-trained large language model to fuse the word embedding representation and the spatiotemporal dependency feature representation to generate a track information feature representation;

[0014] Step 6: Input the track information feature representation into the trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at a future time.

[0015] Furthermore, in step 3, a geocode is generated for the longitude, latitude, and altitude of each track point in the track sequence, specifically including:

[0016] Use the geohash algorithm to encode the longitude and latitude respectively and generate their corresponding binary codes;

[0017] Set a range of altitude values, and divide the altitude into two parts according to the midpoint of the range until a binary code with satisfactory accuracy is obtained.

[0018] Furthermore, in step 3, a BPE word segmenter is used to segment the track position sequence.

[0019] Furthermore, in step 4, a spatiotemporal graph of the track sequence is constructed based on the encoded track sequence, specifically including:

[0020] Slice the track sequence according to the preset time window;

[0021] For each slice, each track point in the slice is used as a node of the graph, and the edges of the graph are constructed according to the distance of the track points. The speed V of the aircraft in the three directions of longitude, latitude, and altitude are calculated. x 、V y 、V z As the attribute of each track point, the corresponding slice graph is constructed; among them, two track points with the same geohash prefix m are regarded as adjacent points and connected edges are established.

[0022] Further, in step 4, the trained track spatiotemporal embedding module is used to extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence; wherein the track spatiotemporal embedding module includes a feature extraction unit, a first fully connected layer, a multi-head attention layer and a second fully connected layer, wherein the feature extraction unit includes a first spatiotemporal convolutional network TCN layer, a GCN layer and a second spatiotemporal convolutional network TCN layer connected in sequence;

[0023] The first and second TCN layers are used to extract the temporal correlation features of the spatiotemporal graph in the time dimension. The encoded track sequence and the output of the first TCN layer are used as the input of the GCN layer, and the GCN layer is used to extract the spatial correlation features of the spatiotemporal graph in the spatial dimension. The input of the first fully connected layer is the word embedding of the track position sequence.

[0024] The spatiotemporal dependency features output by the feature extraction unit are used as the query vector, and the output of the first fully connected layer is used as the key vector and the value vector. The multi-head attention layer is used to map the spatiotemporal dependency features to the latent semantic space associated with the text, and then passes through the second fully connected layer to obtain the spatiotemporal dependency feature representation.

[0025] In a second aspect, the present invention provides a track prediction device based on a pre-trained large language model, comprising:

[0026] A data preprocessing module is used to obtain the historical track data of the target aircraft and perform data preprocessing to obtain a track sequence; wherein the information of each track point in the track sequence includes longitude, latitude, altitude and speed;

[0027] A prompt word embedding representation generation module is used to construct prompt words and input the prompt words into a pre-trained large language model to obtain word embedding representation; wherein the prompt words include a track prediction task description and a track data description;

[0028] Track sequence encoding and word segmentation module, used to generate geocoding for the longitude, latitude and altitude of each track point in the track sequence, obtain the corresponding track position sequence, combine the speed as the attribute of each track point with the track position sequence as the encoded track sequence, and perform word segmentation on the track position sequence;

[0029] A track sequence spatiotemporal graph construction and recoding module, used to construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and extract a spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence;

[0030] A track information feature representation module, used to fuse the word embedding representation and the spatiotemporal dependency feature representation using a pre-trained large language model to generate a track information feature representation;

[0031] The track prediction module is used to input the track information feature representation into the trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at a future time.

[0032] In a third aspect, the present invention 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 method described in the first aspect when executing the program.

[0033] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0034] Beneficial effects of the present invention:

[0035] (1) The present invention re-encodes the track data and maps it to the text space. Therefore, it is possible to use the large language model to learn the semantic relevance of trajectory points and implement the track prediction task without fine-tuning the existing pre-trained large language model. Only a small number of parameters need to be trained, thereby providing more reliable technical support for track prediction.

[0036] (2) In order to utilize the semantic relevance of trajectory point positions to provide more accurate track prediction results, the present invention constructs and recodes the spatiotemporal graph of the track sequence: first, the longitude, latitude, and altitude of the track are converted from numerical values ​​into textual geocoding, and then the trajectory sequence is modeled as a spatiotemporal dynamic graph. The temporal convolutional neural network and graph convolutional neural network are used to learn the temporal dependency and spatial dependency features of the trajectory spatiotemporal graph. Then, the spatiotemporal dependency features are modally aligned with the text modality through the attention mechanism, and the spatiotemporal dependency features of the track points are converted into a latent semantic space associated with the text, so as to facilitate the pre-trained large language model to learn and reason about it.

[0037] (3) The encoding of longitude, latitude and altitude designed by the present invention converts longitude, latitude and altitude from numerical values ​​into text form, thus avoiding the problem of difficulty in model training and damage to prediction accuracy caused by the large difference in longitude, latitude and altitude units.

[0038] (4) The present invention inputs the track data description and prediction task description through prompt words, which can better guide the large model to perform reasoning and also supplement the characteristics of the track sequence to a certain extent. It inputs it together with the re-encoded track sequence embedding vector representation into the pre-trained large language model, which can improve the robustness and accuracy of track prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A schematic flow chart of a track prediction method based on a pre-trained large language model provided by an embodiment of the present invention;

[0040] Figure 2 A track prediction framework diagram based on a pre-trained large language model provided in an embodiment of the present invention;

[0041] Figure 3 A structural block diagram of a track prediction device based on a pre-trained large language model provided by an embodiment of the present invention;

[0042] Figure 4 A structural block diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] Combination Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a track prediction method based on a pre-trained large language model, comprising the following steps:

[0045] S101: Acquire historical track data of the target aircraft and perform data preprocessing to obtain a track sequence; wherein information of each track point in the track sequence includes longitude, latitude, altitude and speed;

[0046] Specifically, the cleaned track sequence is obtained through data preprocessing methods such as data filtering and unified data formatting. Data filtering mainly includes deleting outliers in the track points and directly deleting a track data with too many missing values. Unified data formatting mainly includes normalizing the speed.

[0047] S102: constructing prompt words, and inputting the prompt words into a pre-trained large language model to obtain word embedding representation; wherein the prompt words include a track prediction task description and a track data description;

[0048] Specifically, the pre-trained large language model can use LLAMA 3, GLM4, etc. The track prediction task description can be "Please predict the longitude, latitude, and altitude of the aircraft in the next N time steps based on the sequence of the longitude, latitude, and altitude of the aircraft provided." The track data description is mainly a description of the statistical values ​​of a section of track data, mainly including the range of longitude, latitude, and altitude, the maximum and minimum average values ​​of speed in each direction, the time interval between each track point, and other basic description information.

[0049] After the prompt words of the track prediction task description and the track data description are concatenated, the pre-trained large language model is used to perform word segmentation and generate the word embedding representation P = [p1, p2, .., p n ]∈R n×d , where d is the dimension of the word vector and n is the number of tokens obtained after the prompt word is segmented.

[0050] In addition, if there is aircraft flight information, weather information, etc., it can also be added to the prompt words to improve the accuracy of the model prediction.

[0051] S103: Generate geocoding for the longitude, latitude and altitude of each track point in the track sequence to obtain a corresponding track position sequence, combine the speed as an attribute of each track point with the track position sequence as an encoded track sequence, and perform word segmentation on the track position sequence.

[0052] Specifically, longitude and latitude are encoded using the geohash algorithm, which converts longitude and latitude into K-bit base32 encoded values. The larger the K value, the higher the accuracy of the geohash value. When K=6, the error of geohash is about 0.61km. Geohash is a geographical code and a hierarchical data structure that divides space into grids through a binary method. Geohash is a practical application of the Z-order curve in the space filling curve. Geohash has a property related to the Z-order curve, that is, the geohash strings near a point (but not absolutely) always have a common prefix, and the longer the length of the common prefix, the closer the two points are.

[0053] For the height, the embodiment of the present invention also refers to the binary division idea of ​​geohash to perform encoding processing. In order to reduce the encoding complexity, the height value is first divided by 10, and then a height range [0, max_height] is determined, and binary division is performed according to the midpoint of the value range max_height / 2, and binary division is continued until a binary code that satisfies the accuracy is obtained. Similar to the longitude range of [-180, 180], the point with a longitude of 104.67 during the first binary division is divided into the right half of the interval, and the encoding value is 1; otherwise, it is divided into the left half of the interval, and the encoding value is 0.

[0054] Furthermore, the binary codes corresponding to the longitude, latitude, and altitude are converted into base32 codes, that is, 5 bits of binary are converted into one bit of base32 code.

[0055] It should be noted that the speed sequences in the three directions of longitude, latitude and altitude do not need to be encoded and processed. The speed sequences will be used as attributes of the track points to assist the model in learning.

[0056] After encoding and processing the longitude, latitude, and altitude in the track sequence, a tokenizer is used to segment the track position sequence, and the track position sequence is represented by a token id sequence after segmentation, so that it can be input into the large language model, and the text understanding ability of the large language model is used to learn the semantic relevance between track points. Among them, the tokenizer can use the BPE (BytePair Encoding) tokenizer, which selects the adjacent subwords with the highest frequency and merges them into a token.

[0057] S104: constructing a spatiotemporal graph of the track sequence based on the encoded track sequence, and extracting a spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence.

[0058] Specifically, for constructing the spatiotemporal graph of the track sequence, as an implementable method, the track sequence is sliced ​​according to the time window T, and a graph is constructed for the track sequence in each slice: each track point in each slice is a node of the graph, and the edges of the graph are constructed according to the distance between the track points (two track points with a distance less than the threshold are connected), and the speed V of the aircraft in the three directions of longitude, latitude, and altitude is x 、V y 、V z As the attribute of each track point. Among them, as a feasible implementation method, the common prefix feature of longitude and latitude encoding can be used to measure the position of the track point. Two track points with the same geohash prefix m are considered as adjacent points and can establish an edge; the value of m is determined according to the geohash accuracy.

[0059] For extracting the spatiotemporal dependency feature representation of the track sequence, such as Figure 2As shown, the trained track spatiotemporal embedding module is used to extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence. In the track spatiotemporal embedding module, it mainly includes a feature extraction unit, a first fully connected layer, a multi-head attention layer and a second fully connected layer, wherein the feature extraction unit includes a first spatiotemporal convolutional network TCN layer, a GCN layer and a second spatiotemporal convolutional network TCN layer connected in sequence; the embodiment of the present invention proposes to use two different temporal convolutional network TCN layers to fully extract the temporal correlation features of the spatiotemporal graph in the time dimension, the first TCN layer uses a larger convolution kernel to extract the global temporal features of the spatiotemporal graph, and the second TCN layer uses a smaller convolution kernel to extract the local temporal features of the spatiotemporal graph; the encoded track sequence and the output of the first TCN layer are used as the input of the GCN layer, and the GCN layer is used to extract spatial correlation features on the slice graph of each slice. The input of the first fully connected layer is the word embedding of the track position sequence (the word embedding can be obtained using a pre-trained word vector model, such as the BERT word vector model or the existing LLM generated word embedding). Subsequently, the spatiotemporal dependency features output by the feature extraction unit are used as the query vector, and the output of the first fully connected layer is used as the key vector and value vector. The multi-head attention mechanism is used to map the obtained spatiotemporal dependency features to the latent semantic space associated with the text, thereby obtaining an embedding vector representation of the trajectory sequence after the spatiotemporal graph of the track sequence is aligned with the text modality, which is the final required spatiotemporal dependency feature representation.

[0060] S105: Using a pre-trained large language model, the word embedding representation and the spatiotemporal dependency feature representation are fused to generate a track information feature representation.

[0061] Specifically, the vectors of step 102 and step 104 are concatenated and input into the pre-trained large language model with frozen parameters to obtain the final feature representation of the track information. Through this step, the description of the track prediction task, the basic statistical information of the track data, and the spatiotemporal dependency characteristics of the track points are feature fused through the large language model.

[0062] S106: Input the track information feature representation into a trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at a future time.

[0063] Specifically, during the training process, the track information feature representation of the training set is input into the prediction module, and the model parameters are optimized by minimizing the mean square error loss function to complete the prediction of the longitude, latitude and altitude values ​​of the aircraft at the future time. Among them, the structure of the prediction module is as follows Figure 2 shown.

[0064] The track prediction method provided by the embodiment of the present invention pre-places a track recoding module and adds prompt word guidance on the basis of not changing the existing pre-trained large language model. Only a small number of parameters need to be trained to achieve the track prediction task using the existing large language model, thereby providing more reliable technical support for track prediction.

[0065] The present invention aims at the geocoding of longitude, latitude and altitude, and converts longitude, latitude and altitude from numerical values ​​into text form, thus avoiding the problem of difficulty in model training and damage to prediction accuracy caused by the large difference in longitude, latitude and altitude units.

[0066] Corresponding to the above method, such as Figure 3 As shown, an embodiment of the present invention also provides a track prediction device based on a pre-trained large language model, including: a data preprocessing module, a prompt word embedding representation generation module, a track sequence encoding and segmentation module, a track sequence spatiotemporal graph construction and re-encoding module, a track information feature representation module and a track prediction module.

[0067] Among them, the data preprocessing module is used to obtain the historical track data of the target aircraft and perform data preprocessing to obtain a track sequence; wherein, the information of each track point in the track sequence includes longitude, latitude, altitude and speed. The prompt word embedding representation generation module is used to construct prompt words and input the prompt words into the pre-trained large language model to obtain word embedding representation; wherein, the prompt words contain track prediction task description and track data description. The track sequence encoding and word segmentation module is used to generate geocoding for the longitude, latitude and altitude of each track point in the track sequence to obtain the corresponding track position sequence, and combine the speed as the attribute of each track point with the track position sequence as the encoded track sequence, and perform word segmentation on the track position sequence. The track sequence spatiotemporal graph construction and recoding module is used to construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence. The track information feature representation module is used to fuse the word embedding representation and the spatiotemporal dependency feature representation using a pre-trained large language model to generate a track information feature representation. The track prediction module is used to input the track information feature representation into a trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at a future time.

[0068] The device provided in the embodiment of the present invention is to implement the above method. The specific functions thereof can be referred to the above method embodiment, which will not be described in detail here.

[0069] In the device provided by the embodiment of the present invention, the track sequence spatiotemporal graph construction and recoding module first converts the longitude, latitude and altitude of the track from numerical values ​​into geocoding in text form, and then models the track sequence as a spatiotemporal dynamic graph. The time dependency and space dependency features of the track sequence spatiotemporal graph can be effectively captured by using the temporal convolutional neural network and the graph convolutional neural network. The spatiotemporal dependency features are then modally aligned with the text modality through the attention mechanism, and the spatiotemporal dependency features of the track points are converted into latent semantic space associated with the text, thereby facilitating the pre-trained large language model to learn and reason about it.

[0070] The present invention inputs the basic features of the track data and the description of the prediction task through prompt words, thereby guiding the large language model to perform reasoning, and maps the text and track sequence spatiotemporal graph into a unified semantic space for effective interaction. It has high processing efficiency and can be effectively applied to long- and short-term track prediction tasks.

[0071] Figure 4 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 4 As shown, the electronic device may include: a processor (processor) 401, a communication interface (Communications Interface) 402, a memory (memory) 403 and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404. The processor 401 can call the logic instructions in the memory 403 to execute a track prediction method based on a pre-trained large language model, the method comprising: obtaining the historical track data of the target aircraft and performing data preprocessing to obtain a track sequence; wherein the information of each track point in the track sequence includes longitude, latitude, altitude and speed; constructing a prompt word, and inputting the prompt word into the pre-trained large language model to obtain a word embedding representation; wherein the prompt word contains a track prediction task description and a track data description; generating a geocode for the longitude, latitude, and altitude of each track point in the track sequence to obtain a geocode for the track sequence. The method comprises the following steps: a first step is to generate a corresponding track position sequence, and the speed is combined with the track position sequence as an attribute of each track point as a coded track sequence, and the track position sequence is segmented; a spatiotemporal graph of the track sequence is constructed based on the coded track sequence, and the spatiotemporal dependency feature representation of the track sequence is extracted from the spatiotemporal graph and the track position sequence; a pre-trained large language model is used to fuse the word embedding representation and the spatiotemporal dependency feature representation to generate a track information feature representation; the track information feature representation is input into a trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at a future time.

[0072] In addition, when the logic instructions in the above-mentioned memory 403 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0073] The embodiment of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the methods provided by the above-mentioned method embodiments, for example, including: obtaining the historical track data of the target aircraft and performing data preprocessing to obtain a track sequence; wherein the information of each track point in the track sequence includes longitude, latitude, altitude and speed; constructing a prompt word, and inputting the prompt word into a pre-trained large language model to obtain a word embedding representation; wherein the prompt word includes a track prediction task description and a track data description. ; Generate geocoding for the longitude, latitude and altitude of each track point in the track sequence to obtain the corresponding track position sequence, combine the speed as the attribute of each track point with the track position sequence as the encoded track sequence, and perform word segmentation on the track position sequence; construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence; use a pre-trained large language model to fuse the word embedding representation and the spatiotemporal dependency feature representation to generate a track information feature representation; input the track information feature representation into the trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at the future time.

[0074] The embodiment of the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the methods provided by the above-mentioned method embodiments are implemented, for example, including: obtaining the historical track data of the target aircraft and performing data preprocessing to obtain a track sequence; wherein the information of each track point in the track sequence includes longitude, latitude, altitude and speed; constructing a prompt word, and inputting the prompt word into a pre-trained large language model to obtain a word embedding representation; wherein the prompt word includes a track prediction task description and a track data description; the longitude, latitude, altitude and speed of each track point in the track sequence are calculated; The method generates a geographic code based on the degree to obtain a corresponding track position sequence, combines the speed as the attribute of each track point with the track position sequence as an encoded track sequence, and performs word segmentation on the track position sequence; constructs a spatiotemporal graph of the track sequence based on the encoded track sequence, and extracts the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence; uses a pre-trained large language model to fuse the word embedding representation and the spatiotemporal dependency feature representation to generate a track information feature representation; inputs the track information feature representation into a trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at a future time.

[0075] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A track prediction method based on a pre-trained large language model, characterized in that: include: Step 1: Acquire the historical track data of the target aircraft and perform data preprocessing to obtain a track sequence; wherein the information of each track point in the track sequence includes longitude, latitude, altitude and speed; Step 2: Construct prompt words and input the prompt words into the pre-trained large language model to obtain word embedding representation; wherein the prompt words include the track prediction task description and the track data description; Step 3: Generate geocoding for the longitude, latitude, and altitude of each track point in the track sequence to obtain the corresponding track position sequence, combine the speed as the attribute of each track point with the track position sequence as the encoded track sequence, and perform word segmentation on the track position sequence; Step 4: constructing a spatiotemporal graph of the track sequence based on the encoded track sequence, and extracting a spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence; Step 5: Using a pre-trained large language model to fuse the word embedding representation and the spatiotemporal dependency feature representation to generate a track information feature representation; Step 6: Input the track information feature representation into the trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at a future time.

2. A track prediction method based on a pre-trained large language model according to claim 1, characterized in that: In step 3, geocoding is generated for the longitude, latitude, and altitude of each track point in the track sequence, including: Use the geohash algorithm to encode the longitude and latitude respectively and generate their corresponding binary codes; Set a range of altitude values, and divide the altitude into two parts according to the midpoint of the range until a binary code with satisfactory accuracy is obtained.

3. A track prediction method based on a pre-trained large language model according to claim 1, characterized in that: In step 3, the BPE tokenizer is used to tokenize the track position sequence.

4. A track prediction method based on a pre-trained large language model according to claim 2, characterized in that: In step 4, a spatiotemporal graph of the track sequence is constructed based on the encoded track sequence, which specifically includes: Slice the track sequence according to the preset time window; For each slice, each track point in the slice is used as a node of the graph, and the edges of the graph are constructed according to the distance of the track points. The speed V of the aircraft in the three directions of longitude, latitude, and altitude are calculated. x 、V y 、V z As the attribute of each track point, the corresponding slice graph is constructed; among them, two track points with the same geohash prefix m are regarded as adjacent points and connected edges are established.

5. The method for track prediction based on a pre-trained large language model according to claim 1, characterized in that: In step 4, the trained track spatiotemporal embedding module is used to extract the spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence; wherein the track spatiotemporal embedding module includes a feature extraction unit, a first fully connected layer, a multi-head attention layer and a second fully connected layer, wherein the feature extraction unit includes a first spatiotemporal convolutional network TCN layer, a GCN layer and a second spatiotemporal convolutional network TCN layer connected in sequence; The first and second TCN layers are used to extract the temporal correlation features of the spatiotemporal graph in the time dimension. The encoded track sequence and the output of the first TCN layer are used as the input of the GCN layer, and the GCN layer is used to extract the spatial correlation features of the spatiotemporal graph in the spatial dimension. The input of the first fully connected layer is the word embedding of the track position sequence. The spatiotemporal dependency features output by the feature extraction unit are used as the query vector, and the output of the first fully connected layer is used as the key vector and the value vector. The multi-head attention layer is used to map the spatiotemporal dependency features to the latent semantic space associated with the text, and then passes through the second fully connected layer to obtain the spatiotemporal dependency feature representation.

6. A track prediction device based on a pre-trained large language model, characterized in that: include: A data preprocessing module is used to obtain the historical track data of the target aircraft and perform data preprocessing to obtain a track sequence; wherein the information of each track point in the track sequence includes longitude, latitude, altitude and speed; A prompt word embedding representation generation module is used to construct prompt words and input the prompt words into a pre-trained large language model to obtain word embedding representation; wherein the prompt words include a track prediction task description and a track data description; Track sequence encoding and word segmentation module, used to generate geocoding for the longitude, latitude and altitude of each track point in the track sequence, obtain the corresponding track position sequence, combine the speed as the attribute of each track point with the track position sequence as the encoded track sequence, and perform word segmentation on the track position sequence; A track sequence spatiotemporal graph construction and recoding module, used to construct a spatiotemporal graph of the track sequence based on the encoded track sequence, and extract a spatiotemporal dependency feature representation of the track sequence from the spatiotemporal graph and the track position sequence; A track information feature representation module, used to fuse the word embedding representation and the spatiotemporal dependency feature representation using a pre-trained large language model to generate a track information feature representation; The track prediction module is used to input the track information feature representation into the trained prediction module to obtain the predicted values ​​of the longitude, latitude and altitude of the target aircraft at a future time.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images