Model training method, model prediction method and related devices

Abstract

The application provides a model training method, a model prediction method and related devices, and relates to the technical field of model training and reasoning. The model training method comprises the following steps: inputting track sample data into a track large model to be trained; wherein the track sample data comprises a track point sequence of an aircraft and a task instruction token located before the track point sequence, and the task instruction token is used to indicate different prediction tasks about the aircraft; the different prediction tasks comprise at least two of the following: a track prediction task, a track segmentation task and an aircraft model identification task; through the track large model, a prediction result corresponding to a prediction task indicated by the task instruction token is obtained according to the task instruction token and the track sample data; and the model parameters of the track large model are adjusted according to the prediction result. The technical scheme provided by the application can solve the problem of how to provide effective support for low-altitude flight management in the prior art.

Description

Technical Field

[0001] This application relates to the field of model training and inference technology, and more specifically, to a model training method, a model prediction method, and related apparatus. Background Technology

[0002] With the rapid development of the low-altitude economy, various types of aircraft, such as civilian unmanned aerial vehicles and electric vertical take-off and landing aircraft, are increasingly being used in logistics, surveying and mapping, emergency rescue, and urban commuting. The low-altitude airspace is gradually becoming a high-density, high-dynamic, and complex operating environment.

[0003] The surge in low-altitude flight activities has brought unprecedented challenges to airspace management, and how to provide effective support for low-altitude flight management has become an urgent technical problem to be solved. Summary of the Invention

[0004] Based on the defects and shortcomings of the existing technology, this application proposes a model training method, a model prediction method, and related devices, which can solve the problem of how to provide effective support for low-altitude flight management in the existing technology.

[0005] According to a first aspect of the embodiments of this application, a model training method is provided, the method comprising: The trajectory sample data is input into the trajectory large model to be trained; wherein, the trajectory large model is used to perform prediction tasks related to the aircraft based on the input data; the trajectory sample data includes a sequence of trajectory points of the aircraft and a task indication token located before the sequence of trajectory points, the task indication token being used to indicate different prediction tasks about the aircraft; the different prediction tasks include at least two of the following: trajectory prediction task, trajectory segmentation task, and aircraft type identification task; Using the large trajectory model, based on the task indication token and the trajectory sample data, the prediction result corresponding to the prediction task indicated by the task indication token is obtained; Adjust the model parameters of the large trajectory model based on the prediction results.

[0006] According to a second aspect of the embodiments of this application, a model prediction method is provided, applied to a large trajectory model trained by the model training method described in the first aspect, the method comprising: Receive track data; wherein the track data includes a sequence of waypoints of the aircraft and a mission indication token located before the sequence of waypoints, the mission indication token being used to indicate different prediction tasks about the aircraft; the different prediction tasks include at least two of the following: track prediction task, track segmentation task, and aircraft type identification task; Based on the task instruction token, perform the corresponding prediction task on the track data and output the prediction result.

[0007] According to a third aspect of the embodiments of this application, a model training apparatus is provided, the apparatus comprising: An input module is used to input track sample data into a track model to be trained; wherein, the track model is used to perform prediction tasks related to the aircraft based on the input data; the track sample data includes a sequence of track points of the aircraft and a task indication token located before the sequence of track points, the task indication token being used to indicate different prediction tasks related to the aircraft; the different prediction tasks include at least two of the following: track prediction task, track segmentation task, and aircraft type identification task; The training module is used to obtain the prediction result corresponding to the prediction task indicated by the task indicator token by using the large trajectory model, based on the task indicator token and the trajectory sample data; The adjustment module is used to adjust the model parameters of the large trajectory model based on the prediction results.

[0008] According to a fourth aspect of the embodiments of this application, a model prediction apparatus is provided, applied to a large trajectory model trained by the model training method described in the first aspect, the apparatus comprising: A receiving module is used to receive track data; wherein the track data includes a sequence of waypoints of the aircraft and a task indication token located before the sequence of waypoints, the task indication token being used to indicate different prediction tasks about the aircraft; the different prediction tasks include at least two of the following: track prediction task, track segmentation task, and aircraft type identification task; The prediction module is used to perform a corresponding prediction task on the track data according to the task instruction token, and output the prediction result.

[0009] According to a fifth aspect of the embodiments of this application, an electronic device is provided, including: a memory and a processor; The memory is connected to the processor and is used to store programs; The processor is used to implement the method as described in the first or second aspect by running a program in the memory.

[0010] According to a sixth aspect of the embodiments of this application, a storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the method as described in the first or second aspect.

[0011] According to a seventh aspect of the embodiments of this application, a computer program product or a computer program is provided, the computer program product including the computer program, wherein a processor executes the computer program to implement the steps in the method as described in the first or second aspect.

[0012] In the technical solution provided in this application, by adding task indicator tokens to the track sample data, the track large model can learn the mapping relationship between the task indicator tokens and different prediction tasks during the training process. This enables it to distinguish different prediction tasks based on the task indicator tokens and execute the corresponding prediction tasks, thereby realizing the ability to jointly model multiple tasks. This achieves the goal of understanding and processing multiple complex tasks through a unified model, providing effective support for the management of low-altitude flight. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating a model training method provided in an embodiment of this application.

[0015] Figure 2 This is one of the schematic diagrams of track sample data provided in the embodiments of this application.

[0016] Figure 3 This is a second schematic diagram of the trajectory sample data provided in the embodiments of this application.

[0017] Figure 4 The third schematic diagram of the trajectory sample data provided in the embodiments of this application.

[0018] Figure 5 The fourth schematic diagram of the trajectory sample data provided in the embodiments of this application.

[0019] Figure 6 This is a flowchart illustrating a model prediction method provided in an embodiment of this application.

[0020] Figure 7 This is a block diagram of a model training device provided in an embodiment of this application.

[0021] Figure 8 This is a block diagram of a model prediction device provided in an embodiment of this application.

[0022] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] With the rapid development of the low-altitude economy and the increasing prevalence of various types of aircraft, the surge in low-altitude flight activities has brought unprecedented challenges to airspace management. For example, the lack of accurate prediction capabilities regarding aircraft intentions and future trajectories leads to delayed airspace situational awareness, making it difficult to provide early warnings and guidance for potential flight conflicts and regional congestion. Furthermore, in dense flight environments, the trajectories of different aircraft are prone to intersecting and overlapping (i.e., "trajectory entanglement"), posing significant difficulties for accurate tracking, liability determination, and safety supervision. Finally, the diverse types and performance of low-altitude aircraft render traditional identification methods inefficient and unable to provide effective support for refined traffic flow management.

[0025] In related technologies, most focus on solving single tasks, such as using independent Kalman filters, Convolutional Neural Networks (CNNs), or Recurrent Neural Networks (RNNs) models for trajectory prediction or target recognition. This "information silo" approach lacks the ability to mine and utilize the deep correlations between multimodal and multi-task data from low-altitude flights, making it difficult to fundamentally address the "invisible and uncontrollable" congestion pressures and safety risks associated with low-altitude flights.

[0026] To address this, this application provides a solution that adds a task indicator token to the input data of the large trajectory model. This enables the large trajectory model to distinguish different prediction tasks based on the task indicator token and execute the corresponding prediction tasks, thereby achieving multi-task joint modeling capabilities. This allows the model to understand and process multiple complex tasks through a unified model, providing effective support for the management of low-altitude flight.

[0027] Exemplary methods This application provides a model training method applied to a first electronic device, which can be a terminal device such as a tablet computer, laptop computer, desktop computer, or mobile phone, or a server such as a cloud server. The first electronic device contains a large trajectory model to be trained, which is used to perform aircraft-related prediction tasks based on input data.

[0028] The method is described in detail below through some embodiments. The following embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0029] like Figure 1 As shown, the model training method may include steps 101 to 103, as described below.

[0030] Step 101: Input the track sample data into the track model to be trained.

[0031] The track sample data includes the aircraft's track point sequence and the mission indication token set before the track point sequence.

[0032] The trackpoint sequence includes multiple trackpoints, each of which can include five attribute information: the longitude (lat), latitude (lon), speed (sog), heading (cog), and type of the aircraft.

[0033] In this embodiment of the application, for each attribute att∈{lat,lon,sog,cog,type}, its continuous observations can be discretized as follows: The discretized values ​​of different attributes at the same time constitute a track point in the track sample data. The discretization formula can be shown in formula (1) below: (1) in, Let be the track point at time t. For each attribute, a discretization function is provided. These represent the discretized values ​​of each attribute.

[0034] The task token described herein is used to indicate different prediction tasks related to the aircraft. These different prediction tasks may include, but are not limited to, at least two of the following: track prediction task, track segmentation task, aircraft type identification task, etc.

[0035] The trajectory prediction task refers to the task of predicting trajectories given at times 0 to T. Predict the track point at the next moment. After making L consecutive predictions, we obtain... The value of L can be set according to actual needs.

[0036] Track segmentation, also known as track compositing, refers to a task that involves defining a track from time 0 to T. The trajectory is formed by the superposition of m independent trajectories. The goal of this task is to output m independent trajectories, i.e. ,in .

[0037] Aircraft type identification is the task of identifying the type of aircraft, and it belongs to the category of classification tasks. Aircraft types can include, but are not limited to, electric vertical takeoff and landing (eVTOL) aircraft, unmanned aerial vehicles (UAVs), helicopters, and fixed-wing aircraft. Aircraft type identification can be achieved based on information such as longitude (lat), latitude (lon), airspeed (sog), and heading (cog).

[0038] Optionally, the mission indication token can indicate different prediction tasks for the aircraft through different assignments to its mission indication field. Therefore, the assignments to the mission indication field can include at least two of the following: a first assignment for indicating a track prediction task, a second assignment for indicating a track segmentation task, and a third assignment for indicating an aircraft type identification task.

[0039] For example, the task indicator field in a task indicator token can be represented as According to The values ​​represent different prediction tasks, as described below: Understandably, the values ​​assigned to different tasks can be set according to actual needs; this is just an example.

[0040] Suppose a certain path includes 256 waypoints. When using this path for path prediction, the resulting path sample data can be as follows: Figure 2 As shown. The Task Token is located before the sequence of 256 waypoints, in the task indication field. The value of 1 (in the TASK field of the diagram) indicates that a trajectory prediction task is to be performed. If the task indication field... If the value is 2, it indicates that a track segmentation task is to be performed. If the task indication field... If the value is set to 3, it is used to indicate that an aircraft type identification task is to be performed.

[0041] Depend on Figure 2 It is known that each token in the track sample data includes the same attribute fields. To facilitate the differentiation of Task Tokens, the following can be set: if a token contains... If a field is non-zero and other attribute fields are zero, then the token is a Task Token.

[0042] Step 102: Using the large trajectory model, based on the mission indicator token and trajectory sample data, obtain the prediction result corresponding to the predicted task indicated by the mission indicator token.

[0043] After the trajectory sample data is input into the trajectory big model, the trajectory big model can perform a prediction task on the trajectory sample data based on the task indication token and output the prediction results.

[0044] Step 103: Adjust the model parameters of the large trajectory model based on the prediction results.

[0045] Since the large-scale trajectory model is not a language model and lacks the cue words of a language model, it cannot distinguish between different prediction tasks. Therefore, this application adds task indicator tokens to the trajectory sample data. During training, the large-scale trajectory model can learn the mapping relationship between the task indicator tokens and the corresponding prediction tasks. This allows it to distinguish different prediction tasks based on the task indicator tokens and accurately execute the corresponding prediction tasks. This achieves the goal of understanding and processing multiple complex tasks through a single model, realizing multi-task joint modeling capabilities and providing effective support for the management of low-altitude flight. For example, through trajectory prediction tasks, accurate prediction of the future trajectory and intentions of aircraft can be achieved, thereby providing early warning and guidance for potential flight conflicts and regional congestion. Another example is the trajectory segmentation task, which enables accurate tracking, responsibility identification, and safety supervision of aircraft. Yet another example is the aircraft type identification task, which allows for refined traffic management for different types of aircraft. These all provide effective support for the management of low-altitude flight.

[0046] In some alternative embodiments, for track segmentation tasks, the track sample data may include a mixed sequence of track points from multiple aircraft. For the prediction results of the large track model, track segmentation tokens can be used to distinguish different track segments. These tokens are used to segment the tracks of different aircraft.

[0047] Therefore, step 102: using the large trajectory model, based on the mission indication token and trajectory sample data, to obtain the prediction result corresponding to the predicted task indicated by the mission indication token, may include: Using a large trajectory model, a trajectory segmentation task is performed on the trajectory sample data according to the task instruction token to obtain the trajectory segmentation result.

[0048] The track segmentation result includes at least one track segmentation token.

[0049] The validity of a track split token can be indicated by the assignment of a track split field, which includes this field. For example, the track split field can be represented as... When this field is set to 1, it indicates that the track split token is valid; when this field is set to 0, it indicates that the track split token is invalid.

[0050] Suppose a certain path contains two sub-paths, each containing 128 waypoints. An attempt is made to perform path segmentation on this path, and the prediction result can be as follows: Figure 3 As shown, the two sub-tracks are separated by a Sep Token. The mission indication field in this token... The SEP field is set to 0, and the other attribute fields are also set to 0, indicating that the token is currently invalid.

[0051] Depend on Figure 3 As can be seen, each token in the prediction results includes the same attribute fields. To facilitate differentiation of SepTokens, we can set: if a token contains... If the field is non-zero and all other attribute fields are zero, then the token is a SepToken. If a field is 0, it indicates a continuous flight path.

[0052] In some alternative embodiments, the training process of the large trajectory model may include two stages: a pre-training stage and a fine-tuning stage. In the pre-training stage, only autoregressive prediction of trajectory points is performed. In the fine-tuning stage, fine-tuning is performed for different domain tasks, such as trajectory prediction, trajectory segmentation, and aircraft type recognition.

[0053] To distinguish between the two training phases, the task indicator tokens used to indicate different prediction tasks may also include pre-training tasks. Therefore, step 101: inputting track sample data into the track large model to be trained may include: Input the first track sample data into the track large model; wherein, the task indicator token in the first track sample data is used to indicate the pre-training task.

[0054] After completing the pre-training task, the second track sample data is input into the track large model; the task indicator token in the second track sample data is used to indicate downstream tasks such as track prediction task, track segmentation task or aircraft type identification task.

[0055] Building upon this, the assignment of the task indication field can also include a fourth assignment for indicating the pre-training task, as described below: In the task instruction field When the value is set to 4, it is used to indicate that a pre-training task should be performed.

[0056] Understandably, the values ​​assigned to different tasks can be set according to actual needs; this is just an example.

[0057] Among them, the pre-training task can refer to the use of track sample data by the large track model to perform autoregressive prediction of future track points.

[0058] Autoregressive prediction can specifically be: the model predicts the input... (or In the case of ), the aim is to output the optimal continuous value. .

[0059] The model structure can employ an autoregressive self-attention decoding mechanism (Transformer Decoder), allowing... Probability function of track points at time t. Then you can The prediction problem is modeled as a concatenation of K classification distributions, corresponding to five attributes: {latitude, longitude, speed, heading, and model}. The prediction problem is then transformed into a classification problem with K output heads. The loss function is shown in formula (2) below: (2) In formula (2), CE represents the cross-entropy function. Indicates the loss value. express The track point at any given moment.

[0060] in, The embedding vector represents the track points from time 0 to time T.

[0061] In the discrete values After inputting the large trajectory model, the model can be used for each... Generate high-dimensional real-valued embedding vectors (like By concatenating the embedding vectors of each attribute together (e.g., using the concat function), a joint representation vector of the waypoints can be obtained. The above process can be represented by the following formulas (3) and (4).

[0062] (3) (4) In formula (3), For vector embedding functions. In formula (4), Let be the embedding vector of the track point at time t.

[0063] In some optional embodiments, when the prediction task of the large trajectory model is any one of trajectory prediction, trajectory segmentation, or aircraft type identification, label data is required. Therefore, step 103: adjusting the model parameters of the large trajectory model based on the prediction results may include: Adjust the model parameters of the large trajectory model based on the prediction results and pre-set label data.

[0064] When the prediction task of the large trajectory model is any one of trajectory prediction, trajectory segmentation, or aircraft type identification, the loss value can be calculated based on the prediction results of the large trajectory model and the pre-set label data (i.e., Label data), and then the model parameters can be adjusted based on the loss value.

[0065] In order to enable the large-scale trajectory model to learn the end of such prediction tasks, such as when performing trajectory prediction tasks, the large-scale trajectory model can learn to predict a certain number of future trajectory points before stopping prediction, an end token can be set at the end of the label data to indicate the end of the prediction task.

[0066] The end indicator token can indicate the end of the prediction task by assigning a value to the end indicator field it includes. For example, the end indicator field can be represented as... When this field is set to 1, it indicates that the end indicator token is valid. When this field is set to 0, it indicates that the end indicator token is invalid.

[0067] Suppose that the label data for a certain path prediction task includes 256 path points, such as Figure 4 As shown. In this tag data, the Task Token is located at the beginning of the entire data sequence, and its task indicator field... The Task Token is invalid when the TASK field is set to 0, and all other attribute fields are also set to 0. The End Token is located at the end of the 256 waypoint sequence, which is the last element of the entire data sequence; it is the end indicator field. The value of 1 for the END field indicates the end of the trajectory prediction task, which can also be understood as the end of a trajectory.

[0068] Depend on Figure 4 As can be seen, each token in the track sample data includes the same attribute fields. To facilitate differentiation of End Tokens, we can set: if a token contains... If a field is non-zero and other attribute fields are zero, then the token is an End Token.

[0069] It should be noted that, as Figure 5The image shows the tag data for a certain track segmentation task. In this tag data, the track segmentation field and the end indication field in the token may both be non-zero, making it impossible to determine whether a token is a SepToken or an EndToken. Therefore, based on the positions of the two types of tokens, the EndToken can be determined first, and then the SepToken can be determined.

[0070] It should also be noted that pre-training learns trajectory patterns autonomously through massive amounts of unlabeled trajectory data. Therefore, the trajectory sample data used for pre-training may not contain end indicator tokens.

[0071] Optionally, if the training process of the track large model includes a pre-training stage and a fine-tuning stage, the fine-tuning stage can concatenate the input data of the track large model with the expected output data to form the track sample data of the track large model, and the same applies to the label data.

[0072] However, since the input data is known, there is no need to repeatedly learn this part. Instead, the label data should be fully learned. Therefore, this embodiment can employ a loss mask mechanism to mask the input data portion of the label data. This way, when calculating the loss value, the model will only calculate the loss value of the expected output portion, while the masked portion will be ignored, preventing the input data from affecting gradient updates. The specific implementation is described below.

[0073] Before the aforementioned step of adjusting the model parameters of the large trajectory model based on the prediction results and pre-set label data, the method may further include steps A1 to A3, as described below: Step A1: Generate the input data and expected output data of the large trajectory model based on the actual waypoint sequence of the aircraft.

[0074] Step A2: Concatenate the input data with the expected output data to generate track sample data.

[0075] Step A3: Concatenate the input data with the expected output data, and mask the input data in the concatenation result to generate label data.

[0076] The embodiments of this application can generate input data and expected output data of a large trajectory model based on the actual trajectory point sequence of the aircraft, and then generate trajectory sample data and label data based on the input data and expected output data.

[0077] In this embodiment, a preset category number can be used to replace each track point in the input data, thereby masking a portion of the input data. This preset category number indicates which content should be ignored, and can be -100. When the label data contains a preset category number to be ignored (default is -100), the loss value for that portion will not be calculated, and thus that portion of the data will not affect gradient updates. This allows for the training of a more reliable and effective track model.

[0078] The generated track sample data includes a mission indication token. The generated label data includes an end indication token.

[0079] For the above-described embodiment of generating track sample data and label data, when the prediction task of the large track model includes a track prediction task, step A1: generating input data and expected output data of the large track model based on the aircraft's actual track point sequence may include: Step A111: Randomly select a target waypoint from the real waypoint sequence.

[0080] Step A112: Generate input data based on the sequence of waypoints before the target waypoint, and generate expected output data based on the target waypoint and the sequence of waypoints after the target waypoint.

[0081] In this context, the target waypoint is a waypoint from the sequence of real waypoints. For the path prediction task, the number of waypoints in the input data needs to meet the prediction requirements. Therefore, the target waypoint can be randomly selected from the real waypoint sequence excluding the first M waypoints, where M is the minimum number of known waypoints required for path prediction. Of course, the target waypoint should also avoid being the last waypoint in the real waypoint sequence.

[0082] For example, the actual waypoint sequence is: { A random selection algorithm can be used to randomly select one waypoint as the target waypoint. Let's assume the waypoint is... As the target waypoint, it can be Previous track points As input data, as well as Subsequent waypoints As expected output data.

[0083] Next, input data Compared with expected output data By concatenating the data and adding a mission indicator token at the beginning of the sequence, you can obtain the track sample data. .

[0084] Input data Compared with expected output data Concatenate the sequences, add an end indicator token at the end of the sequence, and utilize preset category numbers. By replacing the input data, you can obtain the label data. .

[0085] Alternatively, a task instruction token can be added during the generation of input data, in which case the input data will be... Similarly, an end indicator token can be added during the generation of the expected output data, then the expected output data will be... Then, the input data is concatenated with the expected output data, and the end indicator token is removed to obtain the track sample data; and the input data is concatenated with the expected output data, and the task indicator token is shifted left by one bit or removed to obtain the label data.

[0086] The basic principle of large-scale trajectory model prediction is to predict the next trajectory point based on one or more known trajectory points. Therefore, trajectory prediction tasks can be performed even if the trajectory sample data includes the expected output data.

[0087] Optionally, input data can be generated based on the target waypoint and the sequence of waypoints preceding the target waypoint, and expected output data can be generated based on the sequence of waypoints following the target waypoint. The specific settings can be configured according to actual needs.

[0088] For the above-described embodiment of generating track sample data and label data, when the prediction task of the large track model includes a track segmentation task, step A1: generating input data and expected output data of the large track model based on the aircraft's actual track point sequence may include: Step A121: Mix the real waypoint sequences of multiple aircraft to generate input data.

[0089] Step A122: For multiple aircraft real waypoint sequences, add a waypoint segmentation token to the end of each aircraft's real waypoint sequence to generate the expected output data.

[0090] In this embodiment, multiple aircraft's actual waypoint sequences within the same time period can be obtained, and then these sequences can be mixed and superimposed to generate input data. Time alignment is required during the mixing and superimposition process.

[0091] For the expected output data, a track segmentation token can be added to the end of the sequence of actual track points for each aircraft to generate the expected output data.

[0092] Assume the input data is The expected output data is Input data Compared with expected output data By concatenating the data and adding a mission indicator token at the beginning of the sequence, you can obtain the track sample data. .

[0093] Input data Compared with expected output data Concatenate the sequences, add an end indicator token at the end of the sequence, and utilize preset category numbers. By replacing the input data, you can obtain the label data. .

[0094] Alternatively, a task instruction token can be added during the generation of input data, in which case the input data will be... Similarly, an end indicator token can be added during the generation of the expected output data, then the expected output data will be... Then, the input data is concatenated with the expected output data, and the end indicator token is removed to obtain the track sample data; and the input data is concatenated with the expected output data, and the task indicator token is shifted left by one bit or removed to obtain the label data.

[0095] The real waypoint sequence includes aircraft model information, such as each waypoint containing aircraft model information. Based on this, when the prediction task of the large-scale trajectory model includes aircraft model identification, step A1: Based on the real waypoint sequence of the aircraft, generate the input data and expected output data of the large-scale trajectory model, which may include: Step A131: Remove the aircraft model information from the real waypoint sequence and generate input data.

[0096] Step A132: Use the actual track point sequence as the expected output data.

[0097] In this embodiment, when generating training data for an aircraft type identification task, the aircraft type information in the real waypoint sequence can be removed to generate input data. The real waypoint sequence is then used as the expected output data. Alternatively, the aircraft type information in the real waypoint sequence can also be used as the expected output data.

[0098] The removal of aircraft model information can be achieved by setting the aircraft model field in each waypoint to 0.

[0099] Assume the input data is , for The token after setting the aircraft model field to 0. Expected output data is... Input data Compared with expected output data By concatenating the data and adding a mission indicator token at the beginning of the sequence, you can obtain the track sample data. .

[0100] Input data Compared with expected output data Concatenate the sequences, add an end indicator token at the end of the sequence, and utilize preset category numbers. By replacing the input data, you can obtain the label data. .

[0101] Alternatively, a task instruction token can be added during the generation of input data, in which case the input data will be... Similarly, an end indicator token can be added during the generation of the expected output data, then the expected output data will be... Then, the input data is concatenated with the expected output data, and the end indicator token is removed to obtain the track sample data; and the input data is concatenated with the expected output data, and the task indicator token is shifted left by one bit or removed to obtain the label data.

[0102] In this embodiment, training datasets for tasks such as trajectory prediction, trajectory segmentation, and aircraft type recognition can be automatically and massively synthesized using existing real trajectory data. This achieves efficient automatic synthesis of large-scale fine-tuning datasets without manual annotation, improving the efficiency of training data generation. Furthermore, the real trajectory sequences used here can be the trajectory point sequences used in the pre-training phase.

[0103] Finally, it should be noted that among all tokens... The field will be encoded into A vector, but it will not participate in the loss calculation. (Among all tokens) Fields The field will be encoded into The vector participates in the loss calculation and is also used as an attribute field in the classification head for prediction.

[0104] In summary, the technical solution provided in this application improves the model's instruction compliance by adding control tokens such as task indicator tokens, end indicator tokens, and track segmentation tokens to the training data (such as track sample data and label data) of the large track model, thereby forcing distribution alignment. Furthermore, this application embodiment can also use a loss mask mechanism to mask the input data portion of the label data in the fine-tuning task, thus achieving consistency between pre-training and fine-tuning task design. Further, this application embodiment can automatically synthesize the training dataset for the fine-tuning task from existing track data, achieving automatic synthesis of the fine-tuning dataset without manual zero-labeling, improving sample generation efficiency. Moreover, in the technical solution provided in this application embodiment, the input data for different prediction tasks are similar, enabling the mining and utilization of deep correlations between low-altitude flight multimodal and multi-task data.

[0105] This application also provides a model prediction method applied to a large trajectory model in a second electronic device. The large trajectory model is pre-trained using the above-described model training method and is used to perform prediction tasks related to the aircraft based on input data.

[0106] The second electronic device can be a terminal device, such as a tablet computer, laptop computer, desktop computer, or mobile phone, or it can be a server, such as a cloud server. The second electronic device can be the same as or different from the first electronic device.

[0107] The method is described in detail below through some embodiments. The following embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0108] like Figure 6 As shown, the model training method may include steps 601 to 602, as described below.

[0109] Step 601: Receive track data.

[0110] The track data includes the aircraft's track point sequence and the mission indication token set before the track point sequence.

[0111] The trackpoint sequence includes multiple trackpoints, each of which can include five attribute information: the longitude (lat), latitude (lon), speed (sog), heading (cog), and type of the aircraft.

[0112] The task token described herein is used to indicate different prediction tasks related to the aircraft. These different prediction tasks may include, but are not limited to, at least two of the following: track prediction task, track segmentation task, aircraft type identification task, etc.

[0113] Optionally, the mission indication token can indicate different prediction tasks for the aircraft through different assignments to its mission indication field. Therefore, the assignments to the mission indication field can include at least two of the following: a first assignment for indicating a track prediction task, a second assignment for indicating a track segmentation task, and a third assignment for indicating an aircraft type identification task.

[0114] For example, the task indicator field in a task indicator token can be represented as According to The values ​​represent different prediction tasks, as described below: Step 602: Based on the task instruction token, perform the corresponding prediction task on the track data and output the prediction results.

[0115] The large trajectory model trained by the model training method in the above embodiments can distinguish different prediction tasks based on the learned mapping relationship between the task indicator tokens and the corresponding prediction tasks, and accurately execute the corresponding prediction tasks. This achieves the goal of understanding and processing multiple complex tasks through a single model, realizes the ability to jointly model multiple tasks, and provides effective support for the management of low-altitude flight.

[0116] In some alternative embodiments, in order to indicate the end of the prediction content, an end indicator token can be set at the end of the prediction result output by the track large model, which indicates the end of the prediction content.

[0117] Optionally, the end indicator token can indicate the end of the prediction task by assigning a value to the end indicator field included therein. For example, the end indicator field can be represented as... When this field is set to 1, it indicates that the end indicator token is valid. When this field is set to 0, it indicates that the end indicator token is invalid.

[0118] In some alternative embodiments, for a track segmentation task, the track data may include a mixed sequence of waypoints from multiple aircraft. To distinguish different track segments in the prediction results for this task, track segmentation tokens can be used. These tokens are used to segment the tracks of different aircraft.

[0119] Based on this, when the task indication token indicates a track segmentation task, step 602: according to the task indication token, perform the corresponding prediction task on the track data and output the prediction result, which may include: Based on the task instruction token, perform a track segmentation task on the track data to obtain the track segmentation result.

[0120] The track segmentation result includes at least one track segmentation token.

[0121] The validity of a track split token can be indicated by the assignment of a track split field, which includes this field. For example, the track split field can be represented as... When this field is set to 1, it indicates that the track split token is valid. When this field is set to 0, it indicates that the track split token is invalid.

[0122] The technical solution provided in this application, by adding task indicator tokens to the input data of the large trajectory model, enables the large trajectory model to distinguish different prediction tasks based on the mapping relationship between the task indicator tokens and different prediction tasks, and to execute the corresponding prediction tasks, thereby achieving multi-task joint modeling capability. In this application embodiment, control tokens such as end indicator tokens or trajectory splitting tokens can also be added to the model's prediction results to improve the readability of the model's prediction results.

[0123] For some technical details regarding the model prediction method provided in this embodiment, please refer to the description of the model training method provided in the above embodiments of this application, which will not be repeated here.

[0124] Exemplary device Accordingly, this application also provides a model training device applied to a first electronic device, which can be a terminal device, such as a tablet computer, laptop computer, desktop computer, or mobile phone, or a server, such as a cloud server. The first electronic device contains a large trajectory model to be trained, which is used to perform prediction tasks related to the aircraft based on input data.

[0125] like Figure 7 As shown, the device may include: Input module 701 is used to input track sample data into the track model to be trained.

[0126] The track sample data includes a sequence of track points for the aircraft and a mission indication token preceding the track point sequence. The mission indication token is used to indicate different prediction tasks for the aircraft. The different prediction tasks include at least two of the following: track prediction task, track segmentation task, and aircraft type identification task.

[0127] Training module 702 is used to obtain a prediction result corresponding to the prediction task indicated by the task indicator token, based on the task indicator token and the track sample data, using the track large model.

[0128] The adjustment module 703 is used to adjust the model parameters of the large trajectory model based on the prediction results.

[0129] Optionally, the task indicator token may also indicate pre-trained tasks as different prediction tasks.

[0130] The input module 701 can be specifically used to: input the first track sample data into the track large model; and after completing the pre-training task, input the second track sample data into the track large model.

[0131] The task indicator token in the first track sample data is used to indicate a pre-training task; the task indicator token in the second track sample data is used to indicate a track prediction task, a track segmentation task, or an aircraft type identification task.

[0132] Optionally, the mission indication token indicates different predicted missions about the aircraft through different assignments to its included mission indication field.

[0133] The assignment of the task indication field includes at least two of the following: a first assignment for indicating a trajectory prediction task, a second assignment for indicating a trajectory segmentation task, a third assignment for indicating an aircraft type identification task, and a fourth assignment for indicating a pre-training task.

[0134] Optionally, when the prediction task of the large trajectory model includes any one of a trajectory prediction task, a trajectory segmentation task, and an aircraft type identification task, the adjustment module 703 can be specifically used to: adjust the model parameters of the large trajectory model according to the prediction result and pre-set label data. The label data ends with an end indicator token, which is used to indicate the end of the prediction task.

[0135] Optionally, the device may further include: The first generation module is used to generate the input data and expected output data of the large trajectory model based on the actual trajectory point sequence of the aircraft.

[0136] The second generation module is used to concatenate the input data with the expected output data to generate the trajectory sample data.

[0137] The third generation module is used to concatenate the input data with the expected output data and mask the input data in the concatenation result to generate the label data.

[0138] Optionally, when the prediction task of the large trajectory model includes a trajectory prediction task, the first generation module may be specifically used to: randomly select a target trajectory point from the real trajectory point sequence; generate the input data based on the trajectory point sequence before the target trajectory point; and generate the expected output data based on the target trajectory point and the trajectory point sequence after the target trajectory point.

[0139] Optionally, when the prediction task of the large trajectory model includes a trajectory segmentation task, the first generation module may be specifically used to: mix the real trajectory point sequences of multiple aircraft to generate the input data; for the real trajectory point sequences of multiple aircraft, add a trajectory segmentation token to the end of the real trajectory point sequence of each aircraft to generate the expected output data; wherein, the trajectory segmentation token is used to segment the trajectories of different aircraft.

[0140] Optionally, the sequence of real waypoints includes the model information of the aircraft.

[0141] When the prediction task of the large trajectory model includes an aircraft type identification task, the first generation module can be specifically used to: remove the aircraft type information from the real trajectory point sequence to generate the input data; and use the real trajectory point sequence as the expected output data.

[0142] Optionally, the track sample data includes a sequence of mixed track points from multiple aircraft.

[0143] When the mission indication token indicates a track segmentation task, the training module 702 can be specifically used to: perform a track segmentation task on the track sample data using the large track model and according to the mission indication token, to obtain a track segmentation result. The track segmentation result includes at least one track segmentation token, which is used to segment the tracks of different aircraft.

[0144] The model training apparatus provided in this embodiment belongs to the same concept as the model training method provided in the above embodiments of this application. It can execute the model training method provided in any of the above embodiments of this application and has the corresponding functional modules and beneficial effects of the execution method. Technical details not described in detail in this embodiment can be found in the specific processing content of the model training method provided in the above embodiments of this application, and will not be repeated here.

[0145] Accordingly, this application also provides a model prediction device for use in a large trajectory model in a second electronic device. The large trajectory model is pre-trained using the model training method described above and is used to perform prediction tasks related to the aircraft based on input data.

[0146] The second electronic device can be a terminal device, such as a tablet computer, laptop computer, desktop computer, or mobile phone, or it can be a server, such as a cloud server. The second electronic device can be the same as or different from the first electronic device.

[0147] like Figure 8 As shown, the device may include: The receiving module 801 is used to receive track data.

[0148] The track data includes a sequence of waypoints for the aircraft and a mission indication token preceding the track point sequence. The mission indication token is used to indicate different prediction tasks for the aircraft. The different prediction tasks include at least two of the following: track prediction task, track segmentation task, and aircraft type identification task.

[0149] The prediction module 802 is used to perform a corresponding prediction task on the track data according to the task instruction token, and output the prediction result.

[0150] Optionally, the mission indication token indicates different predicted missions about the aircraft through different assignments to its included mission indication field.

[0151] The assignment of the task indication field includes at least two of the following: a first assignment for indicating the track prediction task, a second assignment for indicating the track segmentation task, and a third assignment for indicating the aircraft type identification task.

[0152] Optionally, the track data includes a sequence of mixed track points for multiple aircraft.

[0153] When the mission indication token indicates a track segmentation task, the prediction module 802 can be specifically used to: perform a track segmentation task on the track data according to the mission indication token, and obtain a track segmentation result. The track segmentation result includes at least one track segmentation token, which is used to segment the tracks of different aircraft.

[0154] Optionally, an end indicator token is provided at the end of the prediction result, which is used to indicate the end of the prediction result.

[0155] The model prediction apparatus provided in this embodiment belongs to the same concept as the model prediction method provided in the above embodiments of this application. It can execute the model prediction method provided in any of the above embodiments of this application and has the corresponding functional modules and beneficial effects of the execution method. Technical details not described in detail in this embodiment can be found in the specific processing content of the model prediction method provided in the above embodiments of this application, and will not be repeated here.

[0156] It should be understood that the modules in the above map-using device can be implemented by a processor calling software. For example, the device includes a processor connected to a memory containing instructions. The processor calls the instructions stored in the memory to implement any of the above methods or to implement the functions of each unit of the device. The processor can be a general-purpose processor, such as a CPU or microprocessor, and the memory can be internal or external to the device. Alternatively, the units in the device can be implemented as hardware circuits. By designing the hardware circuits, some or all of the unit functions can be implemented. The hardware circuit can be understood as one or more processors. For example, in one implementation, the hardware circuit is an ASIC, and the functions of some or all of the above units are implemented by designing the logical relationships between the components within the circuit. In another implementation, the hardware circuit can be implemented by a PLD, such as an FPGA, which can include a large number of logic gates. The connection relationships between the logic gates are configured through configuration files to implement the functions of some or all of the above units. All units of the above device can be implemented entirely by a processor calling software, entirely by hardware circuits, or partially by a processor calling software with the remaining parts implemented by hardware circuits.

[0157] In this application embodiment, a processor is a circuit with signal processing capabilities. In one implementation, the processor can be a circuit with instruction reading and execution capabilities, such as a CPU, microprocessor, GPU, or DSP. In another implementation, the processor can implement certain functions through the logical relationships of hardware circuits. These logical relationships are fixed or reconfigurable. For example, the processor may be a hardware circuit implemented as an ASIC or PLD, such as an FPGA. In a reconfigurable hardware circuit, the process of the processor loading a configuration document and configuring the hardware circuit can be understood as the processor loading instructions to implement the functions of some or all of the above units. Furthermore, it can also be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as an NPU, TPU, or DPU.

[0158] As can be seen, each unit in the above device can be one or more processors (or processing circuits) configured to implement the above methods, such as: CPU, GPU, NPU, TPU, DPU, microprocessor, DSP, ASIC, FPGA, or a combination of at least two of these processor forms.

[0159] Furthermore, the units in the above devices can be integrated in whole or in part, or they can be implemented independently. In one implementation, these units are integrated together and implemented in the form of a System-on-Chip (SoC). The SoC may include at least one processor for implementing any of the above methods or implementing the functions of the units in the device. The at least one processor may be of different types, such as CPU and FPGA, CPU and artificial intelligence processor, CPU and GPU, etc.

[0160] Exemplary electronic devices This application also provides an electronic device, such as... Figure 9 As shown, the electronic device includes a memory 900 and a processor 910.

[0161] The memory 900 is connected to the processor 910 and is used to store programs.

[0162] The processor 910 is used to implement the model training method or model prediction method in the above embodiments by running the program stored in the memory 900.

[0163] Specifically, the aforementioned electronic device may also include: a communication interface 920, an input device 930, an output device 940, and a bus 950.

[0164] The processor 910, memory 900, communication interface 920, input device 930, and output device 940 are interconnected via a bus. Among them: Bus 950 may include a pathway for transmitting information between various components of a computer system.

[0165] The processor 910 can be a general-purpose processor, such as a general-purpose central processing unit (CPU), a microprocessor, etc., or an application-specific integrated circuit (ASIC), or one or more integrated circuits used to control the execution of the program of the present invention. It can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0166] The processor 910 may include a main processor, as well as a baseband chip, modem, etc.

[0167] The memory 900 stores a program that executes the technical solution of this invention, and may also store an operating system and other key business functions. Specifically, the program may include program code, which includes computer operation instructions. More specifically, the memory 900 may include read-only memory (ROM), other types of static storage devices capable of storing static information and instructions, random access memory (RAM), other types of dynamic storage devices capable of storing information and instructions, disk storage, flash memory, etc.

[0168] Input device 930 may include a device for receiving user input data and information, such as a keyboard, mouse, camera, scanner, light pen, voice input device, touch screen, pedometer, or gravity sensor.

[0169] Output device 940 may include devices that allow information to be output to a user, such as a display screen, printer, speaker, etc.

[0170] The communication interface 920 may include a device that uses any transceiver to communicate with other devices or communication networks, such as Ethernet, Radio Access Network (RAN), Wireless Local Area Network (WLAN), etc.

[0171] The processor 910 executes the program stored in the memory 900 and calls other devices, which can be used to implement the various steps in the model training method or model prediction method provided in the above embodiments of this application.

[0172] Exemplary computer program products and storage media In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the model training method or model prediction method described in the embodiments of this application.

[0173] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0174] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0175] Furthermore, embodiments of this application may also be storage media storing computer programs, which are executed by a processor using the steps of the model training method or model prediction method described in the embodiments of this application.

[0176] Furthermore, embodiments of this application may also be chips, which include processors and data interfaces. The processor reads instructions stored in the memory through the data interface to execute the steps in the model training method or model prediction method described in the embodiments of this application.

[0177] For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0178] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0179] The steps in the methods of the various embodiments of this application can be adjusted, merged, or deleted in order according to actual needs, and the technical features described in each embodiment can be replaced or combined.

[0180] The modules and sub-modules in the devices and terminals in the various embodiments of this application can be merged, divided, and deleted according to actual needs.

[0181] It should be understood that the disclosed terminals, devices, and methods can be implemented in other ways, given the several embodiments provided in this application. For example, the terminal embodiments described above are merely illustrative. For instance, the division of modules or sub-modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple sub-modules or modules may be combined or integrated into another module, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0182] The modules or submodules described as separate components may or may not be physically separate. The components that constitute a module or submodule may or may not be physical modules or submodules; that is, they may be located in one place or distributed across multiple network modules or submodules. Some or all of the modules or submodules can be selected to achieve the purpose of this embodiment's solution, depending on actual needs.

[0183] Furthermore, the functional modules or sub-modules in the various embodiments of this application can be integrated into one processing module, or each module or sub-module can exist physically separately, or two or more modules or sub-modules can be integrated into one module. The integrated modules or sub-modules described above can be implemented in hardware or in the form of software functional modules or sub-modules.

[0184] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0185] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software unit executed by a processor, or a combination of both. The software unit can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0186] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A model training method, characterized in that, The method includes: Track sample data is input into a large-scale track model to be trained; wherein, the large-scale track model is a non-language model, and the large-scale track model is used to perform aircraft-related prediction tasks based on the input data; the track sample data includes a sequence of track points of the aircraft and a task indicator token located before the sequence of track points, the task indicator token is used to indicate different prediction tasks about the aircraft through different assignments of its included task indicator field; the different prediction tasks include at least two of the following: track prediction task, track segmentation task, and aircraft type recognition task; during the training process, the large-scale track model learns the mapping relationship between the task indicator token and the corresponding prediction task, and is used to distinguish different prediction tasks based on the task indicator token and execute the corresponding prediction task; Using the large trajectory model, based on the task indication token and the trajectory sample data, the prediction result corresponding to the prediction task indicated by the task indication token is obtained; Adjust the model parameters of the large trajectory model based on the prediction results.

2. The model training method according to claim 1, characterized in that, The task indicator token is used to indicate different prediction tasks, including pre-training tasks; The process of inputting track sample data into the track model to be trained includes: The first track sample data is input into the track large model; wherein, the task indicator token in the first track sample data is used to indicate the pre-training task; After completing the pre-training task, the second track sample data is input into the track large model; wherein, the task indicator token in the second track sample data is used to indicate the track prediction task, track segmentation task, or aircraft type identification task.

3. The model training method according to claim 2, characterized in that, The assignment of the task indication field includes at least two of the following: a first assignment for indicating a trajectory prediction task, a second assignment for indicating a trajectory segmentation task, a third assignment for indicating an aircraft type identification task, and a fourth assignment for indicating a pre-training task.

4. The model training method according to claim 1 or 2, characterized in that, When the prediction task of the large trajectory model includes any one of trajectory prediction, trajectory segmentation, and aircraft type identification, adjusting the model parameters of the large trajectory model based on the prediction results includes: Based on the prediction results and pre-set label data, the model parameters of the large trajectory model are adjusted; wherein, an end indicator token is set at the end of the label data, and the end indicator token is used to indicate the end of the prediction task.

5. The model training method according to claim 4, characterized in that, Before adjusting the model parameters of the large trajectory model based on the prediction results and pre-set label data, the method further includes: Based on the actual waypoint sequence of the aircraft, the input data and expected output data of the large-scale waypoint model are generated; The input data is concatenated with the expected output data to generate the trajectory sample data; The input data is concatenated with the expected output data, and the input data in the concatenation result is masked to generate the label data.

6. The model training method according to claim 5, characterized in that, When the prediction task of the large-scale trajectory model includes a trajectory prediction task, generating the input data and expected output data of the large-scale trajectory model based on the aircraft's actual waypoint sequence includes: Randomly select a target waypoint from the real waypoint sequence; The input data is generated based on the sequence of waypoints preceding the target waypoint, and the expected output data is generated based on the target waypoint and the sequence of waypoints following the target waypoint.

7. The model training method according to claim 5, characterized in that, When the prediction task of the large-scale trajectory model includes a trajectory segmentation task, generating the input data and expected output data of the large-scale trajectory model based on the aircraft's actual waypoint sequence includes: The input data is generated by mixing the real waypoint sequences of multiple aircraft. For a sequence of real waypoints of multiple aircraft, a waypoint segmentation token is added to the end of the sequence of real waypoints of each aircraft to generate the expected output data; wherein, the waypoint segmentation token is used to segment the waypoints of different aircraft.

8. The model training method according to claim 5, characterized in that, The real waypoint sequence includes the type information of the aircraft; When the prediction task of the large-scale trajectory model includes an aircraft type identification task, the step of generating the input data and expected output data of the large-scale trajectory model based on the aircraft's actual waypoint sequence includes: The aircraft type information in the real waypoint sequence is removed to generate the input data; The actual waypoint sequence is used as the expected output data.

9. The model training method according to claim 1 or 2, characterized in that, The track sample data includes a mixed track point sequence of multiple aircraft; When the task indication token indicates a track segmentation task, the step of obtaining the prediction result corresponding to the prediction task indicated by the task indication token through the track large model, based on the task indication token and the track sample data, includes: Using the large trajectory model, and based on the task instruction token, a trajectory segmentation task is performed on the trajectory sample data to obtain a trajectory segmentation result; wherein, the trajectory segmentation result includes at least one trajectory segmentation token, which is used to segment the trajectories of different aircraft.

10. A model prediction method, characterized in that, Applied to large trajectory models trained by the model training method according to any one of claims 1 to 9; The method includes: Receive track data; wherein the track data includes a sequence of waypoints of the aircraft and a task indication token located before the sequence of waypoints, the task indication token being used to indicate different prediction tasks about the aircraft through different assignments to its task indication field; the different prediction tasks include at least two of the following: track prediction task, track segmentation task, and aircraft type identification task; Based on the task instruction token, perform the corresponding prediction task on the track data and output the prediction result.

11. The model prediction method according to claim 10, characterized in that, The assignment of the task indication field includes at least two of the following: a first assignment for indicating the track prediction task, a second assignment for indicating the track segmentation task, and a third assignment for indicating the aircraft type identification task.

12. The model prediction method according to claim 10 or 11, characterized in that, The track data includes a mixed track point sequence of multiple aircraft; When the task indication token indicates a track segmentation task, the step of performing a corresponding prediction task on the track data according to the task indication token and outputting the prediction result includes: Based on the task instruction token, a track segmentation task is performed on the track data to obtain a track segmentation result; wherein, the track segmentation result includes at least one track segmentation token, which is used to segment the tracks of different aircraft.

13. The model prediction method according to claim 10, characterized in that, The prediction result is marked with an end indicator token, which is used to indicate the end of the prediction result.

14. An electronic device, characterized in that, include: Memory and processor; The memory is connected to the processor and is used to store programs; The processor is configured to implement the model training method as described in any one of claims 1 to 9 or the model prediction method as described in any one of claims 10 to 13 by running a program in the memory.

15. A computer program product, characterized in that, The computer program product stores a computer program, which, when executed by a processor, implements the model training method as described in any one of claims 1 to 9 or the model prediction method as described in any one of claims 10 to 13.

Images