A flight data processing method and system

Abstract

The application relates to the field of aviation technology, in particular to a flight data processing method and system. A flight data processing system comprises a data acquisition module, a flight strategy adjustment module and a flight strategy adjustment model updating module. The application determines the prediction uncertainty of airspace state in space-time analysis through random pruning full connection operation, and guides subsequent flight strategy adjustment according to the prediction uncertainty of airspace state in space-time analysis, so that the aircraft is in a safer environment when performing a flight task, and the safety of the aircraft is ensured. The space-time propagation of the uncertainty of airspace state prediction is analyzed, the subsequent flight strategy adjustment is further guided, and the explainability of airspace state prediction is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aviation technology, and particularly relates to a flight data processing method and system. BACKGROUND

[0002] With the rapid development of aviation technology, the application of aircraft is increasingly widespread, and the tasks performed by the aircraft are becoming more and more complex. Flight data, as a direct reflection of the running state of the aircraft, contains rich flight information, which is of great significance to flight safety, efficiency improvement and airspace management. Effective processing and analysis of flight data, especially airspace state prediction, are the current research focus in the field of aviation.

[0003] The existing airspace state prediction is generally time series prediction, which does not consider the uncertainty brought by spatial dimension change and temporal dimension change, especially in complex weather environment, which leads to safety accidents of the aircraft when performing flight tasks. SUMMARY

[0004] The present application determines the prediction uncertainty of airspace state in spatio-temporal analysis through random pruning full connection operation, and guides the subsequent flight strategy adjustment according to the prediction uncertainty of airspace state in spatio-temporal analysis, so that the aircraft is in a safer environment when performing flight tasks, ensuring the safety of the aircraft. And the spatio-temporal propagation of the uncertainty of airspace state prediction is analyzed, further guiding the subsequent flight strategy adjustment, and enhancing the explainability of airspace state prediction.

[0005] The present application provides a flight data processing method, comprising:

[0006] Obtaining the position data and flight data set of the aircraft at the current monitoring time point, the flight data set comprising an airspace state data set and an aircraft attitude data set, the position data and the airspace state data set at the current monitoring time point being combined to form an airspace state position labeled data set, and the airspace state position labeled data set at the current monitoring time point and the airspace state position labeled data sets at the previous N-1 monitoring time points being combined to form an airspace state spatio-temporal analysis data set at the monitoring time point, and then the airspace state spatio-temporal analysis data set and the aircraft attitude data set at the current monitoring time point being sent to a flight strategy adjustment model for processing to output a flight strategy adjustment data set;

[0007] The flight strategy adjustment model includes an airspace state prediction layer, an uncertainty analysis layer, an uncertainty spatiotemporal propagation layer, and a flight strategy adjustment layer. The airspace state prediction layer includes a Transformer unit and a random pruning fully connected unit. The Transformer unit is used to perform time series analysis on the airspace state spatiotemporal analysis dataset to output the airspace state spatiotemporal analysis vector. The random pruning fully connected unit is used to perform random pruning fully connected operations on the airspace state spatiotemporal analysis vector output by the Transformer unit to output the airspace state prediction vector. The uncertainty analysis layer is used to perform random pruning fully connected operations on the airspace state prediction vectors output by several airspace state prediction layers. Uncertainty analysis is performed to construct an uncertainty vector and a reference airspace state prediction vector; the uncertainty spatiotemporal propagation layer is used to perform time series analysis on the uncertainty vector corresponding to the current monitoring time point and the uncertainty vectors corresponding to the previous N-1 monitoring time points to output the uncertainty spatiotemporal propagation vector; the flight strategy adjustment layer is used to splice the aircraft attitude dataset, the reference airspace state prediction vector, the uncertainty vector and the uncertainty spatiotemporal propagation vector at the current monitoring time point and send them to the multi-layer perceptron for processing to output the flight strategy adjustment dataset; during the use of the flight strategy adjustment model, the flight strategy adjustment model is updated in real time through reinforcement learning.

[0008] Preferably, the random pruning fully connected operation specifically includes the following steps:

[0009] In the process of performing the fully connected operation, whenever any neuron node outputs, a random number δ between (0, 1) is generated by a random function, and then the random number δ is compared with the discard probability P. If the random number δ is greater than the discard probability P, the output of the neuron node is set to zero. If the random number δ is not greater than the discard probability P, the zeroing operation is not performed.

[0010] Preferably, uncertainty analysis is performed on the spatial state prediction vectors output by the multiple spatial state prediction layers through an uncertainty analysis layer to construct an uncertainty vector and a reference spatial state prediction vector, specifically comprising the following steps:

[0011] The operation of sending the spatial state spatiotemporal analysis dataset to the spatial state prediction layer for processing is performed several times to obtain several spatial state prediction vectors;

[0012] A weighted average operation is performed on all airspace state prediction vectors to construct a baseline airspace state prediction vector; then an uncertainty vector is calculated based on all airspace state prediction vectors. The uncertainty vector stores the uncertainty value Fi, i=1, 2, 3, ..., I, where I is the total number of uncertainty values, which is consistent with the total number of data in the airspace state dataset. The uncertainty value Fi is the variance of the i-th data in all airspace state prediction vectors, which is used to characterize the uncertainty of the prediction of the i-th data in the airspace state dataset.

[0013] Preferably, during the use of the flight strategy adjustment model, the flight strategy adjustment model is updated in real time by reinforcement learning, specifically including the following steps:

[0014] At the monitoring time point, the airspace status spatiotemporal analysis dataset and the aircraft attitude dataset at the current monitoring time point are sent to the target flight strategy adjustment model for processing, and the target flight strategy adjustment dataset is output. In the initial state, the flight strategy adjustment model is consistent with the target flight strategy adjustment model. Then, the airspace status spatiotemporal analysis dataset, the aircraft attitude dataset at the current monitoring time point and the target flight strategy adjustment dataset are spliced ​​to construct reward value analysis data. The reward value analysis data is sent to the reward value analysis network for processing, and the reward value data is output. The gradient value of the reward value data for the target flight strategy adjustment dataset is calculated, and then the gradient ascent method is used based on the gradient value to adjust the parameters of the flight strategy adjustment model to achieve real-time update of the flight strategy adjustment model; and the target flight strategy adjustment model is also updated in real time.

[0015] Preferably, the target flight strategy adjustment model is updated in real time, specifically comprising the following steps:

[0016] Whenever the flight strategy adjustment model is updated, the target flight strategy adjustment model is updated in real time using the following formula: θk(new)=δηk+θk(bef), where k=1, 2, 3,…, K, K is the total number of parameter values ​​in the target flight strategy adjustment model, θk(new) is the kth parameter value in the updated target flight strategy adjustment model, δ is the update rate, which is generally 0.01, ηk is the kth parameter value in the updated flight strategy adjustment model, and θk(bef) is the kth parameter value in the target flight strategy adjustment model before the update.

[0017] Preferably, training the flight strategy adjustment model specifically includes the following steps:

[0018] Obtain several flight strategy adjustment training samples, which include an airspace status spatiotemporal analysis dataset and an aircraft attitude dataset at corresponding monitoring time points. Label the flight strategy adjustment training samples using the flight strategy adjustment dataset, and form a flight strategy adjustment training set with all labeled flight strategy adjustment training samples. Train the flight strategy adjustment model using the flight strategy adjustment training set to determine whether the training conditions are met. If so, output the trained flight strategy adjustment model; otherwise, continue training the flight strategy adjustment model using the flight strategy adjustment training set.

[0019] Preferably, training the reward value analysis network specifically includes the following steps:

[0020] Obtain several positive training samples and negative training samples for reward value analysis. The positive training samples for reward value analysis include a spatiotemporal analysis dataset of airspace status, a dataset of aircraft attitudes at corresponding monitoring time points, and a dataset of optimal flight strategy adjustment. The negative training samples for reward value analysis include a spatiotemporal analysis dataset of airspace status, a dataset of aircraft attitudes at corresponding monitoring time points, and a dataset of the worst flight strategy adjustment. Label the positive training samples and negative training samples for reward value analysis using reward value data. Form a reward value analysis training set with all labeled positive training samples and negative training samples for reward value analysis. Train a reward value analysis network using the reward value analysis training set. Use the reward value data as the target output, calculate the loss value, and determine whether the loss value is within a preset range. If the loss value is within the preset range, output the trained reward value analysis network. Otherwise, continue training the reward value analysis network using the reward value analysis training set.

[0021] The present invention also provides a flight data processing system, comprising:

[0022] The data acquisition module is used to obtain the position data and flight data set of the aircraft at the current monitoring time point. The flight data set includes the airspace status data set and the aircraft attitude data set;

[0023] The flight strategy adjustment module is used to combine the position data of the current monitoring time point and the airspace status data set into an airspace status position annotation data set, and combine the airspace status position annotation data set of the current monitoring time point and the airspace status position annotation data sets of the previous N-1 monitoring time points into an airspace status spatiotemporal analysis data set of the monitoring time point. The airspace status spatiotemporal analysis data set and the aircraft attitude data set of the current monitoring time point are then fed into the flight strategy adjustment model for processing to output a flight strategy adjustment data set.

[0024] The flight strategy adjustment model includes an airspace state prediction layer, an uncertainty analysis layer, an uncertainty spatiotemporal propagation layer and a flight strategy adjustment layer, wherein the airspace state prediction layer includes a Transformer unit and a random pruning fully connected unit, wherein the Transformer unit is used to perform time series analysis on the airspace state spatiotemporal analysis dataset to output a airspace state spatiotemporal analysis vector, and the random pruning fully connected unit is used to perform a random pruning fully connected operation on the airspace state spatiotemporal analysis vector output by the Transformer unit to output a airspace state prediction vector; the uncertainty analysis layer is used to perform uncertainty analysis on the airspace state prediction vectors output by several airspace state prediction layers to construct an uncertainty vector and a baseline airspace state prediction vector; the uncertainty spatiotemporal propagation layer is used to perform time series analysis on the uncertainty vector corresponding to the current monitoring time point and the uncertainty vectors corresponding to the previous N-1 monitoring time points to output an uncertainty spatiotemporal propagation vector; the flight strategy adjustment layer is used to splice the aircraft attitude dataset at the current monitoring time point, the baseline airspace state prediction vector, the uncertainty vector and the uncertainty spatiotemporal propagation vector and send them into a multilayer perceptron for processing to output a flight strategy adjustment dataset;

[0025] The flight strategy adjustment model update module is used to update the flight strategy adjustment model in real time through reinforcement learning during the use of the flight strategy adjustment model.

[0026] The present invention has the following advantages:

[0027] The present invention uses a fully connected operation of random pruning to determine the prediction uncertainty of the airspace state in the spatiotemporal analysis, and guides subsequent flight strategy adjustments based on the prediction uncertainty of the airspace state in the spatiotemporal analysis, so that the aircraft is in a safer environment when performing flight missions, thereby ensuring the safety of the aircraft; and analyzes the spatiotemporal propagation of the uncertainty of the airspace state prediction, further guiding subsequent flight strategy adjustments and enhancing the interpretability of the airspace state prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic structural diagram of a flight data processing system used in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0030] Embodiment 1, a flight data processing method, comprising:

[0031] Obtain the position data and flight dataset of the aircraft at the current monitoring time point. The flight dataset includes an airspace state dataset and an aircraft attitude dataset. The airspace state dataset here is environmental data of the aircraft's location at the current monitoring time point, such as wind speed, visibility, and turbulence intensity. These environmental data will affect the aircraft's execution of the flight mission. The aircraft attitude dataset includes the attitude data of the aircraft itself, such as heading angle and flight speed. The position data and the airspace state dataset at the current monitoring time point are combined into an airspace state position annotation dataset, and the airspace state position annotation dataset at the current monitoring time point and the airspace state position annotation datasets at the previous N-1 monitoring time points are combined into an airspace state spatiotemporal analysis dataset at the monitoring time point. The airspace state spatiotemporal analysis dataset and the aircraft attitude dataset at the current monitoring time point are then fed into a flight strategy adjustment model for processing to output a flight strategy adjustment dataset. The flight strategy adjustment dataset corresponds to the aircraft attitude dataset, such as a heading adjustment strategy and a flight speed adjustment strategy. The aircraft is controlled based on the aircraft attitude dataset at the current monitoring time point and the flight strategy adjustment dataset output by the flight strategy adjustment model to achieve adaptive flight state adjustment of the aircraft to the environment.

[0032] The flight strategy adjustment model includes an airspace state prediction layer, an uncertainty analysis layer, an uncertainty spatiotemporal propagation layer, and a flight strategy adjustment layer. The airspace state prediction layer includes a Transformer unit and a random pruned fully connected unit. The Transformer unit is established based on the Transformer model and is used to perform time series analysis on the airspace state spatiotemporal analysis dataset to output the airspace state spatiotemporal analysis vector. Since the airspace state position labeling dataset in the airspace state spatiotemporal analysis dataset includes position data, the spatial changes can also be analyzed during the time series analysis. The random pruned fully connected unit is used to The spatial state spatiotemporal analysis vector output by the Transformer unit is randomly pruned and fully connected. The random pruning fully connected operation here refers to randomly discarding some neuron nodes based on the discard probability during the full connection operation, that is, setting the output of the discarded neuron nodes to zero to output the spatial state prediction vector; the uncertainty analysis layer is used to perform uncertainty analysis on the spatial state prediction vector output by several spatial state prediction layers to construct an uncertainty vector and a reference spatial state prediction vector. Since the space dimension and time dimension of the aircraft will change during the execution of the flight mission, the conventional Time series analysis cannot make accurate predictions. Therefore, the uncertainty of airspace state prediction is introduced. Based on the uncertainty of airspace state prediction, the aircraft strategy is more reasonably controlled to ensure the safety of the aircraft during flight. If the prediction uncertainty is high, the aircraft should output a more conservative flight strategy, such as a safer flight route and a lower flight speed; the uncertainty space-time propagation layer is used to perform time series analysis on the uncertainty vector corresponding to the current monitoring time point and the uncertainty vector corresponding to the previous N-1 monitoring time points to output the uncertainty space-time propagation vector. It should be noted that the uncertainty vectors of N monitoring time points are subjected to time series analysis. , which can analyze the characteristics of the uncertainty propagation of airspace state prediction in time and space. For example, the uncertain change of wind speed will affect the subsequent wind speed prediction along the time dimension, and affect the surrounding wind speed prediction along the spatial dimension, thereby more accurately guiding the subsequent aircraft strategy adjustment. The flight strategy adjustment layer is used to splice the aircraft attitude dataset, the reference airspace state prediction vector, the uncertainty vector and the uncertainty spatiotemporal propagation vector at the current monitoring time point and send them to the multi-layer perceptron for processing to output the flight strategy adjustment dataset. During the use of the flight strategy adjustment model, the flight strategy adjustment model is updated in real time through reinforcement learning.

[0033] The application determines the prediction uncertainty of the space state in the spatio-temporal analysis through the random pruning full connection operation, and guides the subsequent flight strategy adjustment according to the prediction uncertainty of the space state in the spatio-temporal analysis, so that the aircraft is in a safer environment when performing a flight task, and the safety of the aircraft is ensured. The spatio-temporal propagation of the uncertainty of the space state prediction is analyzed, and the subsequent flight strategy adjustment is further guided, and the explainability of the space state prediction is enhanced.

[0034] The random pruning full connection operation specifically includes the following steps:

[0035] In the process of performing the full connection operation, whenever any neuron node outputs, a random number δ between 0 and 1 is generated through a random function, and then the random number δ is compared with the dropout probability P. The dropout probability P is set by the developer and is generally 0.5. If the random number δ is greater than the dropout probability P, the output of the neuron node is set to zero. If the random number δ is not greater than the dropout probability P, no zero setting operation is performed. It should be noted that the full connection operation is generally performed through a plurality of full connection layers, and each full connection layer is composed of a plurality of neuron nodes.

[0036] The uncertainty analysis layer performs uncertainty analysis on the space state prediction vectors output by the plurality of space state prediction layers to construct an uncertainty vector and a benchmark space state prediction vector, specifically including the following steps:

[0037] The space state spatio-temporal analysis data set is sent into the space state prediction layer for processing for a plurality of times, which is 100 times in this embodiment. A plurality of space state prediction vectors are obtained. It should be noted that, due to the random pruning full connection operation, under the condition that the spatio-temporal change is stable, these space state prediction vectors should be consistent. If these space state prediction vectors are inconsistent, it indicates that the accuracy of the time series prediction is low, i.e., the prediction uncertainty.

[0038] All the space state prediction vectors are subjected to a weighted average operation to construct a benchmark space state prediction vector. Then, based on all the space state prediction vectors, an uncertainty vector is calculated. The uncertainty vector stores uncertainty values Fi, i=1, 2, 3, …, I, I being the total number of uncertainty values, which is consistent with the total number of data in the space state data set. The uncertainty value Fi is the variance of the i-th data in all the space state prediction vectors, which is used to represent the prediction uncertainty of the i-th data in the space state data set.

[0039] In the use process of the flight strategy adjustment model, the flight strategy adjustment model is updated in real time through reinforcement learning, specifically including the following steps:

[0040] At the monitoring time point, the airspace state spatiotemporal analysis dataset and the aircraft attitude dataset at the current monitoring time point are fed into the target flight strategy adjustment model for processing, and the target flight strategy adjustment dataset is output. In the initial state, the flight strategy adjustment model is consistent with the target flight strategy adjustment model. Then, the airspace state spatiotemporal analysis dataset, the aircraft attitude dataset at the current monitoring time point, and the target flight strategy adjustment dataset are spliced ​​to construct reward value analysis data. The reward value analysis data is fed into the reward value analysis network for processing. The reward value analysis network is established based on the BP neural network model and outputs reward value data. The reward value data can reflect the safety level of the aircraft after executing the flight strategy adjustment dataset. The gradient value of the reward value data with respect to the target flight strategy adjustment dataset is calculated, and then the parameters of the flight strategy adjustment model are adjusted using the gradient ascent method based on the gradient value, thereby achieving real-time update of the flight strategy adjustment model. It should be noted that performing parameter adjustment using the target flight strategy adjustment dataset output by the target flight strategy adjustment model can ensure that the flight strategy adjustment model is stably updated under the premise that the target flight strategy adjustment model is used as a baseline; and the target flight strategy adjustment model is also updated in real time.

[0041] The model is updated in real time based on the target flight strategy adjustment, which includes the following steps:

[0042] Whenever the flight strategy adjustment model is updated, the target flight strategy adjustment model is updated in real time using the following formula: θ k (new) = δη k +θ k (bef), where k = 1, 2, 3, ..., K, K is the total number of parameter values ​​in the target flight strategy adjustment model, θ k (new) is the kth parameter value in the updated target flight strategy adjustment model, δ is the update rate, which is generally 0.01, and η k Adjust the kth parameter value in the model for the updated flight strategy, θ k (bef) is the kth parameter value in the target flight strategy adjustment model before update;

[0043] Training the flight strategy adjustment model includes the following steps:

[0044] Several flight strategy adjustment training samples are obtained. These include a spatiotemporal analysis dataset of airspace status and an aircraft attitude dataset at corresponding monitoring time points. It should be noted that these spatiotemporal analysis datasets and aircraft attitude datasets at corresponding monitoring time points are obtained by developers based on aircraft simulation flight tests. Aircraft simulation flight tests are conducted using flight simulation software (such as Simulink) and employ a probabilistic random setting of the simulation environment. The flight strategy adjustment training samples are labeled using the flight strategy adjustment dataset. It should be noted that the labeled flight strategy adjustment dataset represents the optimal setting obtained by developers based on expert experience in a random environment. The aircraft simulation flight tests can serve as a sample source for transfer training. In actual use, further training is generally performed based on flight strategy adjustment training samples corresponding to the actual aircraft's operating environment. All labeled flight strategy adjustment training samples form a flight strategy adjustment training set. The flight strategy adjustment model is trained using the flight strategy adjustment training set to determine whether the training conditions are met. The training conditions are generally a certain number of training runs. If the training conditions are met, the trained flight strategy adjustment model is output; otherwise, the flight strategy adjustment model is further trained using the flight strategy adjustment training set.

[0045] Training the reward value analysis network includes the following steps:

[0046] Obtain several positive training samples and negative training samples for reward value analysis. The positive training samples for reward value analysis include the airspace status spatiotemporal analysis dataset, the aircraft attitude dataset at the corresponding monitoring time point, and the optimal flight strategy adjustment dataset. The negative training samples for reward value analysis include the airspace status spatiotemporal analysis dataset, the aircraft attitude dataset at the corresponding monitoring time point, and the worst flight strategy adjustment dataset. The worst flight strategy adjustment dataset here is obtained from an aircraft that has a safety accident in an aircraft simulation flight test. The positive training samples and negative training samples for reward value analysis are labeled using reward value data. Generally, the reward value data corresponding to the positive training sample of reward value analysis is 1, and the reward value data corresponding to the negative training sample of reward value analysis is 0. All labeled positive training samples and negative training samples of reward value analysis form a reward value analysis training set. The reward value analysis network is trained through the reward value analysis training set. The reward value data is used as the target output, and the loss value is calculated to determine whether the loss value is within the preset range. The preset range is set by the developer based on experience. If the loss value is within the preset range, the trained reward value analysis network is output; otherwise, the reward value analysis network is continued to be trained through the reward value analysis training set.

[0047] Example 2, a flight data processing system, such as Figure 1 Shown, including:

[0048] The data acquisition module is used to obtain the aircraft's position data and flight data set at the current monitoring time. The flight data set includes an airspace status data set and an aircraft attitude data set. The airspace status data set here refers to the environmental data of the aircraft's location at the current monitoring time, such as wind speed, visibility, and turbulence intensity. These environmental data will affect the aircraft's flight mission. The aircraft attitude data set includes the aircraft's own attitude data, such as heading angle and flight speed.

[0049] The flight strategy adjustment module is used to combine the airspace status position annotation dataset at the current monitoring time point and the airspace status position annotation datasets at the previous N-1 monitoring time points into an airspace status spatiotemporal analysis dataset at the monitoring time point, and then send the airspace status spatiotemporal analysis dataset and the aircraft attitude dataset at the current monitoring time point into the flight strategy adjustment model for processing to output a flight strategy adjustment dataset. The flight strategy adjustment dataset corresponds to the aircraft attitude dataset, such as the heading adjustment strategy and the flight speed adjustment strategy. The aircraft is controlled based on the aircraft attitude dataset at the current monitoring time point and the flight strategy adjustment dataset output by the flight strategy adjustment model to achieve adaptive flight state adjustment of the aircraft to the environment.

[0050] The flight strategy adjustment model includes an airspace state prediction layer, an uncertainty analysis layer, an uncertainty spatiotemporal propagation layer and a flight strategy adjustment layer, wherein the airspace state prediction layer includes a Transformer unit and a random pruning fully connected unit, wherein the Transformer unit is established based on the Transformer model, and is used to perform time series analysis on the airspace state spatiotemporal analysis dataset to output the airspace state spatiotemporal analysis vector. Since the airspace state position labeling dataset in the airspace state spatiotemporal analysis dataset includes position data, the spatial changes can also be analyzed during the time series analysis. The random pruning fully connected unit is used to perform a random pruning fully connected operation on the airspace state spatiotemporal analysis vector output by the Transformer unit. The random pruning fully connected operation here refers to randomly discarding some neuron nodes based on the discard probability during the full connection operation, that is, setting the output of the discarded neuron nodes to zero to output the airspace state prediction vector; the uncertainty analysis layer is used to perform uncertainty analysis on the airspace state prediction vector output by several airspace state prediction layers to construct an uncertainty vector and a baseline airspace state prediction vector. Since the aircraft is in the process of performing a flight mission, the aircraft Both the spatial and temporal dimensions will change, and conventional time series analysis cannot accurately predict. Therefore, the uncertainty of airspace state prediction is introduced. Based on the uncertainty of airspace state prediction, the aircraft strategy is more reasonably controlled to ensure the safety of the aircraft during flight. If the prediction uncertainty is high, the aircraft should output a more conservative flight strategy, such as a safer flight route and a lower flight speed. The uncertainty spatiotemporal propagation layer is used to perform time series analysis on the uncertainty vector corresponding to the current monitoring time point and the uncertainty vectors corresponding to the previous N-1 monitoring time points to output the uncertainty spatiotemporal propagation vector. It should be noted that the time series analysis of the uncertainty vectors of N monitoring time points can analyze the spatiotemporal propagation characteristics of the uncertainty of airspace state prediction. For example, the uncertain change of wind speed will affect the subsequent wind speed prediction along the time dimension, and affect the surrounding wind speed prediction along the spatial dimension, thereby more accurately guiding the subsequent aircraft strategy adjustment. The flight strategy adjustment layer is used to splice the aircraft attitude dataset of the current monitoring time point, the baseline airspace state prediction vector, the uncertainty vector and the uncertainty spatiotemporal propagation vector, and then send them to the multi-layer perceptron for processing to output the flight strategy adjustment dataset.

[0051] The flight strategy adjustment model update module is used to update the flight strategy adjustment model in real time through reinforcement learning during the use of the flight strategy adjustment model.

[0052] It should be understood that those skilled in the art may make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims. Any portion of this specification not described in detail is prior art known to those skilled in the art.

Claims

1. A flight data processing method, characterized in that: include: Obtain the position data and flight dataset of the aircraft at the current monitoring time point, where the flight dataset includes an airspace status dataset and an aircraft attitude dataset. The position data and the airspace status dataset at the current monitoring time point are combined into an airspace status position annotation dataset. The airspace status position annotation dataset at the current monitoring time point and the airspace status position annotation datasets at the previous N-1 monitoring time points are combined into an airspace status spatiotemporal analysis dataset at the monitoring time point. The airspace status spatiotemporal analysis dataset and the aircraft attitude dataset at the current monitoring time point are then fed into a flight strategy adjustment model for processing to output a flight strategy adjustment dataset. The flight strategy adjustment model includes an airspace state prediction layer, an uncertainty analysis layer, an uncertainty spatiotemporal propagation layer, and a flight strategy adjustment layer. The airspace state prediction layer includes a Transformer unit and a random pruning fully connected unit. The Transformer unit is used to perform time series analysis on the airspace state spatiotemporal analysis dataset to output the airspace state spatiotemporal analysis vector. The random pruning fully connected unit is used to perform random pruning fully connected operations on the airspace state spatiotemporal analysis vector output by the Transformer unit to output the airspace state prediction vector. The uncertainty analysis layer is used to perform random pruning fully connected operations on the airspace state prediction vectors output by several airspace state prediction layers. Uncertainty analysis is performed to construct an uncertainty vector and a baseline airspace state prediction vector. The uncertainty spatiotemporal propagation layer is used to perform time series analysis on the uncertainty vector corresponding to the current monitoring time point and the uncertainty vectors corresponding to the previous N-1 monitoring time points to output the uncertainty spatiotemporal propagation vector. The flight strategy adjustment layer is used to concatenate the aircraft attitude dataset, baseline airspace state prediction vector, uncertainty vector, and uncertainty spatiotemporal propagation vector at the current monitoring time point and send them to the multi-layer perceptron for processing to output the flight strategy adjustment dataset. During the use of the flight strategy adjustment model, the flight strategy adjustment model is updated in real time through reinforcement learning. The random pruning fully connected operation includes the following steps: In the process of executing the full connection operation, whenever any neuron node outputs, a random number δ between (0, 1) is generated by a random function, and then the random number δ is compared with the discard probability P. If the random number δ is greater than the discard probability P, the output of the neuron node is set to zero. If the random number δ is not greater than the discard probability P, the zeroing operation is not performed. The uncertainty analysis layer performs uncertainty analysis on the airspace state prediction vectors output by several airspace state prediction layers to construct uncertainty vectors and reference airspace state prediction vectors. The specific steps include the following: The operation of sending the spatial state spatiotemporal analysis data set to the spatial state prediction layer for processing is performed several times to obtain several spatial state prediction vectors; A weighted average operation is performed on all airspace state prediction vectors to construct a baseline airspace state prediction vector; then an uncertainty vector is calculated based on all airspace state prediction vectors. The uncertainty vector stores the uncertainty value Fi, i=1, 2, 3, ..., I, where I is the total number of uncertainty values, which is consistent with the total number of data in the airspace state dataset. The uncertainty value Fi is the variance of the i-th data in all airspace state prediction vectors, which is used to characterize the uncertainty of the prediction of the i-th data in the airspace state dataset.

2. A flight data processing method according to claim 1, characterized in that: During the use of the flight strategy adjustment model, the flight strategy adjustment model is updated in real time through reinforcement learning. The specific steps include: At the monitoring time point, the airspace status spatiotemporal analysis dataset and the aircraft attitude dataset at the current monitoring time point are sent to the target flight strategy adjustment model for processing, and the target flight strategy adjustment dataset is output. In the initial state, the flight strategy adjustment model is consistent with the target flight strategy adjustment model. Then, the airspace status spatiotemporal analysis dataset, the aircraft attitude dataset at the current monitoring time point and the target flight strategy adjustment dataset are spliced ​​to construct reward value analysis data. The reward value analysis data is sent to the reward value analysis network for processing, and the reward value data is output. The gradient value of the reward value data for the target flight strategy adjustment dataset is calculated, and then the gradient ascent method is used based on the gradient value to adjust the parameters of the flight strategy adjustment model to achieve real-time update of the flight strategy adjustment model; and the target flight strategy adjustment model is also updated in real time.

3. The flight data processing method according to claim 2, characterized in that: The model is updated in real time based on the target flight strategy adjustment, which includes the following steps: Whenever the flight strategy adjustment model is updated, the target flight strategy adjustment model is updated in real time using the following formula: θ k (new) = δη k +θ k (bef), where k = 1, 2, 3, ..., K, K is the total number of parameter values ​​in the target flight strategy adjustment model, θ k (new) is the kth parameter value in the updated target flight strategy adjustment model, δ is the update rate, η k Adjust the kth parameter value in the model for the updated flight strategy, θ k (bef) The kth parameter value in the target flight strategy adjustment model before update.

4. The flight data processing method according to claim 3, characterized in that: Training the flight strategy adjustment model includes the following steps: Obtain several flight strategy adjustment training samples, which include an airspace status spatiotemporal analysis dataset and an aircraft attitude dataset at corresponding monitoring time points. Label the flight strategy adjustment training samples using the flight strategy adjustment dataset, and form a flight strategy adjustment training set with all labeled flight strategy adjustment training samples. Train the flight strategy adjustment model using the flight strategy adjustment training set to determine whether the training conditions are met. If so, output the trained flight strategy adjustment model; otherwise, continue training the flight strategy adjustment model using the flight strategy adjustment training set.

5. The flight data processing method according to claim 4, characterized in that: Training the reward value analysis network includes the following steps: Obtain several positive training samples and negative training samples for reward value analysis. The positive training samples for reward value analysis include a spatiotemporal analysis dataset of airspace status, a dataset of aircraft attitudes at corresponding monitoring time points, and a dataset of optimal flight strategy adjustment. The negative training samples for reward value analysis include a spatiotemporal analysis dataset of airspace status, a dataset of aircraft attitudes at corresponding monitoring time points, and a dataset of the worst flight strategy adjustment. Label the positive training samples and negative training samples for reward value analysis using reward value data. Form a reward value analysis training set with all labeled positive training samples and negative training samples for reward value analysis. Train a reward value analysis network using the reward value analysis training set. Use the reward value data as the target output, calculate the loss value, and determine whether the loss value is within a preset range. If the loss value is within the preset range, output the trained reward value analysis network. Otherwise, continue training the reward value analysis network using the reward value analysis training set.

6. A flight data processing system, characterized in that: The system applies a flight data processing method according to any one of claims 1 to 5, including: The data acquisition module is used to obtain the position data and flight data set of the aircraft at the current monitoring time point. The flight data set includes the airspace status data set and the aircraft attitude data set; The flight strategy adjustment module is used to combine the position data of the current monitoring time point and the airspace status data set into an airspace status position annotation data set, and combine the airspace status position annotation data set of the current monitoring time point and the airspace status position annotation data sets of the previous N-1 monitoring time points into an airspace status spatiotemporal analysis data set of the monitoring time point. The airspace status spatiotemporal analysis data set and the aircraft attitude data set of the current monitoring time point are then fed into the flight strategy adjustment model for processing to output a flight strategy adjustment data set. The flight strategy adjustment model includes an airspace state prediction layer, an uncertainty analysis layer, an uncertainty spatiotemporal propagation layer and a flight strategy adjustment layer, wherein the airspace state prediction layer includes a Transformer unit and a random pruning fully connected unit, wherein the Transformer unit is used to perform time series analysis on the airspace state spatiotemporal analysis dataset to output a airspace state spatiotemporal analysis vector, and the random pruning fully connected unit is used to perform a random pruning fully connected operation on the airspace state spatiotemporal analysis vector output by the Transformer unit to output a airspace state prediction vector; the uncertainty analysis layer is used to perform uncertainty analysis on the airspace state prediction vectors output by several airspace state prediction layers to construct an uncertainty vector and a baseline airspace state prediction vector; the uncertainty spatiotemporal propagation layer is used to perform time series analysis on the uncertainty vector corresponding to the current monitoring time point and the uncertainty vectors corresponding to the previous N-1 monitoring time points to output an uncertainty spatiotemporal propagation vector; the flight strategy adjustment layer is used to splice the aircraft attitude dataset at the current monitoring time point, the baseline airspace state prediction vector, the uncertainty vector and the uncertainty spatiotemporal propagation vector and send them into a multilayer perceptron for processing to output a flight strategy adjustment dataset; The flight strategy adjustment model update module is used to update the flight strategy adjustment model in real time through reinforcement learning during the use of the flight strategy adjustment model.

Images