Flight operation prompting method and system

Abstract

The invention discloses a flight operation prompting method and system, and relates to the technical field of flight auxiliary decision making. Comprising the following steps: acquiring offline data of an aircraft, taking a flight state of the aircraft at any moment as an initial state, determining a gradient corresponding to the initial state, and presetting a strategy network; based on the initial state, selecting a flight operation from the strategy network, determining the flight state of the aircraft at the next moment according to the adopted action, and obtaining the flight state of each historical moment in the preset time period through multi-round iteration; in each round of iteration, the stochastic gradient of the selected flight operation is determined; training the reinforcement learning model to obtain an auxiliary decision-making model; and inputting the flight state of the aircraft at the current moment into the auxiliary decision-making model to obtain an optimal flight operation, and generating an optimal flight operation prompt. According to the method, redundant information can be removed, pilot information overload can be relieved, a pilot can be assisted in fast decision making, and decision making efficiency and operation accuracy are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of flight assistance decision-making technology, and in particular to a flight operation prompting method and system. Background Technology

[0002] Flight operation prompting technology is a technology that utilizes advanced computer science, data analytics, and artificial intelligence to assist pilots in making decisions in complex aviation environments. The goal of this technology is to provide real-time, intelligent, and personalized operational prompts to ensure the safety, efficiency, and successful completion of flight missions.

[0003] In the aviation field, traditional flight operation prompt technology primarily relies on predetermined rules combined with subjective judgment. Before flying, pilots need to thoroughly study and understand the predetermined rules integrated into the flight management system. The flight management system collects flight data and external environmental information in real time through onboard sensors. Based on the collected flight data and predetermined rules, the flight management system automatically performs logical reasoning and calculations to generate corresponding operation prompts.

[0004] Traditional flight operation prompting technology fails to provide efficient information filtering and integration, leaving pilots facing information overload and reduced decision-making efficiency, resulting in insufficient accuracy of the generated operation prompts. Summary of the Invention

[0005] Therefore, it is necessary to provide a flight operation prompting method and system to address the aforementioned technical problems.

[0006] This invention provides a flight operation prompting method, including: Acquire offline data of the aircraft, including: the aircraft's historical flight state and corresponding flight operations at multiple historical moments; take the aircraft's historical flight state at any historical moment as the initial state and determine the gradient of the initial state, the gradient of the initial state represents the initial reward; preset a policy network composed of multiple actions, the actions being flight operations. Based on the initial state, a flight operation is selected from the policy network using a policy iterative reinforcement learning algorithm. The flight state of the aircraft at the next moment is determined based on the action taken. The flight state at each historical moment within a preset time period is obtained through multiple iterations. In each iteration, the stochastic gradient of the selected flight operation is determined, and the stochastic gradient represents the value of the selected flight operation. The reinforcement learning model is trained with the objective function of maximizing the sum of the initial reward and the value at each historical moment to obtain an auxiliary decision model. The real-time flight state of the aircraft is input into the auxiliary decision model to obtain the optimal flight operation. Based on the sequence dependency between real-time flight status and optimal flight operation, a relational output is obtained; the relational output is weighted to determine the text sequence corresponding to the optimal flight operation, thus obtaining the predicted prompt word; the text position of the predicted prompt word is adjusted according to the probability of the action being predicted in the policy network, thus obtaining the optimal flight operation prompt.

[0007] Optionally, the strategy network includes: moving the joystick, operating the pedals, and adjusting the aircraft rudder.

[0008] Optionally, a flight operation is selected from the policy network based on the following formula using an iterative reinforcement learning algorithm: ; in, For policy networks, a For the selected flight operation, For policy network parameters; The stochastic gradient of the selected flight operation is determined based on the following formula: ; in, For stochastic gradients, It is the action value function for randomly sampled flight operations.

[0009] Optionally, the objective function is to maximize the sum of the initial reward and the value at each historical moment, specifically including: The flight state of the aircraft at the next moment is determined based on the actions taken, and the gradient corresponding to the flight state at the next moment is determined based on the following formula: ; ; in, The gradient corresponding to the initial state. β For gradient factor, For an unbiased estimate of the policy gradient, The gradient corresponds to the aircraft's flight state at the next moment; The objective function is determined based on the following formula: ; in, A Represents actions in the policy network. Let the action value function be the action value function under the action. Let be the expectation of the action value function under the policy network parameters for all actions and states that conform to the policy.

[0010] Optionally, optimal flight operation prompts are obtained through a prompt generation model, which includes a generative language module and a fine-tuning module connected in sequence. The generative language module includes a hidden layer and a fully connected layer connected in sequence.

[0011] Optionally, training the cue generation model includes: Retrieve historical flight operations and select the historical best flight operation prompt corresponding to the historical flight operation; The model is generated by inputting historical flight operation prompts to obtain the predicted optimal flight operation prompts. The prompt generation model is trained with the goal of maximizing the likelihood function between the historical best flight operation prompt and the predicted best flight operation prompt, resulting in the trained prompt generation model.

[0012] Alternatively, the likelihood function can be determined based on the following equation: ; Based on the following formula for window word sequences , ; ; ; ; in, for The one-hot vector representation, For word vector matrix, It is a position vector matrix. This represents the total number of layers in the Transformer. Let be the likelihood function. i =1, 2, 3, ... , To indicate the window size for generating the model, For the weights of the fully connected layer, The length of the window word sequence. For the parameters of the neural network model, P ( x i The probability that each word is predicted. Output the relation.

[0013] Optionally, the optimal flight operation prompt is obtained by adjusting the text position of the predicted prompt words based on the probability of the action being predicted in the policy network, according to the following formula: ; in, For optimal flight operation prompts, This represents the probability that an action in the policy network will be predicted.

[0014] This invention also provides a flight operation prompting system, comprising: The computing center is used to acquire offline data of the aircraft, including: the historical flight state of the aircraft at multiple historical moments and the corresponding flight operations; taking the historical flight state of the aircraft at any historical moment as the initial state and determining the gradient of the initial state, the gradient of the initial state represents the initial reward; and pre-setting a policy network composed of multiple actions, with the actions being flight operations. The auxiliary decision-making module is used to select a flight operation from the policy network based on the initial state using a policy iterative reinforcement learning algorithm, and determine the aircraft's flight state at the next moment based on the action taken. Through multiple rounds of iteration, the flight state at each historical moment within a preset time period is obtained. In each round of iteration, the stochastic gradient of the selected flight operation is determined, and the stochastic gradient represents the value of the selected flight operation. The reinforcement learning model is trained with the objective function of maximizing the sum of the initial reward and the value at each historical moment to obtain the auxiliary decision-making model. The real-time flight state of the aircraft is input into the auxiliary decision-making model to obtain the optimal flight operation. The prompt generation model is used to obtain relation output based on the sequence dependency between real-time flight status and optimal flight operation; the relation output is weighted to determine the text sequence corresponding to the optimal flight operation and obtain the predicted prompt words; the text position of the predicted prompt words is adjusted according to the probability of the action being predicted in the policy network to obtain the optimal flight operation prompt.

[0015] Optionally, it also includes a voice device and a decision support display interface; Voice equipment is used to broadcast optimal flight operation prompts; The decision support display interface is used to show optimal flight operation prompts.

[0016] The flight operation prompting method and system provided in this embodiment of the invention have the following advantages compared with the prior art: This invention trains a policy network based on offline flight data and uses a policy iteration algorithm based on reinforcement learning to dynamically optimize the flight operation selection strategy through gradient updates. It can automatically select the optimal flight operation based on the current flight status, realize intelligent filtering and integration of information, remove redundant information, and alleviate information overload for pilots.

[0017] Based on this, the sequence dependency between flight status and optimal operation is captured, and discrete operations are transformed into a coherent text sequence through fully connected layer transformation. This ensures that the prompts conform to the contextual logic of the actual flight scenario and avoids decision fragmentation caused by isolated suggestions. Through reinforcement learning, the coherence of the prompts is also enhanced through sequence modeling, which helps pilots make quick decisions and significantly improves decision efficiency and operational accuracy. Attached Figure Description

[0018] Figure 1This is a flowchart illustrating a flight operation prompting method provided in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In one embodiment, a flight operation prompting system is provided, which includes: a computing center, an auxiliary decision-making module, a prompt generation model, a voice device, and an auxiliary decision-making display interface.

[0021] The computing center acquires offline aircraft data, including historical flight states and corresponding flight maneuvers at multiple historical points in time. The historical flight state at any given time is used as the initial state, and its gradient is determined. This gradient represents the initial reward. A pre-defined policy network consists of various actions, with flight maneuvers as the actions.

[0022] The auxiliary decision-making module selects a flight operation from the policy network based on the initial state using an iterative reinforcement learning algorithm. It then determines the aircraft's flight state at the next moment based on the action taken, obtaining the flight state at each historical moment within a preset time period through multiple iterations. In each iteration, the stochastic gradient of the selected flight operation is determined, representing its value. The reinforcement learning model is trained with the objective function of maximizing the sum of the initial reward and the values ​​at each historical moment, resulting in the auxiliary decision-making model. The aircraft's real-time flight state is then input into the auxiliary decision-making model to obtain the optimal flight operation.

[0023] A prompt generation model is used to obtain relational outputs based on the sequence dependency between real-time flight status and optimal flight maneuvers. The relational outputs are weighted to determine the text sequence corresponding to the optimal flight maneuver, resulting in predicted prompt words. The text positions of the predicted prompt words are adjusted based on the probability of the action being predicted in the policy network to obtain the optimal flight maneuver prompt.

[0024] Voice equipment is used to announce optimal flight operation prompts.

[0025] The decision support display interface is used to show optimal flight operation prompts.

[0026] Its structure comprises two key components: hardware and software. The hardware includes a computing center, voice devices, and a decision support display interface, while the software includes a decision support model and a prompt generation model.

[0027] The computing center acts as the core of the system, undertaking the critical tasks of training and deploying the auxiliary decision-making model and the prompt generation model, ensuring their efficient operation. The trained auxiliary decision-making model and prompt generation model are deployed in the computing center to ensure the system can respond quickly and accurately to pilots' needs in practical applications. The auxiliary decision-making model is trained in the computing center using offline data and policy iterative reinforcement learning algorithms, enabling it to select actions based on the current state, thus providing accurate decision outputs to the pilot. Simultaneously, the prompt generation model is trained in the computing center, using a flight terminology dataset to generate language prompts corresponding to the state and action, providing pilots with intuitive and verbal decision suggestions. Voice devices, connected to the computing center, deliver the generated voice prompts to the pilot. This connection ensures that pilots receive the system-generated voice decision suggestions promptly and clearly. The auxiliary decision-making display interface, also connected to the computing center, displays decision prompts, providing visual support for the pilot and making decision information readily apparent.

[0028] During the offline data acquisition phase, the acquired data is rigorously collected through the civil aviation aircraft flight system. This offline data includes detailed records of flight status and flight operations, with precise monitoring provided by the civil aviation aircraft flight system ensuring the accuracy and comprehensiveness of the data. Regarding flight status, it covers important aircraft parameters such as airspeed, altitude, attitude, and heading, as well as the real-time operational status of various aircraft systems, providing the system with a detailed understanding of the flight environment. As for flight operations, the offline data meticulously records all pilot maneuvers performed during flight, including stick movements, pedal operations, and rudder adjustments.

[0029] During the training phase of the decision support model, the flight state of the offline data is processed into the state in the reinforcement learning algorithm. The flight operations of offline data are processed as actions in reinforcement learning algorithms. Policy Iterative Reinforcement Learning Algorithm for Constructing Policy Networks Policy Network This includes: moving the joystick, operating the pedals, and adjusting the aircraft rudder. The objective function is maximized using a gradient update method, and the objective function is expressed as:

[0030] in, A Represents actions in the policy network. Let the action value function be the action value function under the action. Let be the expectation of the action value function under the policy network parameters for all actions and states that conform to the policy.

[0031] The method to optimize the network is to update the policy network parameters using gradient ascent: in, The gradient corresponding to the initial state. β For gradient factor, For an unbiased estimate of the policy gradient, This represents the gradient corresponding to the final state.

[0032] The policy gradient is as follows: Based on the observed state Using the current strategy, the network randomly samples an action: Then calculate the stochastic gradient: Obviously, stochastic gradient It is an unbiased estimate of the policy gradient. : Therefore, it can be used. Update using stochastic gradient ascent .

[0033] in, For policy networks, a For randomly sampled flight operations, For policy network parameters, The gradient corresponding to the initial state. β For gradient factor, For an unbiased estimate of the policy gradient, The gradient corresponding to the final state. For stochastic gradients, It is the action value function for randomly sampled flight operations.

[0034] During the training phase of the prompt generation model, the input consists of flight status and flight operations, and the output is the language prompts corresponding to the flight status and flight operations. Training is performed using a generative language module combined with a flight terminology dataset. The generative language module consists of a deep neural network, and the training method is as follows:

[0035] Pre-training phase: The generative language module computes text sequences using conventional language modeling methods. Maximum likelihood estimation : For length of window word sequence , in, for The one-hot vector representation, For word vector matrix, It is a position vector matrix. This represents the total number of layers in the Transformer. Let be the likelihood function. To indicate the window size for generating the model, For the weights of the fully connected layer, The length of the window word sequence. For the parameters of the neural network model, P ( x i The probability that each word is predicted. i =1, 2, 3, ... .

[0036] The pre-training phase involves text prediction, which predicts the current word based on existing historical words and outputs the result. The probability of each word being predicted is then used... The maximum likelihood function is calculated, and a loss function is constructed based on it, which allows for the optimization of the language model.

[0037] Downstream task fine-tuning is typically trained and optimized using a flight terminology dataset. The labeled data for downstream tasks is... The input for each sample is The length of the structure is The text sequence, with corresponding labels as The labels are for flight status and flight operations.

[0038] First, the text sequence is input into a pre-trained generative language module. The hidden layer captures the sequence dependencies of the optimal operations, and the relation output is obtained. Relationships are generated and output through fully connected layers. The corresponding text sequence is used to obtain the probability that the optimal operation is predicted.

[0039] The predicted probability of obtaining the optimal operation includes: Relationships are generated and output through fully connected layers. The corresponding text sequence is compared with the content of all texts to obtain the probability that the optimal operation is predicted. Its formula is In the formula, Represents the weights of the fully connected layer ( (Indicates the number of tags) Indicates the length of the window word sequence. Output the relation.

[0040] Finally, the optimal operation prompt is obtained by fine-tuning the probability of the predicted optimal operation through the prompt generation model's fine-tuning module.

[0041] The task fine-tuning process includes: substituting the predicted probability of the optimal operation into the maximum likelihood estimation formula and performing stochastic gradient descent to obtain the maximum likelihood estimate of the optimal operation.

[0042] The formula for calculating the maximum likelihood estimate of the optimal operation is: in, This is the likelihood function between historical prompts and predicted prompts. This represents the probability that a flight operation is predicted in the policy network.

[0043] By using the above methods, the prompt generation model is trained. The trained prompt generation model can generate corresponding optimal operation prompts based on the current flight status of the aircraft and the optimal operation calculated by the auxiliary decision-making model.

[0044] In one embodiment, a flight operation prompting method is provided, such as... Figure 1 As shown, the method includes: Acquire offline aircraft data, including historical flight states and corresponding flight maneuvers at multiple historical moments. Use the aircraft's historical flight state at any given moment as the initial state and determine its gradient, which represents the initial reward. A pre-defined policy network consisting of various actions, with flight maneuvers as the actions.

[0045] Based on the initial state, a flight operation is selected from the policy network using an iterative reinforcement learning algorithm. The flight state of the aircraft at the next moment is determined based on the action taken. Through multiple iterations, the flight state at each historical moment within a preset time period is obtained. In each iteration, the stochastic gradient of the selected flight operation is determined; the stochastic gradient represents the value of the selected flight operation. The reinforcement learning model is trained with the objective function of maximizing the sum of the initial reward and the values ​​at each historical moment, resulting in an auxiliary decision model. The real-time flight state of the aircraft is input into the auxiliary decision model to obtain the optimal flight operation.

[0046] Based on the sequence dependency between real-time flight status and optimal flight maneuvers, a relational output is obtained. This output is weighted to determine the text sequence corresponding to the optimal flight maneuver, resulting in predicted prompts. The text positions of these predicted prompts are then adjusted based on the predicted probability of actions being predicted in the policy network, yielding the optimal flight maneuver prompt.

[0047] The optimal flight operation prompt is obtained through a prompt generation model, which consists of a generative language module and a fine-tuning module connected in sequence. The generative language module includes a hidden layer and a fully connected layer connected in sequence. The hidden layer captures the sequence dependency between the current flight state and the optimal flight operation, resulting in a relation output. The fully connected layer transforms the relation output to obtain the text sequence corresponding to the optimal flight operation, which is the predicted prompt word. The fine-tuning module uses a loss function to fine-tune the predicted prompt word, resulting in the optimal flight operation prompt.

[0048] The training of the prompt generation model specifically includes: acquiring historical flight operations and selecting the historically optimal flight operation prompt corresponding to the historical flight operations; inputting the historical flight operations into the prompt generation model to obtain the predicted optimal flight operation prompt; and training the prompt generation model with the objective of maximizing the likelihood function between the historically optimal flight operation prompt and the predicted optimal flight operation prompt to obtain the trained prompt generation model.

[0049] In the hardware design of the voice equipment, the computing center ensures close collaboration with the voice equipment by establishing an efficient communication connection. This design allows generated voice prompts to be transmitted to the voice equipment in real time, providing pilots with intuitive and real-time voice communication support. When pilots require decision support, the auxiliary decision-making model and prompt generation model generate corresponding decision actions and language prompts at the computing center. Through the connection between the computing center and the voice equipment, these prompts are conveyed to the pilot, providing them with comprehensive and real-time voice communication support.

[0050] In the hardware design of the decision support display interface, the computing center establishes a connection with the interface to transmit generated decision information. Pilots can intuitively access decision-related information, including status, actions, and generated prompts, on the interface. The decision support display interface integrates a high-resolution display to ensure clear display of complex decision information and graphics. Touchscreen technology is employed to provide intuitive user interaction, allowing pilots to operate and confirm information through the interface.

[0051] Once trained on the decision support model and the prompt generation model, these optimized models are embedded into the architecture of the computing center. The computing center is not merely a location for model storage; it is a crucial core for enabling real-time intelligent decision support for the system. Through the application of iterative reinforcement learning algorithms for the decision support model and the comprehensive processing of the flight terminology dataset for the prompt generation model, the computing center undertakes important tasks such as model optimization, decision generation, and prompt generation throughout the entire decision-making chain.

[0052] During system operation, the decision support model, after sensing and analyzing the complex flight conditions, accurately generates the optimal flight maneuvers and transmits the decision information, including the flight status and optimal flight maneuvers, to the input of the prompt generation model. Under the guidance of the prompt generation model, the system generates corresponding voice prompts, presenting the decision information in a clear and concise language. The computing center then transmits the generated voice prompt data to the voice equipment and the decision support display interface. This collaborative process ensures that, during system operation, pilots receive timely and accurate voice prompts and obtain intuitive decision support on the decision support display interface.

[0053] By employing reinforcement learning algorithms and prompt generation models, the decision support system can intelligently select actions based on a large amount of flight data and status, and generate personalized, verbal decision prompts, providing more intelligent and real-time decision support that is far superior to traditional methods based on predetermined rules.

[0054] The deployment and connectivity of the computing center ensures that the system can respond to pilots' needs in real time. The system can process large amounts of flight data, convey information to pilots in the form of language and graphics, and comprehensively process information to solve the problems of information overload and low decision-making efficiency in traditional methods.

[0055] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A flight operation prompting method, characterized in that, include: Acquire offline data of the aircraft, including: the historical flight state of the aircraft at multiple historical moments and the corresponding flight operations; take the historical flight state of the aircraft at any historical moment as the initial state and determine the gradient of the initial state, the gradient of the initial state representing the initial reward; preset a policy network composed of multiple actions, the actions being flight operations; Based on the initial state, a flight operation is selected from the policy network using a policy iterative reinforcement learning algorithm. The flight state of the aircraft at the next moment is determined based on the action taken. The flight state at each historical moment within a preset time period is obtained through multiple iterations. In each iteration, the stochastic gradient of the selected flight operation is determined, and the stochastic gradient represents the value of the selected flight operation. The reinforcement learning model is trained with the objective function of maximizing the sum of the initial reward and the value at each historical moment to obtain an auxiliary decision model. The real-time flight state of the aircraft is input into the auxiliary decision model to obtain the optimal flight operation. Based on the sequence dependency between real-time flight status and optimal flight operation, a relational output is obtained; the relational output is weighted to determine the text sequence corresponding to the optimal flight operation, thus obtaining the predicted prompt word; the text position of the predicted prompt word is adjusted according to the probability of the action being predicted in the policy network, thus obtaining the optimal flight operation prompt.

2. The flight operation prompting method as described in claim 1, characterized in that, The strategy network includes: moving the joystick, operating the pedals, and adjusting the aircraft rudder.

3. The flight operation prompting method as described in claim 1, characterized in that, Based on the following formula, a flight operation is selected from the policy network using a policy iterative reinforcement learning algorithm: ； in, For policy networks, a For the selected flight operation, For policy network parameters; The stochastic gradient of the selected flight operation is determined based on the following formula: ； in, For stochastic gradients, It is the action value function for randomly sampled flight operations.

4. The flight operation prompting method as described in claim 3, characterized in that, The objective function, which maximizes the sum of the initial reward and the value at each historical moment, specifically includes: The flight state of the aircraft at the next moment is determined based on the actions taken, and the gradient corresponding to the flight state at the next moment is determined based on the following formula: ； ； in, The gradient corresponding to the initial state. β For gradient factor, For an unbiased estimate of the policy gradient, The gradient corresponds to the aircraft's flight state at the next moment; The objective function is determined based on the following formula: ； in, A Represents actions in the policy network. Let the action value function be the action value function under the action. Let be the expectation of the action value function under the policy network parameters for all actions and states that conform to the policy.

5. The flight operation prompting method as described in claim 1, characterized in that, The optimal flight operation prompts are obtained through a prompt generation model, which includes a generative language module and a fine-tuning module connected in sequence. The generative language module includes a hidden layer and a fully connected layer connected in sequence.

6. The flight operation prompting method as described in claim 5, characterized in that, Training the prompt generation model specifically includes: Retrieve historical flight operations and select the historical best flight operation prompt corresponding to the historical flight operation; The model is generated by inputting historical flight operation prompts to obtain the predicted optimal flight operation prompts. The prompt generation model is trained with the goal of maximizing the likelihood function between the historical best flight operation prompt and the predicted best flight operation prompt, resulting in the trained prompt generation model.

7. The flight operation prompting method as described in claim 6, characterized in that, The likelihood function is determined based on the following formula: ； Based on the following formula for window word sequences , ； ； ； ； in, for The one-hot vector representation, For word vector matrix, It is a position vector matrix. This represents the total number of layers in the Transformer. Let be the likelihood function. i =1, 2, 3, ... , To indicate the window size for generating the model, For the weights of the fully connected layer, The length of the window word sequence. For the parameters of the neural network model, P ( x i The probability that each word is predicted. Output the relation.

8. The flight operation prompting method as described in claim 7, characterized in that, The optimal flight operation prompt is obtained by adjusting the text position of the predicted prompt words based on the probability of the action being predicted in the policy network, as shown in the following formula: ； in, For optimal flight operation prompts, This represents the probability that an action in the policy network will be predicted.

9. A flight operation prompting system, characterized in that, include: The computing center is used to acquire offline data of the aircraft, including: the historical flight state of the aircraft at multiple historical moments and the corresponding flight operations; taking the historical flight state of the aircraft at any historical moment as the initial state and determining the gradient of the initial state, wherein the gradient of the initial state represents the initial reward; and pre-setting a policy network composed of multiple actions, wherein the actions are flight operations. The auxiliary decision-making module is used to select a flight operation from the policy network based on the initial state using a policy iterative reinforcement learning algorithm, and determine the aircraft's flight state at the next moment based on the action taken. Through multiple iterations, the flight state at each historical moment within a preset time period is obtained. In each iteration, the stochastic gradient of the selected flight operation is determined, whereby the stochastic gradient represents the value of the selected flight operation. The reinforcement learning model is trained with the objective function of maximizing the sum of the initial reward and the values ​​at each historical moment to obtain the auxiliary decision-making model. The real-time flight state of the aircraft is input into the auxiliary decision-making model to obtain the optimal flight operation. The prompt generation model is used to obtain relation output based on the sequence dependency between real-time flight status and optimal flight operation; the relation output is weighted to determine the text sequence corresponding to the optimal flight operation and obtain the predicted prompt words; the text position of the predicted prompt words is adjusted according to the probability of the action being predicted in the policy network to obtain the optimal flight operation prompt.

10. A flight operation prompting system as described in claim 9, characterized in that, It also includes voice devices and decision support display interfaces; A voice device is used to broadcast the optimal flight operation prompts; The auxiliary decision display interface is used to display the optimal flight operation prompts.

Images