Aircraft landing assistance decision-making method, device and product based on deep learning and MPC

By combining deep learning and MPC in an aircraft landing assistance decision-making method, and utilizing a dynamic surrogate model with a hybrid coding module and an autoregressive decoding module, the problem of prediction error divergence and computational complexity in traditional methods is solved. This achieves high-precision, real-time aircraft landing assistance decision-making, improving the safety and stability of aircraft landing.

CN123113971APending Publication Date: 2026-11-17BEIHANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611233865.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional flight data analysis and pure data-driven deep learning models lack physical constraints during aircraft landing, leading to divergent prediction errors. Furthermore, multi-scale feature fusion is difficult, and traditional model predictive control methods are computationally complex and lack real-time performance.

Method used

An aircraft landing auxiliary decision-making method based on deep learning and MPC is adopted. By iteratively optimizing the control sequence, combining the dynamic surrogate model of the hybrid coding module and the autoregressive decoding module, and using analytical aerodynamic formulas to calculate aerodynamic characteristics for physical constraints, an end-to-end closed-loop decision-making system is constructed.

Benefits of technology

It improves the prediction accuracy and physical consistency of the aircraft landing process, reduces computational complexity, meets the needs of real-time decision-making, and enhances the safety and smoothness of landing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure FT_3
    Figure FT_3
Patent Text Reader

Abstract

This application discloses an aircraft landing aid decision-making method, device, and product based on deep learning and MPC. To address the challenge of simultaneously achieving prediction accuracy, physical consistency, and real-time decision-making, this method iteratively optimizes future control sequences within a model predictive control framework to guide the aircraft back to the glide path. This optimization utilizes a deep learning dynamics surrogate model to predict future flight states. This surrogate model processes historical flight data through a hybrid encoding module and fuses historical encodings, current data, and aerodynamic characteristics calculated using analytical aerodynamic formulas through an autoregressive decoding module for prediction. Finally, the first control variable of the optimized sequence is output as a manipulation suggestion. This invention integrates physical priors, improving prediction accuracy and physical consistency while ensuring real-time decision-making.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of flight control, and in particular to an aircraft landing assistance decision-making method, device and product based on deep learning and MPC. Background Technology

[0002] Aircraft landings are extremely risky. Traditional flight data analysis and purely data-driven deep learning models face challenges in dynamic modeling, including a lack of physical constraints leading to divergent prediction errors and difficulties in fusing continuous and discrete multi-scale data. Currently, deep learning performs exceptionally well in time-series predictions. However, purely data-driven models face two major challenges in flight dynamics applications: first, the lack of physical constraints causes long-range prediction errors to diverge rapidly when the data volume is small; second, the differences in multi-scale features make it difficult to effectively fuse continuous aircraft state parameters with discrete control inputs. Meanwhile, while Model Predictive Control (MPC) is suitable for landing assistance decision-making, its implementation is relatively difficult due to the overly complex calculations of aerodynamic formulas in traditional physical models. Summary of the Invention

[0003] The purpose of this application is to provide an aircraft landing auxiliary decision-making method, device and product based on deep learning and MPC, so as to solve the problems of low prediction accuracy and poor physical consistency of pure data-driven deep learning models in the prior art, as well as the computational complexity and insufficient real-time performance of traditional model predictive control methods.

[0004] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an aircraft landing assistance decision-making method based on deep learning and MPC, including: In the model predictive control framework, the control sequence in the future prediction time domain is iteratively optimized to enable the aircraft to return to the preset landing glide path. The control sequence consists of a series of control variables. The iterative optimization utilizes a deep learning-based dynamic surrogate model to predict future flight states and optimizes the control sequence based on the prediction results. The dynamic surrogate model adopts an architecture combining a hybrid encoding module and an autoregressive decoding module. The hybrid encoding module generates context codes by encoding historical flight data, which includes the control variables and the flight states. The autoregressive decoding module makes predictions based on aerodynamic features, the context codes, and the current flight data. The aerodynamic features are calculated from the current flight data by analyzing aerodynamic formulas. The first control variable in the optimized control sequence is output as the current manipulation suggestion.

[0005] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the aircraft landing assistance decision method based on deep learning and MPC as described above.

[0006] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aircraft landing assistance decision-making method based on deep learning and MPC as described above.

[0007] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aircraft landing assistance decision-making method based on deep learning and MPC as described above.

[0008] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides an aircraft landing auxiliary decision-making method based on deep learning and MPC. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the architecture of a deep learning-based dynamic agent model in one embodiment of this application; Figure 2 This is a flowchart illustrating the cold / hot start strategy in one embodiment of this application; Figure 3 This is a schematic diagram illustrating the principle of the cold / hot start strategy in one embodiment of this application. Detailed Implementation

[0011] 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.

[0012] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] This application provides a method, device, and product for aircraft landing assistance decision-making based on deep learning and MPC, aiming to solve the technical problems of prediction divergence caused by the lack of physical constraints in pure data-driven models and the difficulty in real-time decision-making due to the computational complexity of traditional physical models in existing technologies. This application achieves real-time and accurate assistance decision-making for the aircraft landing process by constructing a deep learning dynamics surrogate model that integrates physical priors and embedding it into the model predictive control (MPC) framework.

[0014] In one specific implementation, the aircraft landing assistance decision-making method based on deep learning and MPC provided in this application specifically includes: Step S110: In the model predictive control framework, the control sequence in the future prediction time domain is iteratively optimized to enable the aircraft to return to the preset landing glide path. The control sequence consists of a series of control variables. The iterative optimization utilizes a deep learning-based dynamic surrogate model to predict future flight states and optimizes the control sequence based on the prediction results. The dynamic surrogate model adopts an architecture combining a hybrid encoding module and an autoregressive decoding module. The hybrid encoding module generates context codes by encoding historical flight data, which includes control variables and flight states. The autoregressive decoding module makes predictions based on aerodynamic characteristics, context codes, and current flight data. The aerodynamic characteristics are calculated from the current flight data by analyzing aerodynamic formulas.

[0015] Step S120: Output the first control variable in the optimized control sequence as the current manipulation suggestion.

[0016] Specifically, the core of this method lies in the iterative optimization process within the model predictive control framework. At each decision point, within a pre-defined future prediction time domain, the system repeatedly adjusts a candidate control sequence, aiming to find a control strategy that most effectively guides the aircraft back to the pre-defined landing glide path in the future time domain. This control sequence consists of a series of control variables applied at different future time steps, such as elevator and aileron deflection angle commands.

[0017] To achieve efficient and accurate iterative optimization, this process utilizes a deep learning-based dynamics surrogate model to predict future flight states. This surrogate model replaces the computationally complex rigid body dynamics equations in traditional MPC, enabling it to quickly calculate the aircraft's flight trajectory in the future prediction time domain given the current state and a set of candidate control sequences. The optimizer comprehensively evaluates the merits of the current candidate control sequences based on factors such as the deviation between the trajectory predicted by the surrogate model and the target glide path, as well as the smoothness of control, and adjusts the sequences accordingly for the next iteration until the optimal solution is found. In this way, the efficient predictive capability of the deep learning model and the optimization control capability of MPC are organically combined, resolving the contradiction between accuracy and real-time performance that traditional methods struggle to balance.

[0018] This dynamic surrogate model features a unique architecture combining a hybrid encoding module and an autoregressive decoding module. This design aims to fully utilize the temporal information and physical laws within the aircraft flight data. The hybrid encoding module processes historical flight data, extracting key patterns and dynamic features that will influence the future from past flight states and control operations using a deep neural network structure, and compressing this into a compact context code, also known as a context code. Flight data is multimodal, including both continuously changing flight states (such as aircraft position, velocity, and attitude angles) and discrete or continuous control variables (such as control surface deflection angle and engine thrust).

[0019] The autoregressive decoding module is responsible for generating the future flight state step by step within the prediction time domain. At each prediction step, it utilizes not only the contextual encoding generated by the hybrid encoding module, which contains historical information, but also current flight data (i.e., the predicted state from the previous step and the planned control input for the current step), as well as a set of aerodynamic features calculated using analytical aerodynamic formulas. The introduction of aerodynamic features is one of the key innovations of this scheme. It explicitly injects prior knowledge of flight dynamics into the data-driven model, providing physical constraints for the model's predictions, thereby significantly improving the model's stability and accuracy in long-term predictions. Finally, after iterative optimization throughout the entire prediction time domain, the system outputs the first control input from the optimized control sequence as the current manipulation recommendation to the pilot or flight control system, achieving a closed loop from state awareness to decision recommendation.

[0020] In summary, this invention explicitly injects aerodynamic features calculated using analytical aerodynamic formulas into a deep learning model, combining prior physical knowledge with a data-driven model. This ensures that model predictions not only follow patterns in the data but also conform to fundamental flight dynamics laws, effectively avoiding the divergence problem caused by error accumulation in long-term predictions of pure data models. This significantly improves prediction accuracy and physical consistency. Furthermore, this invention constructs a complete closed loop from historical data analysis and future state prediction to the generation of optimal control recommendations. This directly transforms complex flight state deviations into specific, actionable control suggestions, providing efficient end-to-end intelligent decision support for pilots or autopilot systems, contributing to improved safety and smoothness during landing.

[0021] In some embodiments, the hybrid encoding module includes a TCN encoder for extracting local temporal features and a Transformer encoder for establishing global temporal dependencies, with the output of the TCN encoder serving as the input of the Transformer encoder.

[0022] Reference Figure 1 During operation, the input time-series data sequence (historical flight data) first passes through the TCN encoder. The TCN encoder, through its causal convolution and dilated convolution structure, efficiently captures local dependencies and short-term dynamic patterns in the data, such as the rapid response of the aircraft attitude after the pilot manipulates the control surfaces. Subsequently, the output of the TCN encoder is fed into the Transformer encoder. The Transformer encoder, utilizing its core self-attention mechanism, can establish dependencies between any two time steps in the input sequence, thereby capturing global, long-term temporal dependencies. By concatenating the TCN and Transformer, this hybrid encoding module can take into account both local and global dynamic characteristics, providing richer and more comprehensive historical information encoding for subsequent decoding and prediction—that is, context encoding.

[0023] Preferably, the TCN encoder is a two-layer causal dilated convolutional network. The first layer of the TCN encoder has a dilation rate of 1, and the second layer has a dilation rate of 2. After the two layers are stacked, the total receptive field is 13 frames, which is used to capture short-term local dynamic patterns such as the instantaneous attitude response after the control surface deflection. The Transformer encoder also adopts a two-layer structure. Each layer consists of a Pre-LN multi-head self-attention mechanism and a feedforward network. During training, a causal mask can be applied to ensure that the self-attention does not cross the temporal causal boundary. The Transformer encoder is responsible for establishing global temporal dependencies outside the receptive field of the TCN encoder, and finally outputs the context encoding.

[0024] Reference Figure 1In some embodiments, prediction is made based on aerodynamic features, context encoding, and current flight data. Specifically, this includes: fusing context encoding and current flight data using a cross-attention mechanism to obtain fused features; injecting aerodynamic features into the fused features and making predictions based on the fused features after injection.

[0025] Specifically, to effectively fuse historical contextual encodings and current flight data, this module employs a cross-attention mechanism. Specifically, when predicting each future time step, the current flight data (current frame) "queries" the contextual encodings generated by the encoder, which contain historical information. It dynamically filters and aggregates the historical information most relevant to the current prediction through attention weights, generating a fused feature that integrates historical and current information. Simultaneously, a separate aerodynamic feature calculation module calculates key aerodynamic features based on the current frame's flight data by analyzing aerodynamic formulas. These aerodynamic features are then "injected" into the aforementioned fused feature, collectively used to predict the flight state at the next moment. Through this design, the model not only utilizes the temporal correlations in the data but also explicitly integrates physical laws, making the prediction process both data-driven flexibility and consistent with the physical model.

[0026] Furthermore, the autoregressive decoding module includes a decoupled longitudinal expert network and a lateral expert network; it injects aerodynamic features into the fused features and then performs predictions based on the fused features, specifically including: splitting the aerodynamic features into longitudinal features and lateral features, injecting the longitudinal features and lateral features into the fused features respectively; processing the fused features injected with longitudinal features through the longitudinal expert network; processing the fused features injected with lateral features through the lateral expert network; and performing predictions based on the outputs of the longitudinal expert network and the lateral expert network.

[0027] This design draws upon an important prior knowledge in flight mechanics: under small disturbances, the longitudinal motion (e.g., pitch, climb) and lateral motion (e.g., roll, yaw) of an aircraft are largely independent. Therefore, when injecting aerodynamic features, the aerodynamic features calculated from these features are first decomposed into longitudinal features (e.g., lift coefficient) and lateral features (e.g., side force coefficient, roll moment coefficient). Then, the longitudinal features are combined with the fused features and fed into a longitudinal expert network for processing to predict longitudinal state changes; simultaneously, the lateral features are combined with the fused features and fed into a lateral expert network for processing to predict lateral state changes. Finally, the outputs of the two expert networks are combined to obtain a complete flight state prediction. This decoupled design reduces the learning difficulty of individual networks, enabling the model to more accurately capture the dynamic characteristics of different motion modes and improving overall prediction accuracy.

[0028] In some embodiments, a cold-start strategy is employed during the iterative optimization process to improve optimization efficiency.

[0029] The cold-start and warm-start strategy aims to provide a better initial guess for each round of optimization, thereby reducing the number of iterations and accelerating convergence. This is a key approach to solving the problem of time-consuming online computation in MPC. By adopting cold-start and warm-start strategies, decision-making time can be significantly shortened, meeting the stringent real-time requirements of the aircraft landing phase.

[0030] Furthermore, the cold and hot start strategy includes: at the initial decision time, initializing the control sequence at each time step in the future prediction time domain to a preset initial manipulation quantity; at subsequent decision times, shifting the optimal control sequence obtained from the previous decision time optimization and using it as the initial control sequence for the current optimization process.

[0031] Specifically, cold start occurs at the initial decision point or when the system needs to be reset. At this time, since no prior optimization information is available, the system initializes the control sequence for each time step in the future prediction time domain (i.e., cold start initialization) to a preset initial manipulation variable, such as filling it with the last value of the historical manipulation sequence. Warm start, on the other hand, is applied to subsequent consecutive decision points. After optimization is completed at time t, the system has obtained an optimal control sequence from t to t+N. When a new round of optimization begins at time t+1, the system no longer blindly initializes; instead, it shifts the optimal sequence obtained at time t, using the control variables from steps 2 to N at time t as the initial values ​​for steps 1 to N-1 at time t+1 (i.e., warm start initialization), and copies the last step to fill the Nth step. This warm start method provides an initial point very close to the optimal solution for the optimization process, thus greatly accelerating the convergence process and ultimately obtaining the output.

[0032] In some embodiments, iterative optimization is achieved by minimizing a cost value, which includes a tracking term of the predicted trajectory against a standard glide path and an operational smoothing term of the control sequence, the predicted trajectory being determined based on the predicted flight state.

[0033] Specifically, iterative optimization essentially solves an optimization problem, namely, minimizing a cost function by adjusting the control sequence. This cost function is designed to include at least two parts: a tracking term between the predicted trajectory and the standard glide path, and an operational smoothing term for the control sequence. The tracking term penalizes the deviation between the predicted flight trajectory and the target reference trajectory (i.e., the standard glide path); a smaller value indicates that the aircraft is closer to the target trajectory. The operational smoothing term penalizes drastic changes between adjacent control variables in the control sequence; a smaller value indicates smoother control suggestions that are easier for the pilot to execute. By minimizing the weighted sum of these two terms, this method can ensure the effectiveness of trajectory correction while avoiding overly aggressive or erratic control commands, thus improving the practicality and safety of decision-making.

[0034] To ensure that the generated control suggestions always remain within the aircraft's safety envelope and the pilot's comfort zone, the cost function includes a constraint penalty term. This term penalizes any violation of preset flight state constraints (such as maximum permissible angle of attack and minimum flight speed) or control input constraints (such as maximum control surface deflection and engine thrust rate of change limits). When the predicted flight state or planned control input exceeds these constraints, the constraint penalty term incurs a significant cost, guiding the optimization process to find a suboptimal solution that does not violate the constraints. By introducing the constraint penalty term, the control suggestions generated by this method effectively correct the flight path while strictly ensuring flight safety, making them more reliable in practical applications.

[0035] In some embodiments, the flight state includes position, attitude, and speed; the control quantities include at least one of elevator deflection command, aileron deflection command, rudder deflection command, and engine thrust command; the flight data also includes environmental observations, including wind speed, wind direction, and ground contact status; and the aerodynamic characteristics include at least one of lift coefficient, drag coefficient, and side force coefficient.

[0036] Specifically, flight data may include, but is not limited to, at least one or more combinations of the following: aircraft position (e.g., longitude, latitude, altitude), speed (e.g., ground speed, airspeed, vertical speed), attitude angles (e.g., pitch angle, roll angle, yaw angle), angular rates (e.g., pitch rate, roll rate, yaw rate), control surface deflection angles (e.g., elevator, aileron, rudder deflection angles), and engine thrust (e.g., throttle position or engine speed). These data collectively constitute complete information describing the aircraft's motion state and control intentions.

[0037] Accordingly, the control sequence can include, but is not limited to, various combinations of elevator deflection commands, aileron deflection commands, rudder deflection commands, and engine thrust commands at each time step. These commands directly correspond to equipment that the pilot can operate, making the recommendations output by this method directly executable.

[0038] Aerodynamic characteristics calculated by analyzing aerodynamic formulas and injected into the autoregressive decoding module can include, but are not limited to, at least one or more combinations of the aircraft's lift coefficient, drag coefficient, and side force coefficient. These coefficients are core parameters describing aircraft aerodynamics; injecting them as physical priors into the model can greatly improve the model's physical consistency and predictive accuracy.

[0039] As can be seen from the above description, the present invention has the following technical effects: 1. By explicitly injecting the aerodynamic features calculated by analytical aerodynamic formulas into the deep learning model, the combination of physical prior knowledge and data-driven model makes the model prediction not only follow the patterns in the data, but also conform to the basic laws of flight dynamics. This effectively avoids the divergence problem caused by error accumulation in long-term prediction of pure data models, and significantly improves the accuracy and physical consistency of prediction.

[0040] 2. By adopting a lightweight deep learning dynamics proxy model to replace the computationally cumbersome physical model in traditional MPC, the computational cost of a single prediction is significantly reduced. At the same time, the combination of cold and hot start strategy provides a better initial solution for iterative optimization, which significantly accelerates the convergence speed of the optimization process, thereby meeting the stringent requirements for real-time auxiliary decision-making during the aircraft landing phase.

[0041] 3. A complete closed loop has been constructed, from historical data analysis and future state prediction to the generation of optimal control suggestions. This loop can directly transform complex flight state deviations into specific and executable control suggestions, providing efficient end-to-end intelligent decision support for pilots or autopilot systems, and helping to improve the safety and smoothness of the landing process.

[0042] By designing an expert network architecture that decouples the longitudinal and lateral directions, and drawing on prior knowledge of motion approximation decoupling in flight mechanics, the learning complexity of the model is effectively reduced, enabling the model to better generalize and adapt to various complex flight modes, such as stable approach and large-amplitude maneuvers.

[0043] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0044] For example, in one embodiment, a complete aircraft landing assistance decision-making process is provided. The core of this process is to use a deep learning model that incorporates prior physical knowledge as a high-precision predictor and place it in the closed loop of model predictive control (MPC). Through online iterative optimization, it generates control suggestions in real time to correct the aircraft's landing trajectory.

[0045] First, the system prepares the data, such as collecting flight data of a certain type of passenger aircraft during landing at a specific airport, including state variables such as position, attitude, and speed, control variables such as ailerons, elevators, and throttle, as well as environmental and aerodynamic observations such as angle of attack, sideslip angle, and wind speed, to form a time series dataset, which includes flight data at each time step.

[0046] Next, a deep learning dynamics agent model, including an encoder and a decoder, is constructed and trained. (See reference...) Figure 1 The encoder takes historical flight data as input. It consists of a cascaded two-layer causal dilated convolutional network (TCN) and a two-layer Transformer encoder. The first layer of the TCN encoder has a dilation rate of 1, and the second layer has a dilation rate of 2, used to capture short-term local dynamics. The Transformer encoder has eight attention heads per layer to establish global temporal dependencies, ultimately outputting contextual encoding. The decoder employs an autoregressive structure, with a core four-head cross-attention mechanism module for fusing contextual encoding and current flight data. Simultaneously, an aerodynamic feature calculation module calculates 18-dimensional aerodynamic features by analyzing aerodynamic formulas and injects them into the position prediction expert network. It also outputs attitude information such as angle of attack / sideslip angle through an auxiliary head, serving as the attitude condition for the next moment. The position prediction expert network is internally decoupled into longitudinal and lateral expert networks, each performing predictions and outputting the current flight state through an output head. Finally, position prediction is performed using velocity-time integration to update the aircraft's position; the positions of multiple consecutive future time steps constitute the predicted flight trajectory.

[0047] The above describes the training process of a deep learning dynamics surrogate model.

[0048] Then, the trained surrogate model is embedded into the MPC controller. The controller uses the aforementioned surrogate model as its internal prediction function (surrogate model forward prediction). The cost function is designed to include two parts: one part is the mean square error between the predicted trajectory and the standard glide path reference trajectory (tracking term), and the other part is the rate of change of the control sequence (operational smoothing term).

[0049] Reference Figure 2 and Figure 3During the aircraft approach, the system acquires the flight data sequence of the most recent 100 frames in real time. At the first decision point, the MPC controller initiates a cold start: it initializes all control sequences within the future prediction time domain (e.g., 50 time steps) with the values ​​of the last frame of the historical control sequence, thus forming a cold start initialization.

[0050] Subsequently, the system enters the MPC optimization loop. First, a surrogate model is used for forward prediction, forecasting the aircraft's flight trajectory for the next 50 time steps based on the current control sequence. Then, the cost value is calculated. Next, the gradient of the cost function with respect to the control sequence is calculated via backpropagation, and an optimizer (such as Adam) is used to update the control sequence based on the gradient. This process is repeated until the maximum number of iterations is reached or the cost function converges, resulting in the optimized control sequence. Finally, the control input for the first time step of the optimized control sequence is output as the optimal suggestion. At the next decision point, the MPC controller initiates a warm start: the optimal control sequence (steps 2 to 50) optimized at the previous decision point is shifted as the initial control sequence for the current optimization, and then the optimization loop is repeated. This implementation can predict the aircraft trajectory in real time and accurately, and generate smooth and effective control suggestions, guiding the aircraft smoothly back from a deviation to a standard glide path.

[0051] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0052] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0053] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0054] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0055] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0056] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0057] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A landing assistance decision-making method for aircraft based on deep learning and MPC, characterized in that, include: In the model predictive control framework, the control sequence in the future prediction time domain is iteratively optimized to enable the aircraft to return to the preset landing glide path. The control sequence consists of a series of control variables. The iterative optimization utilizes a deep learning-based dynamic surrogate model to predict future flight states and optimizes the control sequence based on the prediction results. The dynamic surrogate model adopts an architecture combining a hybrid encoding module and an autoregressive decoding module. The hybrid encoding module generates context codes by encoding historical flight data, which includes the control variables and the flight states. The autoregressive decoding module makes predictions based on aerodynamic features, the context codes, and the current flight data. The aerodynamic features are calculated from the current flight data by analyzing aerodynamic formulas. The first control variable in the optimized control sequence is output as the current manipulation suggestion.

2. The aircraft landing auxiliary decision-making method based on deep learning and MPC according to claim 1, characterized in that, The hybrid encoding module includes a TCN encoder for extracting local temporal features and a Transformer encoder for establishing global temporal dependencies. The output of the TCN encoder is the input of the Transformer encoder.

3. The aircraft landing auxiliary decision-making method based on deep learning and MPC according to claim 1, characterized in that, Prediction is made based on aerodynamic characteristics, the context encoding, and current flight data, specifically including: A cross-attention mechanism is used to fuse the context encoding with the current flight data to obtain fused features; The aerodynamic features are injected into the fused features, and predictions are made based on the fused features after injection.

4. The aircraft landing auxiliary decision-making method based on deep learning and MPC according to claim 3, characterized in that, The autoregressive decoding module includes a decoupled vertical expert network and a horizontal expert network. Injecting the aerodynamic features into the fused features and then performing predictions based on the fused features after injection specifically includes: The aerodynamic feature is split into longitudinal feature and lateral feature, and the longitudinal feature and the lateral feature are respectively injected into the fused feature; The fused features injected with the vertical features are processed through the vertical expert network; The fused features injected with the lateral features are processed through the lateral expert network; Prediction is made based on the outputs of the vertical expert network and the horizontal expert network.

5. The aircraft landing auxiliary decision-making method based on deep learning and MPC according to claim 1, characterized in that, A cold-start strategy is employed during the iterative optimization process to improve optimization efficiency.

6. The aircraft landing auxiliary decision-making method based on deep learning and MPC according to claim 5, characterized in that, The cold / hot start strategy includes: At the initial decision moment, the control sequence at each time step in the future prediction time domain is initialized to the preset initial manipulation quantity; At subsequent decision moments, the optimal control sequence obtained from the previous decision moment is shifted over time and used as the initial control sequence for the current optimization process.

7. The aircraft landing auxiliary decision-making method based on deep learning and MPC according to claim 1, characterized in that, The iterative optimization is achieved by minimizing a cost value, which includes a tracking term of the predicted trajectory to the standard glide path and an operational smoothing term of the control sequence, the predicted trajectory being determined based on the predicted flight state.

8. The aircraft landing auxiliary decision-making method based on deep learning and MPC according to claim 1, characterized in that, The flight status includes position, attitude, and velocity; The control quantity includes at least one of elevator deflection command, aileron deflection command, rudder deflection command, and engine thrust command; The flight data also includes environmental observations, such as wind speed, wind direction, and grounding status.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the aircraft landing assistance decision method based on deep learning and MPC as described in any one of claims 1-8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the aircraft landing assistance decision method based on deep learning and MPC as described in any one of claims 1-8.