CN-121999651-A - Multi-landing-field low-altitude aircraft autonomous arrival method based on sector division and deep reinforcement learning

CN121999651ACN 121999651 ACN121999651 ACN 121999651ACN-121999651-A

Abstract

The invention discloses an autonomous arrival method of a multi-landing-field low-altitude aircraft based on sector division and deep reinforcement learning, which relates to the technical field of intelligent arrival control of low-altitude multi-landing-field aircraft, and comprises the steps of establishing a circular terminal airspace model, uniformly configuring a plurality of landing fields in an airspace, wherein each landing field has a dedicated landing zone and executing single-machine occupation constraint; dividing a terminal airspace into orthogonal sectors with the number equal to the number of landing fields, each sector covering a specific angle range, determining the sector to which the aircraft belongs according to the azimuth angle of the aircraft relative to the airspace center, establishing a deterministic mapping relation between the sector and the landing fields based on a rotation offset rule, constructing a multi-flow space-time attention neural network architecture, training the neural network by adopting a dual-delay depth deterministic strategy gradient algorithm, and decoupling collision avoidance learning and navigation targets by adopting a staged task allocation protocol. The invention obviously improves the safety, efficiency, energy management and robustness through the cooperation of sector division and reinforcement learning.

Inventors

WANG RUIXIN
ZHANG LU
WANG JUBO
ZHU TIANCHEN
LI DAQING

Assignees

中国民航大学

Dates

Publication Date: 20260508
Application Date: 20260225

Claims (10)

1. The method for autonomously arriving the multi-landing-field low-altitude aircraft based on sector division and deep reinforcement learning is characterized by comprising the following steps of: Establishing a circular terminal airspace model which takes the airspace center as the circle center and has a preset operation radius and a regeneration boundary radius, and uniformly configuring a plurality of landing fields in the airspace, wherein each landing field has a dedicated landing area and performs single-machine occupation constraint; Dividing the terminal airspace into orthogonal sectors with the number equal to the number of landing fields, wherein each sector covers a specific angle range, determining the sector to which the aircraft belongs according to the azimuth angle of the aircraft relative to the airspace center, and establishing a deterministic mapping relation between the sector and the landing fields based on a rotation offset rule; Constructing a multi-stream space-time attention neural network architecture, wherein the architecture comprises a special encoder for respectively processing an aircraft dynamic state stream, a task information stream and a social information stream, the task information stream encoder adopts a task-oriented weighted aggregation mechanism to strengthen the distribution of target characteristics, and the social information stream encoder processes variable-scale neighbor information through a long-period memory network; Training the neural network by adopting a dual-delay depth deterministic strategy gradient algorithm, applying a multi-objective rewarding function comprising layering security penalty, objective achievement rewarding, operation efficiency rewarding and control comfort degree penalty in a training process, and decoupling collision avoidance learning and navigation objectives through a phased task allocation protocol.
2. The multi-takeoff and landing field low altitude aircraft autonomous arrival method of claim 1 wherein said division of orthogonal sectors is accomplished by calculating aircraft azimuth angles and applying a sector assignment function, the sector angle range being 360 degrees divided by the total number of sectors, the sector number being determined by the azimuth integer division result.
3. The autonomous arrival method of multiple take-off and landing low-altitude aircraft according to claim 1, wherein the deterministic mapping of the sectors and the take-off and landing fields adopts a rotation offset rule, and three priority allocation strategies are implemented under sector constraint, wherein the three priority allocation strategies comprise a distance priority strategy which only considers the aircraft of which the sector matches a take-off and landing field designated source sector, selects the aircraft with the minimum Euclidean distance to a target take-off and landing field, a first-in first-out strategy which selects the unallocated aircraft of which the sector matches and enters an airspace time earliest, and a battery priority strategy which selects the aircraft of which the sector matches and the percentage of the residual battery power is lowest.
4. The multi-landing low-altitude aircraft autonomous arrival method of claim 1, wherein the dynamic state stream comprises aircraft position, speed, and heading angle, the mission information stream comprises each landing field relative position, euclidean distance, and assigned state mask, and the social information stream comprises neighbor distance, relative position, relative speed, heading intersection angle, and approach point space-time parameters.
5. The autonomous arrival method of a multi-takeoff and landing low-altitude aircraft according to claim 1, wherein in the multi-stream space-time attention neural network architecture, a task information stream encoder performs multi-head self-attention computation, generates task guide weights by using an allocation state mask, highlights allocated target features through weighted aggregation, a dynamic state stream encoder adopts a multi-layer perceptron, a social information stream encoder inputs long-term and short-term memory network extraction features after sorting adjacent machines according to distance, and each stream feature is fused by a transducer to output action instructions after time encoding.
6. The multi-takeoff and landing low-altitude aircraft autonomous arrival method of claim 1, wherein said tiered security penalty includes a three-level mechanism that applies an accident-level penalty when the minimum paired distance is below a collision threshold, an event-level penalty when the minimum paired distance is in a collision alert interval, and a gaussian potential field-based approach warning penalty when the minimum paired distance is above the collision threshold but there is a potential risk.
7. The multi-takeoff and landing low-altitude aircraft autonomous arrival method of claim 1, wherein said target achievement rewards include a distance reduction reward calculated by differential distance reduction, a heading alignment reward that encourages aircraft to fly toward the assigned target, and a restricted zone violation penalty that prohibits unassigned aircraft from approaching the takeoff and landing.
8. The multi-takeoff and landing low-altitude aircraft autonomous arrival method according to claim 1, wherein said control comfort penalty employs a fourth-order function of heading change rate, significantly suppressing aggressive maneuver to reduce energy consumption and promote trajectory smoothness, and said operational efficiency reward is calculated based on a delay difference of actual arrival time and theoretical straight flight time upon successful descent.
9. The multi-take-off and landing low-altitude aircraft autonomous arrival method according to claim 1, wherein the phased task allocation protocol disables target allocation in an initial training stage to enable an agent to concentrate on learning a basic collision avoidance behavior, and a subsequent stage adopts a cyclic selection strategy to allocate different take-off and landing sites for the aircraft to ensure that training data covers all destination scenes.
10. The method for autonomously arriving a multi-takeoff and landing low-altitude aircraft according to claim 1, wherein the training process adopts a course learning strategy to increase the task complexity by gradually increasing the number of the aircraft in the space, and simultaneously implements a boundary constraint mechanism to reset the position of the aircraft to the previous state when the aircraft exceeds the regeneration boundary.

Description

Multi-landing-field low-altitude aircraft autonomous arrival method based on sector division and deep reinforcement learning Technical Field The invention relates to the technical field of intelligent arrival control of low-altitude multi-takeoff and landing field aircrafts, in particular to an autonomous arrival method of a multi-takeoff and landing field low-altitude aircrafts based on sector division and deep reinforcement learning. Background Urban air traffic (Urban Air Mobility, UAM) is used as an innovative traffic mode, and an electric vertical take-off and landing aircraft (eVTOL) is utilized to flexibly operate in a three-dimensional airspace, so that a solution is provided for relieving ground traffic jam. This mode is expected to break through the limitations of traditional ground traffic. However, unlike traditional civil aviation operating in controlled high-altitude corridors, urban air traffic presents operational challenges. The traditional air traffic management rule is formulated for high-altitude operation and centralized control structure, and the core of the air traffic management rule comprises the characteristics of fixed navigation, large-interval separation, manual control, post analysis and the like. These rules cannot accommodate the unique needs of low-altitude environments, particularly high-density traffic, distributed take-off and landing network coordination, highly autonomous operation, and transition of risk management to proactive prevention. The most prominent problem in the prior art is the lack of a specific enforceable digital flight rule for multi-takeoff and landing field coordination. While related organizations build high-level conceptual frameworks that define the goals of urban air traffic systems, these frameworks fail to specify behavior specifications for autonomous aircraft in specific operational scenarios. Particularly in a multi-landing scenario, an aircraft needs to dynamically select a path between competing destinations while managing lateral cross-conflicts with other aircraft traveling to different landing sites. The existing single take-off and landing self-organizing arrival framework realizes the decentralized autonomous arrival management, but is only suitable for scenes in which all aircrafts share the same destination, and space conflict mainly occurs along the radial inbound direction. When expanding to a plurality of take-off and landing sites, the pure learning method faces the difficulty of rapid expansion of exploration space, the action space for covering destination selection, path planning and microscopic collision prevention is increased, and the emerging behavior lacks a structural mechanism for providing operation guarantee for a supervision mechanism and cannot adapt to heterogeneous constraints such as battery power difference and priority. Disclosure of Invention In view of the above, the invention provides an autonomous arrival method of a multi-landing-field low-altitude aircraft based on sector division and deep reinforcement learning, so as to solve the problems of dynamic selection and conflict management of the multi-landing-field low-altitude aircraft in the prior art. The specific technical scheme of the invention is as follows: a multi-landing-field low-altitude aircraft autonomous arrival method based on sector division and deep reinforcement learning comprises the following steps: Establishing a circular terminal airspace model which takes the airspace center as the circle center and has a preset operation radius and a regeneration boundary radius, and uniformly configuring a plurality of landing fields in the airspace, wherein each landing field has a dedicated landing area and performs single-machine occupation constraint; dividing a terminal airspace into orthogonal sectors with the number equal to the number of landing fields, wherein each sector covers a specific angle range, determining the sector to which the aircraft belongs according to the azimuth angle of the aircraft relative to the airspace center, and establishing a deterministic mapping relation between the sector and the landing fields based on a rotation offset rule; Constructing a multi-stream space-time attention neural network architecture, wherein the architecture comprises a special encoder for respectively processing an aircraft dynamic state stream, a task information stream and a social information stream, the task information stream encoder adopts a task-oriented weighted aggregation mechanism to strengthen the distribution of target characteristics, and the social information stream encoder processes variable-scale neighbor information through a long-period memory network; The neural network is trained by adopting a dual-delay depth deterministic strategy gradient algorithm, a multi-objective rewarding function comprising layering security penalty, objective achievement rewarding, operation efficiency rewarding and control comfort degree penal