Unmanned aerial vehicle flight control method and system based on Mangbar and dual-channel attention mechanism under wind field disturbance

Abstract

The invention discloses an unmanned aerial vehicle flight control method and system based on a Mangbar and a dual-channel attention mechanism under wind field disturbance, and particularly relates to the technical field of unmanned aerial vehicle autonomous navigation and intelligent control. The method comprises the following steps: constructing a wind field disturbance model based on computational fluid mechanics, and simulating an unsteady wind field environment among urban building groups; designing a deep reinforcement learning network fusing a Mama framework and a double-channel attention mechanism, wherein the deep reinforcement learning network is used for realizing dynamic modeling of long-sequence wind field features and attention weighting of key spatial-temporal features; a dynamic road sign point guiding mechanism and a multi-stage attenuation greedy strategy are put forward, the sparse reward problem is relieved, and the local path optimization capability is improved; and a multi-target composite reward function including energy constraint, safety distance and path efficiency is constructed, and global path planning of energy consumption perception is realized. The method can improve the flight stability, path smoothness and energy efficiency of the unmanned aerial vehicle in a strong wind disturbance environment, and is suitable for unmanned aerial vehicle autonomous navigation tasks in an urban complex environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle autonomous navigation and intelligent control, in particular to an unmanned aerial vehicle flight control method and system based on Mamba and double-channel attention mechanism under wind field disturbance. BACKGROUND

[0002] As a flexible aerial platform, unmanned aerial vehicles (UAVs) are increasingly used in military reconnaissance, logistics distribution, city inspection, and other fields. Path planning is one of the core technologies to ensure the autonomous, safe, and efficient execution of UAV tasks. However, in complex environments such as cities, UAV flight faces severe challenges, with the unsteady, spatially heterogeneous wind field disturbance caused by building flow being a major difficulty.

[0003] Existing UAV flight control methods mainly include traditional algorithms and deep reinforcement learning (DRL) based methods. Traditional algorithms such as A* and Dijkstra, while theoretically complete, face computational efficiency bottlenecks in dynamic environments due to the need for frequent re-planning, and are difficult to effectively handle continuous dynamic physical disturbances such as wind fields. In recent years, deep reinforcement learning methods have shown excellent environmental adaptability by combining the perception capabilities of deep learning with the decision-making advantages of reinforcement learning. For example, DeepQ-Network (DQN) and its improved algorithms (such as DoubleDQN, DDQN) have been applied to path planning. To handle temporal dynamics, long short-term memory networks (LSTMs) are introduced into the DRL framework to capture historical state information. In addition, attention mechanisms improve the model's ability to focus on important environmental information (such as dynamic obstacles) by adaptively weighting key features.

[0004] However, when faced with complex wind field disturbances, existing DRL-based methods still have obvious shortcomings: 1. Recurrent neural networks such as LSTM have gradient vanishing / explosion problems, limited modeling capabilities for long-term wind field spatio-temporal dependencies, and their serial computing mechanism may become a performance bottleneck in rapidly changing wind fields with high real-time requirements; 2. Many studies do not fully consider the effects of micro-scale wind field disturbances generated by building turbulence, or use overly simplified wind field models that do not accurately reflect the real physical environment, resulting in poor planning strategies in practice; 3. Path planning needs to balance multiple goals such as shortest path, obstacle avoidance safety, and minimum energy consumption. Existing methods often fail to effectively balance these goals, especially under wind field disturbance, frequent attitude adjustments can lead to a sharp increase in energy consumption, and most studies lack explicit modeling and optimization of energy management; 4. In a wide environment, the UAV can only obtain significant rewards when it reaches the target, which leads to low learning efficiency and slow convergence. Summary of the Invention

[0005] The purpose of this invention is to propose a UAV flight control method and system based on the Mamba and dual-channel attention mechanisms under wind field disturbances, aiming to solve the problems of inaccurate modeling of complex wind field disturbances, insufficient ability to capture long-term time-series dependencies, and difficulty in coordinating the optimization of path length, safety and energy consumption in existing methods.

[0006] According to a first aspect of the present disclosure, a method for unmanned aerial vehicle (UAV) flight control based on Mamba and dual-channel attention mechanisms under wind field disturbances is provided, comprising the following steps: Based on computational fluid dynamics models, a dynamic environmental model reflecting the disturbance of the wind field around a building is constructed. The design of the reinforcement learning decision framework defines the state space, action space, and multi-objective composite reward function of the UAV based on the dynamic environment model. The state space includes the UAV state, obstacle information, target point information, and wind field information. The action space is a discrete flight direction command. The multi-objective composite reward function includes path reward, energy penalty, wind direction reward, and landmark reward. A deep reinforcement learning network SMA-DDQN is constructed. This network uses a reinforcement learning decision framework as a constraint, extracts the spatial and temporal features of the input state in parallel through a dual-channel attention module, and uses a Mamba module to model the temporal dependencies of long sequences to output the optimal action value function. The deep reinforcement learning network SMA-DDQN was trained using a dynamic landmark generation mechanism and a multi-stage decay greedy exploration strategy to obtain a path planning model. By inputting real-time environmental conditions into the path planning model, flight control commands are obtained, driving the UAV to complete path tracking and obstacle avoidance in windy and turbulent environments.

[0007] According to a second aspect of the present disclosure, a drone flight control system based on Mamba and dual-channel attention mechanisms under wind field disturbances is provided, comprising: The dynamic environment modeling module, based on computational fluid dynamics models, constructs dynamic environment models that reflect the disturbances of the wind field around a building. The decision framework design module designs a reinforcement learning decision framework. Based on the dynamic environment model, it defines the state space, action space, and multi-objective composite reward function of the UAV. The state space includes the UAV state, obstacle information, target point information, and wind field information. The action space is a discrete flight direction command. The multi-objective composite reward function includes path reward, energy penalty, wind direction reward, and landmark reward. The deep reinforcement learning network construction module constructs the SMA-DDQN deep reinforcement learning network. This network uses the reinforcement learning decision framework as a constraint, extracts the spatial and temporal features of the input state in parallel through a dual-channel attention module, and uses the Mamba module to model the temporal dependencies of long sequences to output the optimal action value function. The network training module employs a dynamic landmark generation mechanism and a multi-stage decaying greedy exploration strategy to train the deep reinforcement learning network SMA-DDQN, thereby obtaining a path planning model. The online flight control execution module inputs real-time environmental conditions into the path planning model to obtain flight control commands, driving the UAV to complete path tracking and obstacle avoidance in windy and turbulent environments.

[0008] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned UAV flight control method based on Mamba and dual-channel attention mechanisms under wind field disturbance.

[0009] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned method for unmanned aerial vehicle (UAV) flight control based on Mamba and dual-channel attention mechanisms under wind field disturbances.

[0010] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention simulates unsteady wind field disturbances caused by flow around urban building clusters by coupling the Reynolds-averaged Navier-Stokes (RANS) equations in computational fluid dynamics (CFD) with a k-ε turbulence model. Simultaneously, it incorporates a UAV energy consumption physics model including base power consumption, aerodynamic drag power, and steering energy consumption, constructing a multi-physics coupled training environment that closely approximates real-world physical laws. This environment realistically reflects the dynamic correlation between wind field disturbances and UAV energy consumption, providing a reliable physical basis for subsequent decision-making, significantly reducing performance discrepancies between simulation and real-world scenarios, and substantially improving practicality and scenario transferability.

[0011] 2. This invention is the first to introduce the Mamba architecture into the field of UAV path planning. Its Selective State-Space Model (SSM), through an input-dependent dynamic parameter generation mechanism, can more efficiently and accurately capture the long-sequence temporal dependencies of wind field disturbances. Compared with traditional Long Short-Term Memory (LSTM) networks, the Mamba architecture not only overcomes the problem of insufficient long-term memory capacity caused by gradient vanishing / exploding, but also solves the efficiency bottleneck of LSTM's serial computation with linear computational complexity. This enables UAVs to more accurately predict wind field change trends and provide timely and reliable temporal data for dynamic path adjustment.

[0012] 3. This invention combines a dual-channel attention mechanism with the Mamba module. The dual-channel attention mechanism can adaptively weight the spatial features (such as the location and size of key obstacles) and temporal features (such as the temporal changes in wind intensity and direction) of the input state in parallel, focusing on core decision-making information. The Mamba module further models long-term temporal dependencies on this basis, achieving deep fusion of spatiotemporal features. This fusion mechanism enables the intelligent agent to simultaneously and accurately focus on the dynamics of key obstacles and wind fields, avoiding decision-making biases caused by dependence on a single feature. The generated flight path has continuous curvature and no redundant steering, maintaining stable flight control even in sudden turbulent environments and reducing the frequency of attitude adjustment.

[0013] 4. This invention designs a state space incorporating energy perception (integrating real-time remaining energy information of the UAV) and a multi-objective composite reward function (including energy penalty terms, wind direction reward terms, etc.), making energy consumption optimization one of the core decision-making objectives. The energy penalty term incentivizes the UAV to reach the target before its energy is depleted, while the wind direction reward term guides the UAV to actively utilize tailwinds, reducing additional energy consumption caused by headwinds or frequent turns. While ensuring obstacle avoidance safety and path efficiency, this optimization mechanism significantly reduces UAV energy consumption, effectively extending its endurance and meeting the needs of long-term operations in complex urban environments.

[0014] 5. This invention employs a dynamic landmark generation mechanism, setting landmarks along the current optimal path at a fixed ratio and dynamically updating them as better paths emerge. This provides the agent with phased guidance signals, effectively alleviating the sparse reward problem commonly encountered in reinforcement learning. Simultaneously, it combines a multi-stage decaying greedy exploration strategy to prevent the model from getting trapped in local optima or experiencing low exploration efficiency. The synergistic effect of these two strategies significantly accelerates the convergence speed of the SMA-DDQN network, improves the model's exploration efficiency in complex wind field and obstacle-coupled environments, and shortens the algorithm's training cycle. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0016] Figure 1 This is a schematic diagram of a building wind field CFD simulation constructed in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram illustrating the influence of wind speed in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the state space structure defined in the embodiments of the present invention.

[0019] Figure 4This is a schematic diagram of the action space defined in the embodiments of the present invention.

[0020] Figure 5 This is a diagram of the SMA-DDQN network structure proposed in this invention.

[0021] Figure 6 This is a flowchart of the improved DDQN algorithm of this invention.

[0022] Figure 7 This is a schematic diagram of the dynamic landmark generation mechanism used in this embodiment of the invention.

[0023] Figure 8 This is a comparison chart of algorithm reward convergence curves under complex environments in embodiments of the present invention.

[0024] Figure 9 This is a comparison curve of energy consumption under complex environments in an embodiment of the present invention.

[0025] Figure 10 This is a visualization of the path planning results under complex environment A in this embodiment of the invention.

[0026] Figure 11 This is a visualization of the path planning results under complex environment B in this embodiment of the invention.

[0027] Figure 12 This is a wind speed and direction distribution map of a real urban scene in an embodiment of the present invention.

[0028] Figure 13 This is a path planning result diagram of a real urban scene in an embodiment of the present invention. Detailed Implementation

[0029] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0030] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0031] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0032] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0033] Example 1: This embodiment provides a UAV flight control method based on Mamba and dual-channel attention mechanisms under wind field disturbances, including the following steps: S1. Based on computational fluid dynamics, a dynamic environmental model reflecting the disturbance of the wind field around the building is constructed; Specifically, a geometric model based on the actual urban building layout was established using CFD software. Velocity inlet boundary conditions were used at the computational domain inlet, and the incoming wind speed was set. Steady-state calculations were performed using the Reynolds-averaged Navier-Stokes (RANS) equations and the standard k-ε turbulence model to solve the flow field around the buildings, obtaining three-dimensional wind speed vector data for each spatial point within the computational domain. This allowed for the accurate simulation of typical wind field disturbances, such as the high-pressure zone on the windward side of the building, the vortex shedding zone on the leeward side, and the acceleration zone on the crosswind side. Figure 1 As shown.

[0034] Meanwhile, this invention comprehensively considers the dynamic characteristics of aircraft, wind field interference effects, and energy consumption laws to construct a multi-physics coupled energy consumption model to simulate the dynamic decay process of UAV battery capacity. The energy consumption model is as follows: in, These represent the base power consumption and aerodynamic drag power, respectively. It is the time step It is the steering energy consumption coefficient; This refers to the steering angle. Base power consumption is the minimum power required for the drone to maintain its internal components and fly; the aerodynamic drag power is: in, Let represent air density, drag coefficient, and equivalent frontal area, respectively; considering the coupling effect of wind speed and flight direction, As shown in equation 3: in, These represent the drone's speed and the wind speed, respectively. The angle between the wind direction and the drone, These represent the wind direction angle and the drone's flight angle, respectively. Figure 2 As shown.

[0035] S2. Design a reinforcement learning decision framework. Based on the dynamic environment model, define the state space, action space, and multi-objective composite reward function of the UAV. The state space includes the UAV state, obstacle information, target point information, and wind field information. The action space is a discrete flight direction command. The multi-objective composite reward function includes path reward, energy penalty, wind direction reward, and landmark reward. 1) The state space of the UAV: ​​The state space is the set of all possible states. The state space is designed as follows: , Information representing obstacles Classified according to the detection angle of the laser sensor Information representing the target point includes the straight-line distance and direction from the drone to the target point; Information representing the wind field includes its direction and intensity; information representing the drone's energy, such as... Figure 3 As shown.

[0036] 2) Action Space of UAVs: Considering the motion characteristics of fixed-wing UAVs and wind information, a discrete action space was designed. ,like Figure 4 As shown.

[0037] After selecting an action based on the current position, the formula for updating the next position is as follows: in, Indicates the speed of the drone. and This represents the wind speed along the x and y axes. Indicates a time interval.

[0038] 3) Multi-objective composite reward function: In reinforcement learning, the reward function is a core mechanism. It defines the immediate reward obtained by the agent when performing an action in a specific state, guiding the agent to optimize the policy by maximizing the cumulative reward. The reward function of this invention is as follows: The completion reward is the reward for the drone to complete the mission or fail. A reward is given for completing the mission, while a penalty is given for failing the mission, as shown in equation (6): The energy penalty is based on the remaining energy, which encourages the drone to reach the target point before its energy is depleted, as shown in equation (7): in, It is the remaining energy; It is the maximum energy.

[0039] Greedy rewards are based on changes in the distance and direction between the agent and the target, encouraging the drone to move toward the target, as shown in equation (8): in, , , and These represent the change in distance between the current position and the previous step, and the change in the angle between the current position and the target direction, respectively. This indicates the distance a drone travels in a fixed amount of time.

[0040] Path reward is a reward based on path length after completing a task, encouraging drones to reach their destinations via shorter paths, as shown in equation (9): in, It is the distance of the flight path; It is the maximum distance from the starting point to the end point.

[0041] The wind direction reward is based on the angle between the current direction and the wind direction, which encourages the drone to move with the wind, as shown in equation (10): in, It is the angle between the agent's current direction and the wind direction, defined as follows: .

[0042] The reward for passing specific landmarks encourages drones to continuously correct their optimal routes, as shown in equation (11): in, ,and It represents the Euclidean distance from the landmark to the destination. The closer the landmark is to the destination, the more important it is. ,and This represents the number of times a landmark has been visited. As the number of visits increases, the decay factor gradually decreases, thus reducing the reward for landmarks that have been visited multiple times.

[0043] S3. Construct a deep reinforcement learning network SMA-DDQN. This network uses a reinforcement learning decision framework as a constraint, extracts the spatial and temporal features of the input state in parallel through a dual-channel attention module, models the temporal dependencies of long sequences using a Mamba module, and outputs the optimal action value function. Specifically, the SMA-DDQN network adopts a dual-Q network (DDQN) architecture, and its feature extraction layer includes a dual-channel attention module and a Mamba module. The dual-channel attention module calculates spatial attention and temporal attention weights in parallel, adaptively weights the features of the input state, and focuses on key obstacles and wind field features. like Figure 5 As shown, the temporal attention channel employs a joint representation of historical state sequences and the current state, dynamically generating temporal weights through temporal correlation analysis; the spatial attention mechanism constructs state feature dimensional information and dynamically generates spatial weights through interaction calculations between features. This dual-channel attention mechanism, through adaptive feature selection and time-point focusing, achieves precise capture of key decision-making elements, thereby effectively enhancing the pertinence and robustness of obstacle avoidance decisions.

[0044] The Mamba module, as a selective state-space model, introduces a selective input-dependency mechanism and hardware-aware parallel computing, playing a core role in temporal feature fusion and dynamic modeling. Compared to deep reinforcement learning using LSTM, whose fixed gating mechanism is less adaptable to complex dynamic scenarios, Mamba, through its selective input-dependency mechanism, can adjust its temporal modeling strategy in real time according to the state, thus adapting more flexibly to dynamically changing path environments. Its SSM framework can more effectively capture long-distance spatiotemporal dependencies while dynamically filtering noise interference and indefinitely remembering useful information, while LSTM struggles to accurately retain long-term useful information and is limited by the gradient vanishing problem, resulting in gradual performance degradation in long sequence modeling.

[0045] To address the inherent information loss problem in state-space models during temporal feature compression, a joint representation construction method is employed. This method concatenates Mamba predictions and latent states along the feature dimension to construct a complete joint representation space. This effectively offsets the information loss in state-space models and alleviates the loss of discernibility caused by representation dimension compression. The final action is obtained after passing through a fully connected layer, such as... Figure 5 As shown. By modeling the relationship between obstacle distance and value function, safety constraints and prediction loss are integrated into the reinforcement learning objective, avoiding the difficulty of learning obstacle avoidance strategies in traditional DDQN. This invention proposes a multi-objective loss function, as shown in equation (12).

[0046] In the above formula, It is the loss of standard DDQN; ,in This represents the highest output of the action value function. This represents the safety factor, which causes the agent to lower its value assessment in dangerous situations, preventing the agent from risking approaching obstacles in pursuit of higher value. ,in, Indicates the predicted next state. It is the true next state, and combined with Mamba, it enables the network to learn accurate predictions of state representations. Based on the above improvements, such as... Figure 6 As shown.

[0047] S4. The deep reinforcement learning network SMA-DDQN is trained using a dynamic landmark generation mechanism and a multi-stage decay greedy exploration strategy to obtain a path planning model; Specifically, waypoints are defined as spatial location markers with navigational value, serving as important references for optimizing flight paths and adjusting flight strategies, such as... Figure 7 The sequence of flight trajectory points is as follows: , It is the first A trajectory point, containing coordinates and timestamp The spacing between adjacent trajectory points is determined by both the kinematic model and wind field disturbances. Landmark points are generated as shown in equation (13): in, It is the interval distance; That is, the Euclidean distance between all adjacent trajectory points, such as Figure 1 As shown. Furthermore, to avoid repeatedly exploring known path areas and improve planning efficiency, the current path is optimized... , replace .

[0048] To prevent getting trapped in local optima during the search path, a decay strategy was designed, as shown in equation (14): In the above formula, This represents the threshold for switching between each stage; The exponential decay coefficient is... The slope represents the linear decay.

[0049] S5. Input the real-time environmental status into the path planning model to obtain flight control commands, and drive the UAV to complete path tracking and obstacle avoidance in a wind-turbulent environment.

[0050] In simulation environment and real scene ( Figure 12 The test was conducted in a certain area of ​​Dalian City. The real-time status was input into the trained path planning model, and the model output action commands to control the flight of the drone.

[0051] Based on the above scheme, the experimental parameters were set as follows: The experiment was conducted in a simulated environment of size 1200*1200. The drone would restart if it reached the finish line, flew out of bounds, touched an obstacle, exceeded the maximum number of steps, or ran out of energy. All experiments were conducted in a simulated environment using an NVIDIA Quadro RTX 4060 GPU, and the framework used was the PyTorch architecture. Several popular reinforcement learning algorithms were selected for comparison: DQN, DDQN, DDQN-attention (with an attention mechanism), DDQN-LSTM (with an LSTM mechanism), LA-DDQN (combining LSTM and attention mechanisms), and the proposed SMA-DDQN.

[0052] The simple environment test aimed to evaluate the algorithm's path planning capabilities in basic scenarios, including three scenarios: no obstacles, large single obstacle, and large double obstacle. 800 tests were performed in each scenario, and the results are shown in Table 1.

[0053] Table 1 Performance Comparison under Simple Environments In simple obstacle scenarios, SMA-DDQN exhibits optimal path planning efficiency, generating smooth and direct flight paths. This advantage is primarily attributed to the dynamic landmark guidance mechanism and the design of the composite reward function, enabling the UAV to utilize wind field dynamics for trajectory optimization. As environmental complexity increases, SMA-DDQN demonstrates superior environmental adaptability. Traditional DQN and DDQN exhibit significant performance limitations in simple environments, struggling to effectively handle the coupling effects between obstacles and wind fields. LSTM modules improve the temporal modeling capabilities of traditional DDQN, but still have limitations in handling complex wind field coupling effects; while attention mechanisms enhance spatial feature extraction, they cannot effectively model long-term temporal dependencies; and the performance of LA-DDQN, which combines both, is inferior to SMA-DDQN. This indicates that Mamba captures long-sequence wind field dynamics and, combined with an attention mechanism, focuses on key obstacle features, achieving efficient modeling of spatiotemporal features.

[0054] The complex environment test included two different dynamic wind field and obstacle configurations, focusing on evaluating the algorithm's robustness and adaptability under dynamic disturbances. Three sets of experiments were conducted for each environment, with 2400 experiments per set. The results are shown in Table 2 and... Figures 8-11 As shown.

[0055] The algorithms showed significant differences in training convergence. For example... Figure 8 As shown, all algorithms exhibited varying degrees of fluctuation in the initial training phase. SMA-DDQN demonstrated the fastest convergence speed, stabilizing after 500 iterations and eventually reaching equilibrium. DQN and DDQN exhibited volatility throughout the entire training period, failing to fully stabilize even in the later stages, indicating that their underlying algorithm architectures struggled to effectively handle the complex coupling effects between dynamic wind fields and obstacles. Notably, LA-DDQN showed slightly slower convergence in the early stages, and its reward value after stabilization was lower and more volatile than that of SMA-DDQN.

[0056] In energy consumption, such as Figure 9 As shown, SMA-DDQN has lower energy consumption and significantly smaller fluctuations than the comparative algorithms, demonstrating stable energy-saving performance. In contrast, traditional algorithms exhibit low energy efficiency in dynamic wind fields due to path redundancy and frequent turning. They not only show greater energy consumption volatility but also exhibit obvious fluctuation characteristics. Furthermore, while LA-DDQN's energy consumption is significantly lower than DQN and DDQN, its mean energy consumption and fluctuation amplitude are still higher than SMA-DDQN.

[0057] Table 2 shows a comparison of performance metrics for different algorithms, with SMA-DDQN outperforming all others. The integration of Attention and LSTM results in better performance than traditional algorithms, indicating that the attention mechanism and LSTM are highly adaptable to complex environments. Furthermore, LA-DDQN significantly outperforms DDQN using Attention and LSTM in both success rate and average number of steps, but its overall performance is still inferior to SMA-DDQN. This suggests that the combined module is insufficient to fully capture the dynamic coupling of the wind field and obstacles, further highlighting the contribution of the Mamba architecture and dynamic guidance strategy to improving overall performance.

[0058] Table 2 Performance Comparison under Complex Environments exist Figure 10 and Figure 11The path planning visualization results revealed that the DQN path exhibits significant redundancy and frequent turns, especially showing an "oscillating flight" phenomenon in complex wind fields. The DDQN path shows some improvement, but still exhibits a hesitant, jagged trajectory in narrow channels. DDQN-Attention enhances the identification of key obstacles but does not completely solve the wind field coupling problem. DDQN-LSTM improves long-term planning but is insufficiently responsive to wind field disturbances. LA-DDQN, after fusing LSTM and attention mechanisms, improves both trajectory smoothness and path length; however, it still exhibits slight lateral oscillations in strong turbulence areas and fails to fully utilize the downwind area to reduce energy consumption. In contrast, the SMA-DDQN path is the shortest and has continuous curvature, avoiding obstacles in advance and significantly favoring the downwind area, indicating that it achieves efficient and low-energy navigation in wind field-obstacle coupling environments through the Mamba module and dynamic guidance strategy.

[0059] To further verify the effectiveness of the proposed algorithm in real-world urban environments with wind field fluid dynamics data, a real urban market scene in Dalian, Liaoning Province, was selected. Air at 25°C under standard atmospheric pressure was chosen as the fluid, and the algorithm was used to simulate its path. Figure 13 As shown in the path results diagram, the path is smooth overall with no obvious redundant turns. The algorithm effectively combines wind field information and obstacle avoidance. Furthermore, in densely built-up areas, the path detours in advance to avoid emergency obstacle avoidance, reflecting Mamba's predictive ability for long-term time series. The path remains smooth even in narrow passages, indicating that dynamic landmark guidance and attention mechanisms synergistically optimize local decision-making. The path responds quickly to sudden obstacles, and the safety loss function forces the UAV to maintain a safe distance.

[0060] Example 2: This embodiment provides a drone flight control system based on the Mamba and dual-channel attention mechanism under wind field disturbance, including: The dynamic environment modeling module, based on computational fluid dynamics models, constructs dynamic environment models that reflect the disturbances of the wind field around a building. The decision framework design module designs a reinforcement learning decision framework. Based on the dynamic environment model, it defines the state space, action space, and multi-objective composite reward function of the UAV. The state space includes the UAV state, obstacle information, target point information, and wind field information. The action space is a discrete flight direction command. The multi-objective composite reward function includes path reward, energy penalty, wind direction reward, and landmark reward. The deep reinforcement learning network construction module constructs the SMA-DDQN deep reinforcement learning network. This network uses the reinforcement learning decision framework as a constraint, extracts the spatial and temporal features of the input state in parallel through a dual-channel attention module, and uses the Mamba module to model the temporal dependencies of long sequences to output the optimal action value function. The network training module employs a dynamic landmark generation mechanism and a multi-stage decaying greedy exploration strategy to train the deep reinforcement learning network SMA-DDQN, thereby obtaining a path planning model. The online flight control execution module inputs real-time environmental conditions into the path planning model to obtain flight control commands, driving the UAV to complete path tracking and obstacle avoidance in windy and turbulent environments.

[0061] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0062] Example 3: An electronic device is provided for running the aforementioned "A UAV Flight Control Method Based on Mamba and Dual-Channel Attention Mechanism under Wind Disturbance". The electronic device includes: a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S5 of the method described in Embodiment 1. The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0063] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S5 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0064] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0065] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0066] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for UAV flight control based on Mamba and dual-channel attention mechanisms under wind field disturbance, characterized in that, Includes the following steps: Based on computational fluid dynamics models, a dynamic environmental model reflecting the disturbance of the wind field around a building is constructed. The design of the reinforcement learning decision framework defines the state space, action space, and multi-objective composite reward function of the UAV based on the dynamic environment model. The state space includes the UAV state, obstacle information, target point information, and wind field information. The action space is a discrete flight direction command. The multi-objective composite reward function includes path reward, energy penalty, wind direction reward, and landmark reward. A deep reinforcement learning network SMA-DDQN is constructed. This network uses a reinforcement learning decision framework as a constraint, extracts the spatial and temporal features of the input state in parallel through a dual-channel attention module, and uses a Mamba module to model the temporal dependencies of long sequences to output the optimal action value function. The deep reinforcement learning network SMA-DDQN was trained using a dynamic landmark generation mechanism and a multi-stage decay greedy exploration strategy to obtain a path planning model. By inputting real-time environmental conditions into the path planning model, flight control commands are obtained, driving the UAV to complete path tracking and obstacle avoidance in windy and turbulent environments.

2. The UAV flight control method based on Mamba and dual-channel attention mechanism under wind field disturbance as described in claim 1, characterized in that, When constructing the dynamic environment model, the Reynolds-averaged Navier-Stokes equations coupled with the k-ε turbulence model are used to simulate the wind field. Specifically, the wind field control equations are solved to obtain the wind speed vector distribution at each point in space, thereby characterizing the turbulent wind field disturbances caused by the separation of building wakes and corner flows.

3. The UAV flight control method based on Mamba and dual-channel attention mechanism under wind field disturbance as described in claim 1, characterized in that, When constructing the dynamic environment model, a UAV energy consumption model that considers wind field interference is simultaneously established. The specific model is as follows: in, For steering energy consumption coefficient, This is the absolute value of the steering angle. It is the time step. These represent the base power consumption and aerodynamic drag power, respectively.

4. The UAV flight control method based on Mamba and dual-channel attention mechanism under wind field disturbance as described in claim 1, characterized in that, The multi-objective composite reward function The specific form is: Among them, energy penalty item , For the remaining energy, For maximum energy, These are the weighting coefficients; Greedy Rewards , , as well as These represent the change in distance between the current position and the previous step, and the change in the angle between the current position and the target direction, respectively. This indicates the distance the drone travels in a fixed amount of time. These are the weighting coefficients; Path Rewards , This is the actual flight path length. This represents the maximum possible distance from the starting point to the ending point. These are the weighting coefficients; Wind direction bonus , The angle between the drone's heading and the wind direction. These are the weighting coefficients; Complete the reward items , These are the weighting coefficients; Road sign rewards , , This represents the Euclidean distance from the landmark to the destination. , Indicates the number of times a landmark has been visited. These are the weighting coefficients.

5. The UAV flight control method based on Mamba and dual-channel attention mechanism under wind field disturbance as described in claim 1, characterized in that, The dual-channel attention module performs adaptive weighted processing on the features of the input state by computing spatial attention weights and temporal attention weights in parallel, thereby focusing on key obstacles and wind field features. The Mamba module, acting as a selective state-space model, receives attention-weighted feature sequences and utilizes its own input-dependent selective mechanism to model the dynamics of long-sequence wind field disturbances. Its selective state-space model is as follows: ,in This parameter is determined by the current input. Dynamic generation via linear projection ensures that the model can adapt to temporal dynamics.

6. The UAV flight control method based on Mamba and dual-channel attention mechanism under wind field disturbance as described in claim 1, characterized in that, The dynamic landmark generation mechanism is as follows: Along the currently known optimal path, with the total length of that path fixed ratio Road signs are set at intervals, where the distance between road signs is... When a shorter path is found, the landmark sequence is updated with the new optimal path.

7. The UAV flight control method based on Mamba and dual-channel attention mechanism under wind field disturbance as described in claim 1, characterized in that, Multi-stage decaying greedy exploration strategy Exploration rate The attenuation rule is defined by the following piecewise function: in, , These are the weighting coefficients. Represents the number of training rounds. This represents the threshold for switching between each stage. The exponential decay coefficient is... The slope represents the linear decay.

8. A UAV flight control system based on Mamba and dual-channel attention mechanisms under wind field disturbance, characterized in that, include: The dynamic environment modeling module, based on computational fluid dynamics models, constructs dynamic environment models that reflect the disturbances of the wind field around a building. The decision framework design module designs a reinforcement learning decision framework. Based on the dynamic environment model, it defines the state space, action space, and multi-objective composite reward function of the UAV. The state space includes the UAV state, obstacle information, target point information, and wind field information. The action space is a discrete flight direction command. The multi-objective composite reward function includes path reward, energy penalty, wind direction reward, and landmark reward. The deep reinforcement learning network construction module constructs the SMA-DDQN deep reinforcement learning network. This network uses the reinforcement learning decision framework as a constraint, extracts the spatial and temporal features of the input state in parallel through a dual-channel attention module, and uses the Mamba module to model the temporal dependencies of long sequences to output the optimal action value function. The network training module employs a dynamic landmark generation mechanism and a multi-stage decay greedy exploration strategy to train the deep reinforcement learning network SMA-DDQN, thereby obtaining a path planning model. The online flight control execution module inputs real-time environmental conditions into the path planning model to obtain flight control commands, driving the UAV to complete path tracking and obstacle avoidance in windy and turbulent environments.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the UAV flight control method based on Mamba and dual-channel attention mechanism under wind field disturbance as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the UAV flight control method based on the Mamba and dual-channel attention mechanism under wind field disturbance as described in any one of claims 1-7.

Images