Coaxial double-propeller multi-rotor unmanned aerial vehicle stability control method based on aerodynamic disturbance optimization

Abstract

The invention discloses a coaxial double-propeller multi-rotor unmanned aerial vehicle stability control method based on aerodynamic disturbance optimization, and the method comprises the following steps: S1, collecting flight data, preprocessing the data, and generating an aerodynamic disturbance training data set; s2, based on the aerodynamic disturbance training data set, constructing an improved Fourier neural operator network model; s3, optimizing parameters of the model by adopting an improved chaos grey wolf optimization algorithm; s4, embedding the optimized model into an unmanned aerial vehicle flight control system; s5, according to the predicted aerodynamic disturbance data, a model prediction control algorithm is adopted to adjust the thrust and the attitude; and S6, verifying the stability of the control method through a simulation test and an actual flight test. The optimization method for improving the flight stability of the unmanned aerial vehicle in the complex aerodynamic environment is described in detail, and the method has the advantages of being high in calculation efficiency, accurate in aerodynamic disturbance prediction and high in control system robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) control, and in particular to a coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization. Background Art

[0002] Coaxial twin-propeller multi-rotor drones (UAVs) are widely used in military reconnaissance, logistics and transportation, agricultural plant protection, and urban air mobility due to their high thrust-to-weight ratio, compact structure, and high maneuverability. However, these drones are significantly affected by complex aerodynamic disturbances during flight, particularly under conditions of high speed, sudden wind changes, turbulent environments, and ground effect. These conditions pose significant challenges to attitude stability and control accuracy. Traditional control methods, such as PID control, robust control, and model predictive control, can compensate for external disturbances to a certain extent. However, due to the unique aerodynamic characteristics of coaxial twin-propeller systems, these methods often struggle to maintain stability under extreme disturbances, resulting in reduced UAV flight control reliability.

[0003] In the existing technology, the research on aerodynamic interference optimization mainly focuses on two aspects: one is the flow field analysis method based on computational fluid dynamics (CFD) simulation, and the other is the control optimization strategy based on data-driven. CFD simulation can finely model the aerodynamic characteristics of the coaxial twin-propeller system, including the wake effect between the upper and lower blades, vortex separation, and the influence of the environmental wind field. However, CFD simulation is computationally intensive and usually requires high-performance computing resources. Even with efficient numerical methods, it is difficult to meet the needs of real-time flight control. In addition, the generalization ability of the CFD model in the actual flight environment is weak, and it cannot adapt to the rapid changes under different flight conditions, making it difficult for the control strategy based on CFD modeling to meet the needs of dynamic compensation.

[0004] On the other hand, data-driven control optimization strategies mainly rely on deep learning or reinforcement learning methods, which are trained through large amounts of flight data to establish disturbance prediction models. However, existing neural network models such as convolutional neural networks, long short-term memory networks, and variational autoencoders have certain limitations when processing high-dimensional aerodynamic disturbance data. These methods usually have difficulty capturing global information when learning complex flow field relationships, resulting in insufficient accuracy in aerodynamic disturbance prediction. In addition, data-driven methods usually require a large amount of labeled data, and the cost of obtaining disturbance data of UAVs in different flight environments is high, especially under adverse weather conditions. The safety and feasibility of data collection are restricted, further limiting the promotion and application of data-driven models.

[0005] To address these issues, the Fourier neural operator, a novel deep learning method, has recently demonstrated superior performance in solving complex partial differential equations. The Fourier neural operator can directly learn function mappings in the frequency domain and extract global features through Fourier transforms, thereby more efficiently handling the nonlinear characteristics of aerodynamic disturbances. Compared to traditional neural networks, the Fourier neural operator has stronger generalization capabilities and can achieve high-precision predictions at a lower computational cost. Therefore, using the Fourier neural operator to construct aerodynamic disturbance prediction models can effectively overcome the computational burden of the Fourier neural operator modeling, while also improving the adaptability of data-driven models to diverse flight conditions. However, the standard Fourier neural operator still has certain limitations when processing high-dimensional aerodynamic flow field data. For example, it is less able to capture local disturbance features, which can lead to prediction errors in complex environments with strong turbulence effects. Therefore, optimizing the Fourier neural operator structure to balance global flow field modeling capabilities with accurate prediction of local disturbance features remains an urgent problem.

[0006] Furthermore, the performance of neural network models is highly dependent on the choice of hyperparameters, such as the number of network layers, Fourier frequency domain weights, and learning rate. Traditional hyperparameter optimization methods primarily include grid search, random search, and Bayesian optimization. However, these methods suffer from low search efficiency and are prone to getting stuck in local optima in high-dimensional parameter spaces. The Gray Wolf Optimization (GWA), a swarm intelligence optimization algorithm, offers significant advantages in optimizing complex nonlinear systems due to its excellent global search capabilities. However, the traditional GWA still suffers from low search efficiency during the optimization process, particularly in the late convergence phase, where its local search capability is insufficient, which can lead to premature convergence. Furthermore, the GWA's poor balance between exploration and exploitation can result in significant fitness fluctuations in certain complex optimization tasks. Therefore, improving the GWA to enhance its global search and local optimization capabilities, making it more suitable for neural network parameter optimization, is a worthy research topic.

[0007] In summary, existing CFD simulation-based methods are computationally intensive and difficult to meet real-time requirements. Traditional neural networks lack precision and generalization in aerodynamic disturbance modeling. While the standard Fourier neural network operator has strong global modeling capabilities, it still lacks the ability to extract local disturbance features. The Gray Wolf optimization algorithm has low search efficiency when optimizing neural network parameters and is prone to falling into local optima. To address these issues, the present invention proposes a stable control method for a coaxial twin-propeller multi-rotor unmanned aerial vehicle (UAV) based on an improved Fourier neural network operator and an improved chaotic Gray Wolf optimization algorithm. By constructing a multi-resolution Fourier neural network operator and improving its frequency domain mapping, it not only models global flow field information but also enhances its ability to learn local disturbances, thereby improving aerodynamic disturbance prediction accuracy. Furthermore, the Fourier neural network operator hyperparameters are optimized using the chaotic Gray Wolf optimization algorithm (F4). The detection wolf mechanism and Levy flight step size strategy are introduced to enhance the Gray Wolf optimization algorithm's global search and local optimization capabilities, enabling more efficient optimization of neural network parameters. In addition, an adaptive control algorithm based on the combination of MPC and sliding mode control can realize dynamic adjustment of the UAV thrust distribution and attitude compensation, thereby improving the stability and control accuracy of the UAV in complex aerodynamic environments.

[0008] This invention not only solves the problem of high CFD calculation cost, but also improves the accuracy and generalization ability of the neural network model in aerodynamic disturbance prediction. At the same time, by improving the Gray Wolf optimization algorithm, the training efficiency and performance optimization ability of the model are improved, and ultimately the precise and stable control of the coaxial twin-propeller multi-rotor UAV in complex environments is achieved.

[0009] Therefore, how to provide a coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization is a problem that technicians in this field urgently need to solve. Summary of the Invention

[0010] One purpose of the present invention is to propose a coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization. The present invention describes in detail an optimization method for improving the flight stability of the UAV in a complex aerodynamic environment, which has the advantages of high computational efficiency, accurate prediction of aerodynamic disturbances, and strong robustness of the control system.

[0011] According to an embodiment of the present invention, a coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization includes the following steps:

[0012] S1. Collect flight data, pre-process the data, and generate an aerodynamic disturbance training dataset;

[0013] S2. Based on the aerodynamic disturbance training dataset, an improved Fourier neural operator network model is constructed;

[0014] S3. Using the improved chaotic grey wolf optimization algorithm, the parameters of the improved Fourier neural operator network model are globally searched and locally optimized to update the parameters;

[0015] S4. Embed the optimized improved Fourier neural operator network model into the UAV flight control system to receive flight data in real time and predict aerodynamic disturbance data;

[0016] S5. Based on the real-time predicted aerodynamic disturbance data, the model predictive control algorithm is used to achieve dynamic adjustment of the thrust of each rotor and compensation of the flight attitude;

[0017] S6. Verify the stability and real-time performance of the control method under various aerodynamic interference conditions through simulation tests and actual flight tests.

[0018] The coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization according to claim 1 is characterized in that S2 specifically includes:

[0019] S21. Define the input parameter space and assume that the input data for the UAV aerodynamic disturbance prediction is u(x,t);

[0020] S22. Aiming at the complex flow field disturbance characteristics of coaxial twin-propeller multi-rotor UAVs, a multi-resolution Fourier neural operator is used to improve the frequency domain mapping method of the traditional Fourier neural operator network. Specifically,

[0021] S221, the first layer, the local Fourier transform layer, uses the local window Fourier transform to extract the local flow field disturbance characteristics under different scale windows. is the local Fourier transform operator, then the local spectrum under different scale windows is expressed as:

[0022]

[0023] Where l represents the number of local windows, is the Fourier frequency domain representation under the local window;

[0024] S222, the second layer, the global Fourier transform layer, represents the local spectrum Perform weighted global fusion:

[0025]

[0026] Among them, α l is the learnable fusion weight;

[0027] S223, the third layer, the adaptive spectrum filtering layer, sets W(k) as the frequency domain weight matrix, uses adaptive spectrum filtering to dynamically adjust the weights of high and low frequency information:

[0028]

[0029] Where b is the bias term, and W(k) is learned using the dynamic spectral attention mechanism;

[0030] S224, the fourth layer, the inverse Fourier transform layer, restores the frequency domain information to the physical space through the inverse Fourier transform:

[0031]

[0032] in, is the inverse Fourier transform, d(x,t) is the predicted data of the UAV aerodynamic disturbance;

[0033] S23. Construct cross-scale perturbations, including:

[0034] S231, low-frequency disturbance compensation, uses a low-pass filter to extract low-frequency disturbance features and perform nonlinear transformation:

[0035] d low (x,t)=ReLU(W low d(x,t)+b low );

[0036] Among them, ReLU is the activation function, W low is the low-frequency weight matrix, b low is the low frequency compensation term;

[0037] S232, high-frequency disturbance enhancement, uses a high-pass filter to extract high-frequency disturbance information, combined with adaptive gain adjustment:

[0038] d high (x,t)=tanh(W high d(x,t)+b high );

[0039] Among them, tanh(·) is the hyperbolic tangent function, W high is the high-frequency weight matrix, b high is the high frequency compensation term;

[0040] S233, cross-scale disturbance fusion, combining low-frequency disturbance compensation and high-frequency disturbance enhancement for disturbance enhancement:

[0041] d final (x,t)=β low d low (x,t)+β high d high (x,t);

[0042] Among them, β low and β high is the learnable fusion weight;

[0043] S24. Define the output of the improved Fourier neural operator model and set the final aerodynamic disturbance prediction result as d final (x, t), the complete mapping relationship can be expressed as:

[0044]

[0045] in, is the trained Fourier neural operator, θ is the optimized model parameter, u(x,t) is the input data, d final (x, t) is the final aerodynamic disturbance prediction output value.

[0046] The coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization according to claim 1 is characterized in that S3 specifically includes:

[0047] S31. Initialize the parameters of the improved gray wolf optimization algorithm, set the optimization group size N and the maximum number of iterations T, and set the four wolf individual position vectors X to represent the parameters to be optimized of the improved Fourier neural operator network model, where:

[0048] X i ={W(k),b,W low ,b low ,W high ,b high ,β low ,β high};

[0049] Initialize the initial positions of the gray wolf group to be evenly distributed in the search space, randomly initialize the weight matrix, set four wolf individuals, the main wolf X α 、Deputy Wolf X β , Second Deputy Wolf X δ , Detection Wolf X θ ;

[0050] S32. Calculate the fitness function and set the fitness function J to measure the error of the model parameters in predicting aerodynamic disturbances. The objective function is:

[0051]

[0052] Among them, MSE(·) is the mean square error loss function, SpectralError(·) is the spectral error loss function, d final is the aerodynamic disturbance prediction output value of the improved Fourier neural operator network model, d true is the real aerodynamic disturbance data in the aerodynamic disturbance training dataset, d final ,d trueThe Fourier spectrum of λ1, λ2, and λ3 are weight factors, and Diversity(X) is a population diversity maintenance item to prevent the algorithm from premature convergence.

[0053] S33, using chaotic mapping to generate the initial position of the population

[0054] S34, update the position of individual gray wolves, calculate the dominant wolf X of the current population α 、Deputy Wolf X β , Second Deputy Wolf X δ , Detection Wolf X θ , the first three wolves search according to the classic gray wolf optimization algorithm:

[0055]

[0056] Among them, the main wolf X α 、Deputy Wolf X β , Second Deputy Wolf X δ Will be updated with each iteration according to the new population fitness value, is the position of the gray wolf in generation t, c1, c2, c3 are the dynamic adjustment control factors;

[0057] Detecting wolves for global exploration:

[0058]

[0059] in, is the position of the individual wolf detected in the tth generation, Levy(λ) is a random variable generated from the Levy distribution, γ is the global search weight, X rand is a random individual;

[0060] S35, perform convergence judgment and parameter update, calculate the optimal fitness value J of each generation α , if J α When convergence or the maximum number of iterations T is reached, the current optimal solution X is output. α As the final optimization parameter, if it does not converge, it returns to continue iterative optimization, and the final updated parameter is expressed as:

[0061] θ opt ={W opt (k),b opt ,W low,opt ,b low,opt ,W high,opt ,b high,opt ,β low,opt ,β high,opt};

[0062] S36, using the optimized parameter θ optRe-adjust the improved Fourier neural operator network model to achieve the optimal effect.

[0063] The beneficial effects of the present invention are:

[0064] (1) The present invention uses a multi-resolution Fourier neural operator to model aerodynamic disturbances. Compared with traditional neural network methods, it can directly learn disturbance features in the frequency domain. In combination with a local Fourier transform layer and an adaptive spectrum filter layer, it enhances the learning ability of local disturbance features, making aerodynamic disturbance prediction more accurate. Compared with aerodynamic modeling methods based on computational fluid dynamics (CFD), the present invention can reduce computing resource consumption and improve computing efficiency while ensuring high-precision prediction, meet the real-time prediction requirements of UAVs for aerodynamic disturbances in complex environments, and ensure rapid response of flight control.

[0065] (2) The present invention uses an improved chaotic gray wolf optimization algorithm to optimize the hyperparameters of the Fourier neural operator model, introduces a detection wolf mechanism to enhance global search capabilities, combines a random walk strategy with a gradient-guided learning strategy to improve local optimization capabilities, and uses a chaotic perturbation mechanism to enhance the diversity of the optimization process, so that the optimization algorithm has stronger search capabilities and convergence stability during training. Compared with traditional optimization algorithms, the present invention can effectively avoid the local optimal problem in the optimization process, improve the adaptability of the neural operator model in different flight environments, ensure that the aerodynamic disturbance prediction can maintain stable accuracy under different flight conditions, and improve the generalization ability and robustness of the UAV system.

[0066] (3) The present invention adopts an adaptive control strategy based on the combination of model predictive control and sliding mode control, in which model predictive control is used to calculate the optimal rotor thrust distribution strategy, and sliding mode control is used to compensate for attitude deviations caused by complex aerodynamic disturbances. The combination of the two can ensure that the UAV maintains stable flight under complex aerodynamic disturbance conditions such as different wind speeds, air pressure changes, turbulent environments, and ground effects. Compared with the traditional PID control method, the control strategy of the present invention can adaptively adjust control parameters based on real-time predicted aerodynamic disturbance data, improve attitude control accuracy, enhance the flight stability of the UAV in complex environments, and ensure that the UAV can stably complete flight missions in various harsh environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0068] Figure 1 This is the overall flow chart of the coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization proposed in the present invention;

[0069] Figure 2 This is a schematic diagram of the optimization process of the improved chaotic gray wolf optimization algorithm proposed in the present invention. DETAILED DESCRIPTION

[0070] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0071] refer to Figure 1 and Figure 2 The stable control method of a coaxial twin-propeller multi-rotor UAV based on aerodynamic interference optimization includes the following steps:

[0072] S1. Collect flight data, pre-process the data, and generate an aerodynamic disturbance training dataset;

[0073] S2. Based on the aerodynamic disturbance training dataset, an improved Fourier neural operator network model is constructed;

[0074] S3. Using the improved chaotic grey wolf optimization algorithm, the parameters of the improved Fourier neural operator network model are globally searched and locally optimized to update the parameters;

[0075] S4. Embed the optimized improved Fourier neural operator network model into the UAV flight control system to receive flight data in real time and predict aerodynamic disturbance data;

[0076] S5. Based on the real-time predicted aerodynamic disturbance data, the model predictive control algorithm is used to achieve dynamic adjustment of the thrust of each rotor and compensation of the flight attitude;

[0077] S6. Verify the stability and real-time performance of the control method under various aerodynamic interference conditions through simulation tests and actual flight tests.

[0078] In this embodiment, S2 specifically includes:

[0079] S21. Define the input parameter space and assume that the input data for the UAV aerodynamic disturbance prediction is u(x,t);

[0080] S22. Aiming at the complex flow field disturbance characteristics of coaxial twin-propeller multi-rotor UAVs, a multi-resolution Fourier neural operator is used to improve the frequency domain mapping method of the traditional Fourier neural operator network. Specifically,

[0081] S221, the first layer, the local Fourier transform layer, uses the local window Fourier transform to extract the local flow field disturbance characteristics under different scale windows. is the local Fourier transform operator, then the local spectrum under different scale windows is expressed as:

[0082]

[0083] Where l represents the number of local windows, is the Fourier frequency domain representation under the local window;

[0084] S222, the second layer, the global Fourier transform layer, represents the local spectrum Perform weighted global fusion:

[0085]

[0086] Among them, α l is the learnable fusion weight;

[0087] S223, the third layer, the adaptive spectrum filtering layer, sets W(k) as the frequency domain weight matrix, uses adaptive spectrum filtering to dynamically adjust the weights of high and low frequency information:

[0088]

[0089] Where b is the bias term, and W(k) is learned using the dynamic spectral attention mechanism;

[0090] S224, the fourth layer, the inverse Fourier transform layer, restores the frequency domain information to the physical space through the inverse Fourier transform:

[0091]

[0092] in, is the inverse Fourier transform, d(x,t) is the predicted data of the UAV aerodynamic disturbance;

[0093] S23. Construct cross-scale perturbations, including:

[0094] S231, low-frequency disturbance compensation, uses a low-pass filter to extract low-frequency disturbance features and perform nonlinear transformation:

[0095] d low (x,t)=ReLU(W low d(x,t)+b low );

[0096] Among them, ReLU is the activation function, W low is the low-frequency weight matrix, b low is the low frequency compensation term;

[0097] S232, high-frequency disturbance enhancement, uses a high-pass filter to extract high-frequency disturbance information, combined with adaptive gain adjustment:

[0098] d high (x,t)=tanh(W high d(x,t)+b high);

[0099] Among them, tanh(·) is the hyperbolic tangent function, W high is the high-frequency weight matrix, b high is the high frequency compensation term;

[0100] S233, cross-scale disturbance fusion, combining low-frequency disturbance compensation and high-frequency disturbance enhancement for disturbance enhancement:

[0101] d final (x,t)=β low d low (x,t)+β high d high (x,t);

[0102] Among them, β low and β high is the learnable fusion weight;

[0103] S24. Define the output of the improved Fourier neural operator model and set the final aerodynamic disturbance prediction result as d final (x, t), the complete mapping relationship can be expressed as:

[0104]

[0105] in, is the trained Fourier neural operator, θ is the optimized model parameter, u(x,t) is the input data, d final (x, t) is the final aerodynamic disturbance prediction output value.

[0106] In this embodiment, S3 specifically includes:

[0107] S31. Initialize the parameters of the improved gray wolf optimization algorithm, set the optimization group size N and the maximum number of iterations T, and set the four wolf individual position vectors X to represent the parameters to be optimized of the improved Fourier neural operator network model, where:

[0108] X i ={W(k),b,W low ,b low ,W high ,b high ,β low ,β high};

[0109] Initialize the initial positions of the gray wolf group to be evenly distributed in the search space, randomly initialize the weight matrix, set four wolf individuals, the main wolf X α 、Deputy Wolf X β , Second Deputy Wolf X δ , Detection Wolf X θ ;

[0110] S32. Calculate the fitness function and set the fitness function J to measure the error of the model parameters in predicting aerodynamic disturbances. The objective function is:

[0111]

[0112] Among them, MSE(·) is the mean square error loss function, SpectralError(·) is the spectral error loss function, d final is the aerodynamic disturbance prediction output value of the improved Fourier neural operator network model, d true is the real aerodynamic disturbance data in the aerodynamic disturbance training dataset, d final ,d true The Fourier spectrum of λ1, λ2, and λ3 are weight factors, and Diversity(X) is a population diversity maintenance item to prevent the algorithm from premature convergence.

[0113] S33, using chaotic mapping to generate the initial position of the population

[0114] S34, update the position of individual gray wolves, calculate the dominant wolf X of the current population α 、Deputy Wolf X β , Second Deputy Wolf X δ , Detection Wolf X θ , the first three wolves search according to the classic gray wolf optimization algorithm:

[0115]

[0116] Among them, the main wolf X α 、Deputy Wolf X β , Second Deputy Wolf X δ Will be updated with each iteration according to the new population fitness value, is the position of the gray wolf in generation t, c1, c2, c3 are the dynamic adjustment control factors;

[0117] Detecting wolves for global exploration:

[0118]

[0119] in, is the position of the individual wolf detected in the tth generation, Levy(λ) is a random variable generated from the Levy distribution, γ is the global search weight, X rand is a random individual;

[0120] S35, perform convergence judgment and parameter update, calculate the optimal fitness value J of each generation α , if J αWhen convergence or the maximum number of iterations T is reached, the current optimal solution X is output. α As the final optimization parameter, if it does not converge, it returns to continue iterative optimization, and the final updated parameter is expressed as:

[0121] θ opt ={W opt (k),b opt ,W low,opt ,b low,opt ,W high,opt ,b high,opt ,β low,opt ,β high,opt};

[0122] S36, using the optimized parameter θ opt Re-adjust the improved Fourier neural operator network model to achieve the optimal effect.

[0123] Example:

[0124] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a stable flight experiment of a coaxial twin-propeller multi-rotor UAV in a complex wind field environment in a coastal area. The wind speed in the area varies dramatically, and the interaction between the sea breeze and the terrain produces complex turbulence, which poses a great challenge to the flight stability of the UAV. Traditional control methods are often difficult to maintain the stability of the UAV in such an environment, resulting in problems such as large flight trajectory deviation, high energy consumption, and delayed attitude adjustment. To this end, the present invention adopts an improved Fourier neural operator to predict aerodynamic disturbances, and combines it with an improved chaotic gray wolf optimization algorithm to optimize the model parameters, and finally integrates it into an adaptive control strategy based on the combination of model predictive control and sliding mode control to optimize the flight stability and energy efficiency of the UAV.

[0125] For the experimental scenario, we selected a coaxial twin-propeller hexacopter drone with a 1.2-meter wingspan, a maximum payload of 2.5 kg, and a maximum flight time of 45 minutes. It was equipped with a high-precision IMU, a barometric pressure sensor, a wind speed measurement device, and a real-time positioning system. The experiment involved two control schemes: a traditional control method (baseline group) and the control method proposed in this invention (experimental group). Flight tests were conducted under various wind speed conditions to evaluate flight trajectory error, attitude stability, energy consumption, and real-time disturbance prediction accuracy.

[0126] During the specific application process, the experimental group's drone first collected real-time flight data, including parameters such as wind speed, air pressure, and blade speed, and input them into an optimized Fourier neural operator model to calculate real-time aerodynamic disturbance data. Subsequently, the optimal thrust distribution strategy was calculated based on the model predictive control algorithm, and the attitude was rapidly adjusted through sliding mode control. Under strong wind disturbances, the present invention can predict wind field changes in advance and actively optimize the drone's thrust and attitude, allowing it to maintain high-precision trajectory control even in unstable aerodynamic environments.

[0127] In an environment with high wind speed and high turbulence, the drones in the baseline group showed significant flight trajectory deviation, with the maximum deviation reaching 1.8 meters, and there was large attitude jitter. Especially when the wind speed changed rapidly, there was a delay of 300-500 milliseconds in attitude adjustment. In contrast, the experimental group of drones used an improved Fourier neural operator to predict disturbances and combined it with an adaptive control strategy for adjustment. Its maximum trajectory deviation was only 0.4 meters, and the attitude adjustment delay was reduced to 80-120 milliseconds, significantly improving flight stability. In addition, within the same flight time (15 minutes), the average energy consumption of the drones in the experimental group decreased by about 12.5%, effectively extending the flight time. The experimental data is as follows:

[0128] Table 1: Flight stability test data of coaxial twin-propeller multi-rotor UAV in complex wind field

[0129] Test items Traditional method (benchmark group) The method of the present invention (experimental group) Maximum flight path deviation (m) 1.8 0.4 Posture adjustment delay (milliseconds) 300～500 80～120 Energy consumption reduction (%) - 12.5 Maximum wind resistance (m / s) 6.5 10.2 Flight time (minutes) 35 40

[0130] Experimental data demonstrates that the present invention significantly improves the flight stability, wind resistance, and energy efficiency of coaxial twin-propeller multi-rotor drones in complex aerodynamic environments. In terms of flight trajectory control, the experimental group's maximum trajectory deviation was reduced by 77.8% compared to traditional control methods, from 1.8 meters to 0.4 meters. This demonstrates that the present invention's improved Fourier neural operator can accurately predict wind field disturbances and, combined with an adaptive control strategy, effectively adjust thrust distribution and attitude compensation, enabling the drone to maintain a precise flight path even in complex wind environments.

[0131] In terms of attitude adjustment response speed, the attitude adjustment delay of traditional control methods is 300 to 500 milliseconds, while the present invention reduces it to 80 to 120 milliseconds, an improvement of approximately 73%. This optimization is mainly due to the improved Fourier neural operator that can predict wind speed changes in advance, and combined with the improved chaotic gray wolf optimization algorithm to optimize control parameters, the adaptive control system can adjust the rotor thrust more quickly and achieve precise control of the flight attitude. This improvement is crucial for UAVs flying in sudden wind speed changes or turbulent environments. It can reduce the jitter caused by wind field disturbances during flight and ensure stability.

[0132] In terms of energy consumption, this invention utilizes an optimized thrust distribution strategy, reducing the average energy consumption of drones by 12.5%. Compared to traditional control methods, this optimization effectively extends flight time, allowing drones to perform longer missions with the same amount of power. This is particularly true in high-wind conditions, where traditional methods often require significant additional power consumption to combat wind forces. However, this invention, through precise prediction of aerodynamic disturbances, can proactively adjust thrust, enabling drones to maintain stable flight while maintaining optimal energy consumption, thereby improving flight endurance.

[0133] Furthermore, the control method of the present invention achieves a maximum wind resistance of 10.2 m / s, while the conventional control method only has a maximum wind resistance of 6.5 m / s, an improvement of over 57%. This result demonstrates that the present invention not only improves the stability of the drone in moderate wind speeds, but also enables it to maintain flight in higher wind speeds. This improvement is of great significance for drone missions in strong wind environments such as maritime patrols and mountain monitoring, enabling drones to operate safely and stably in a wider range of application scenarios.

[0134] Overall, the experimental data fully demonstrates the advantages of this invention. Compared to traditional control methods, this invention improves the prediction accuracy of aerodynamic disturbances through an improved Fourier neural operator. Combined with an improved Chaos Gray Wolf optimization algorithm to optimize control parameters, this significantly enhances the flight stability of drones in complex wind environments, resulting in faster attitude adjustment response, lower energy consumption, longer endurance, and greater wind resistance. This provides more reliable technical support for future drone applications in high-wind environments.

[0135] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization is characterized by: The steps include: S1. Collect flight data, pre-process the data, and generate an aerodynamic disturbance training dataset; S2. Based on the aerodynamic disturbance training dataset, an improved Fourier neural operator network model is constructed; S3. Using the improved chaotic grey wolf optimization algorithm, the parameters of the improved Fourier neural operator network model are globally searched and locally optimized to update the parameters; S4. Embed the optimized improved Fourier neural operator network model into the UAV flight control system to receive flight data in real time and predict aerodynamic disturbance data; S5. Based on the real-time predicted aerodynamic disturbance data, the model predictive control algorithm is used to achieve dynamic adjustment of the thrust of each rotor and compensation of the flight attitude; S6. Verify the stability and real-time performance of the control method under various aerodynamic interference conditions through simulation tests and actual flight tests.

2. The coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization according to claim 1 is characterized in that: The S2 specifically includes: S21. Define the input parameter space and assume that the input data for the UAV aerodynamic disturbance prediction is u(x,t); S22. Aiming at the complex flow field disturbance characteristics of coaxial twin-propeller multi-rotor UAVs, a multi-resolution Fourier neural operator is used to improve the frequency domain mapping method of the traditional Fourier neural operator network. Specifically, S221, the first layer, the local Fourier transform layer, uses the local window Fourier transform to extract the local flow field disturbance characteristics under different scale windows. is the local Fourier transform operator, then the local spectrum under different scale windows is expressed as: Where l represents the number of local windows, is the Fourier frequency domain representation under the local window; S222, the second layer, the global Fourier transform layer, represents the local spectrum Perform weighted global fusion: Among them, α l is the learnable fusion weight; S223, the third layer, the adaptive spectrum filtering layer, sets W(k) as the frequency domain weight matrix, uses adaptive spectrum filtering to dynamically adjust the weights of high and low frequency information: Where b is the bias term, and W(k) is learned using the dynamic spectral attention mechanism; S224, the fourth layer, the inverse Fourier transform layer, restores the frequency domain information to the physical space through the inverse Fourier transform: in, is the inverse Fourier transform, d(x,t) is the predicted data of the UAV aerodynamic disturbance; S23. Construct cross-scale perturbations, including: S231, low-frequency disturbance compensation, uses a low-pass filter to extract low-frequency disturbance features and perform nonlinear transformation: d low (x,t))ReLU(W low d(x,t)+b low )4 Among them, ReLU is the activation function, W low is the low-frequency weight matrix, b low is the low frequency compensation term; S232, high-frequency disturbance enhancement, uses a high-pass filter to extract high-frequency disturbance information, combined with adaptive gain adjustment: d high (x,t)=tanh(W high d(x,t)+b high ); Among them, tanh(·) is the hyperbolic tangent function, W high is the high-frequency weight matrix, b high is the high frequency compensation term; S233, cross-scale disturbance fusion, combining low-frequency disturbance compensation and high-frequency disturbance enhancement for disturbance enhancement: d final (x,t)=β low d low (x,t)+β high d high (x,t); Among them, β low and β high is the learnable fusion weight; S24. Define the output of the improved Fourier neural operator model and set the final aerodynamic disturbance prediction result as d final (x, t), the complete mapping relationship can be expressed as: in, is the trained Fourier neural operator, θ is the optimized model parameter, u(x,t) is the input data, d final (x, t) is the final aerodynamic disturbance prediction output value.

3. The coaxial twin-propeller multi-rotor UAV stabilization control method based on aerodynamic interference optimization according to claim 1 is characterized in that: The S3 specifically includes: S31. Initialize the parameters of the improved gray wolf optimization algorithm, set the optimization group size N and the maximum number of iterations T, and set the four wolf individual position vectors X to represent the parameters to be optimized of the improved Fourier neural operator network model, where: X i ={W(k),b,W low ,b low ,W high ,b high ,β low ,β high }; Initialize the initial positions of the gray wolf group to be evenly distributed in the search space, randomly initialize the weight matrix, set four wolf individuals, the main wolf X α 、Deputy Wolf X β , Second Deputy Wolf X δ , Detection Wolf X θ ; S32. Calculate the fitness function and set the fitness function J to measure the error of the model parameters in predicting aerodynamic disturbances. The objective function is: Among them, MSE(·) is the mean square error loss function, SpectralError(·) is the spectral error loss function, d final is the aerodynamic disturbance prediction output value of the improved Fourier neural operator network model, d true is the real aerodynamic disturbance data in the aerodynamic disturbance training dataset, d final ,d true The Fourier spectrum of λ1, λ2, and λ3 are weight factors, and Diversity(X) is a population diversity maintenance item to prevent the algorithm from premature convergence. S33, using chaotic mapping to generate the initial position of the population S34, update the position of individual gray wolves, calculate the dominant wolf X of the current population α 、Deputy Wolf X β , Second Deputy Wolf X δ , Detection Wolf X θ , the first three wolves search according to the classic gray wolf optimization algorithm: Among them, the main wolf X α 、Deputy Wolf X β , Second Deputy Wolf X δ Will be updated with each iteration according to the new population fitness value, is the position of the gray wolf in generation t, c1, c2, c3 are the dynamic adjustment control factors; Detecting wolves for global exploration: in, is the position of the individual wolf detected in the tth generation, Levy(λ) is a random variable generated from the Levy distribution, γ is the global search weight, X rand is a random individual; S35, perform convergence judgment and parameter update, calculate the optimal fitness value J of each generation α , if J α When convergence or the maximum number of iterations T is reached, the current optimal solution X is output. α As the final optimization parameter, if it does not converge, it returns to continue iterative optimization. The final updated parameter is expressed as: i opt ={W opt (k),b opt ,W low,opt ,b low,opt ,W high,opt ,b high,opt ,b low,opt ,b high,opt }; S36, using the optimized parameter θ opt Re-adjust the improved Fourier neural operator network model to achieve the optimal effect.

Images