Adaptive Control Method for Trajectory Tracking of Rotorcraft Incorporating AI Deep Learning

Abstract

The present invention discloses a trajectory tracking adaptive control method for a rotary-wing aircraft integrating AI deep learning. By using an algorithm that combines AI deep learning and adaptive control, an aerodynamic term of wind interference that is not accurately modeled in a wind interference environment is obtained, and trajectory tracking adaptive control is performed on the rotary-wing aircraft in different wind speed interference environments; combining the basis function and the adaptive parameter, an aerodynamic term of wind interference that is not accurately modeled in a wind interference environment is obtained; combining feedback control, in a flight reference trajectory tracking controller of a reference trajectory that the aircraft needs to track and is pre-designed, the aerodynamic terms of wind interference that are not accurately modeled and are adaptively identified online are compensated in real time. The present invention introduces a normalized linear coefficient adaptive estimation strategy, which improves the anti-interference robustness of the system while proving the stable convergence of the error between the unmodeled dynamic term and the adaptive term of the rotary-wing aircraft system under the influence of wind interference.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft navigation, positioning and control, and particularly relates to a trajectory tracking adaptive control method for a rotary-wing aircraft integrating AI deep learning. Background Technique

[0002] Rotary-wing aircraft have the advantages of simple structural principle, low cost, vertical takeoff and landing, fixed-point hovering and convenient maintenance, and can complete many tasks that manned aircraft cannot. Especially in recent years, combined with emerging technologies such as artificial intelligence, they have been widely used in military and civilian fields.

[0003] With the development of technology and the continuous change of the military pattern, higher requirements are put forward for the performance of rotary-wing aircraft. Especially, the quality of the precise flight trajectory tracking control performance in high-altitude or offshore strong wind interference environments is one of the key indicators determining its future practical expansion ability, and this problem has not been effectively solved at present. Therefore, researching and solving the problem of precise flight trajectory tracking control of rotary-wing aircraft in wind interference environments has great application prospects.

[0004] In an external wind interference environment, it is difficult to establish an accurate dynamic model of a rotary-wing aircraft. The external wind environment brings serious non-linear and strongly coupled interference unmodeled dynamic effects to the aircraft model. Whether the wind interference can be effectively compensated during the design process of the controller is the key to its ability to accurately track the flight trajectory. At the same time, the wind speeds in different wind fields are different, resulting in different wind interference forces. To improve the adaptability and robustness of the control system of rotary-wing aircraft in unknown wind field environments, it is urgent to solve the difficult problem of designing a flight trajectory tracking control method suitable for environments with different wind speeds. Summary of the Invention

[0005] The purpose of the present invention is to solve the above-mentioned difficult problems, and to provide a trajectory tracking adaptive control method for a rotary-wing aircraft integrating AI deep learning for the task application of theoretical flight trajectory tracking control of rotary-wing aircraft in different wind speed interference environments.

[0006] To achieve the above purpose, the present invention provides the following method:

[0007] A trajectory tracking adaptive control method for a rotary-wing aircraft integrating AI deep learning uses an algorithm combining AI deep learning and adaptive control to obtain the unaccurately modeled wind interference aerodynamic terms in a wind interference environment, and performs trajectory tracking adaptive control on the rotary-wing aircraft in different wind speed interference environments, including the following steps:

[0008] Offline learning of the deep neural network is carried out through the collected flight data, and the basis function of the unaccurately modeled wind interference aerodynamic terms in the wind interference environment is trained and approximately fitted;

[0009] Update the adaptive parameters by using the information of the reference model prediction error term in adaptive control;

[0010] Combine the basis function and the adaptive parameters to obtain the aerodynamic terms of the wind disturbance that are not accurately modeled in the wind disturbance environment;

[0011] Combine feedback control and compensate the aerodynamic terms of the wind disturbance that are not accurately modeled and identified online in real time in the flight reference trajectory tracking controller of the reference trajectory that the aircraft needs to track and is pre-designed.

[0012] Furthermore, use an algorithm that combines AI deep learning and adaptive control to obtain the aerodynamic terms of the wind disturbance that are not accurately modeled in the wind disturbance environment

[0013] including

[0014] The dynamic equation of the aerodynamic terms of the wind disturbance that are not accurately modeled is:

[0015]

[0016] Among them, the meanings of each state are as follows:

[0017] Position state vector in the earth horizontal coordinate system: P fs =[X, Y, Z] T , the superscript T represents the matrix transpose, X represents the north coordinate, Y represents the east coordinate, Z represents the earth coordinate, and the origin is set as the take-off point;

[0018] Attitude Euler angle state: Θ fs =[φ, θ, ψ], roll angle φ, yaw angle ψ, pitch angle θ, roll angular velocity Yaw angular velocity Pitch angular velocity

[0019] Moment of inertia: Moment of inertia about the roll axis Moment of inertia about the yaw axis Moment of inertia about the pitch axis

[0020] Mass of the aircraft: m;

[0021] Gravitational acceleration: g;

[0022] Aircraft control command term: U = [U1 U2 U3 U4] T , U1 is the control quantity of the rotor thrust, U2 is the control quantity of the rotor roll moment, U3 is the control quantity of the rotor pitch moment, and U4 is the control quantity of the rotor yaw moment;

[0023] The unmodeled wind disturbance aerodynamic term f(x, μ) caused by different external wind conditions: f(x, μ) = [f X f Y f Z T are the components of the wind disturbance aerodynamic term in the three directions of the northward X-axis, eastward Y-axis, and downward Z-axis of the earth horizontal coordinate system. This unmodeled aerodynamic term is related to the state vector x of the aircraft and the wind field vector μ. The state vector is selected as the three-dimensional velocity of the aircraft (three-dimensional), the pitch angle, roll angle, and yaw angle of the aircraft attitude represented by Euler angles (three-dimensional), and the four-dimensional combination of the rotational speed control commands of the four motors of the rotor represented by pulse width modulation signals (PWM), where the four dimensions are the rotational speed control signals of the four motors of the rotorcraft. Assume that n different types of wind speed wind fields are respectively represented as μ1, μ2,..., μ n , and a specific k-th type of wind speed is set as μ k ;

[0024] By fusing AI deep learning and adaptive control algorithms, the inaccurately modeled wind disturbance aerodynamic term f(x, μ) is fitted to have the following expression:

[0025]

[0026] where, is the representation basis function shared by all different wind conditions learned through the deep neural network, that is, used to learn and capture the common dependence term of the unmodeled dynamics on the UAV state. This quantity is independent of the magnitude of the wind speed. λ(μ) is a set of linear coefficients, briefly denoted as λ, which have different values under different wind conditions and are online identified through the adaptive control algorithm, is multiplied by λ(μ) to approximately fit the inaccurately modeled wind disturbance aerodynamic term f(x, μ);

[0027] When the rotor UAV is under online control, the approximately fitted inaccurately modeled wind disturbance aerodynamic term f(x, μ) is added to the control force for real-time compensation to eliminate the inaccurately modeled dynamic influence and enable the aircraft to accurately track the reference trajectory.

[0028] Furthermore, the deep neural network is offline learned through the pre-collected flight data to train and approximately fit the representation basis function shared by all different wind conditions in the inaccurately modeled wind disturbance aerodynamic term under the wind disturbance environment

[0029] including

[0030] ​Introduce the adversarial transfer learning deep network in the deep neural network learning algorithm. Under the condition of few-shot data, perform effective learning to represent the aerodynamics independent of the wind field with a deep neural network. Through the characteristic of the deep neural network to fit any function, the output of the deep neural network is regarded as a set of basis functions representing the wind interference aerodynamic effect, that is, the shared aerodynamic model basis functions under different wind conditions. And introduce the adversarial training step to eliminate the influence brought by the changes of different wind conditions with a classification network. Specifically,

[0031] The adversarial transfer learning deep network adopts a fully connected form and trains two networks in total. And ξ, the two are two different DNNs; the network The output of Is the aerodynamic representation function shared by different wind conditions, that is, Is independent of the wind condition; the adaptive parameter λ(μ) is a linear coefficient related to the wind condition. Each different wind condition has a specific parameter λ(μ), and the k-th wind condition has a specific parameter λ(μ k ).); Train a deep neural network So that the product of it and λ(μ) can approximately represent the uncertain wind interference aerodynamic term f(x, μ) under different wind conditions; Since the training set is the data of the flight state quantities x such as the speed, attitude, and motor speed of the aircraft under different wind conditions, in actual flight, the state quantity x of the aircraft itself will change due to the change of the wind condition and wind speed. Directly training with a simple least square error loss function The neural network is likely to learn the aerodynamic model related to the wind condition. Therefore, in order to eliminate the influence brought by the change of the wind condition and wind speed and truly let the network Learn the shared aerodynamic model independent of the wind condition; introduce the classification network ξ. The role of the network ξ is to input the Under different wind conditions and be able to output the corresponding wind condition and wind speed category k, and introduce adversarial training. Use the classification network ξ as the discriminator and train And the two DNN networks of ξ at the same time. Adversarial training means that in each round of iteration, first fix the network parameters of the classification network ξ and update the network Then fix the network parameters of the network And update the network ξ. When the two reach the Nash equilibrium, the training is completed.

[0032] Furthermore, the adversarial training algorithm adopted by the adversarial transfer learning deep network specifically includes

[0033] The deep neural network, i.e., the adversarial transfer learning deep network, is trained using the least squares and stochastic gradient descent methods. The input state variables of the neural network are the three-dimensional velocity of the aircraft (three-dimensional), the pitch angle, roll angle, and yaw angle of the aircraft attitude represented by Euler angles (three-dimensional), and the four motor speed control commands of the rotor represented by pulse width modulation signals (PWM) (four-dimensional); the output state of the neural network is the representation basis function shared by different wind conditions (five-dimensional row vector); the linear coefficient λ(μ) = [λ X (μ)λ Y (μ)λ Z (μ)] T (fifteen-dimensional, the adaptive parameter vectors λ X (μ), λ Y (μ), and λ Z (μ) are five-dimensional column vectors respectively), the five-dimensional output of the neural network and the first five dimensions λ X (μ), the middle five dimensions λ Y (μ), and the last five dimensions λ Z (μ) of the fifteen-dimensional linear coefficient updated by the adaptive control are multiplied respectively, and exactly three numbers are obtained. These three numbers are used to approximate the three components of the wind field on the three axes in the geodetic coordinate system, namely the northward X-axis, the eastward Y-axis, and the downward Z-axis, representing the three-axis unmodeled aerodynamic terms of the aircraft in the geodetic coordinate system f(x, μ) = [f X f Y f Z T (three-dimensional) are as follows:

[0034]

[0035] The loss function of the adversarial transfer learning deep network is as follows:

[0036]

[0037] where k represents the single data collected in n different wind condition experiments, N K represents the total number of sample points collected in the k-th wind condition experiment, i represents a sample point among the total number of sample points collected in the k-th wind condition experiment, represents the aircraft state variables collected in the k-th wind condition and wind speed experiment, represents the residual force error term between the predicted aircraft acceleration of the aircraft nominal reference model and the actually measured aircraft acceleration in the k-th wind condition and wind speed experiment, that is, the unmodeled wind interference aerodynamic term; is a typical least square error loss function,

[0038] ​ is the loss function of the K - classification network, usually the cross - entropy loss function; the loss function introduces classification errors, and during training it enables the classification model to accurately distinguish the current wind condition category k, ensuring the training can learn a shared aerodynamic model independent of wind speed and wind conditions.

[0039] Furthermore, updating the adaptive parameters using the information of the reference model prediction error term in adaptive control

[0040] includes

[0041] In the online adaptive control stage, a set of linear coefficients λ(μ) are updated. The linear coefficients λ(μ) represent a set of constant vectors that vary with different wind disturbances and different wind field vectors μ, and are identified online through an adaptive control algorithm;

[0042] The linear coefficients λ(μ) are obtained by an adaptive estimation method based on the normalized least - squares method. The adaptive estimation method includes a reference trajectory feed - forward term, a basic PID control feedback term, an adaptive parameter update rate, and a time - varying matrix covariance update rate based on the normalized least - squares method,

[0043] Specifically as follows:

[0044]

[0045]

[0046] where, u F is the control rate of the rotorcraft, λ is the adaptive linear parameter updated online, R is the time - varying covariance matrix, e is the reference trajectory tracking error term, PID(e) is the PID controller designed according to the reference trajectory tracking error term e, the superscript T is the transpose symbol, r is the position reference trajectory command in the earth coordinate system that the aircraft needs to track preset in advance, k r is the reference trajectory feed - forward control coefficient designed according to the tracking reference trajectory, k r ranges from 0 to 1 and is adjusted according to the actual reference trajectory tracking situation. y is the residual force error term obtained by subtracting the acceleration of the aircraft predicted by the nominal reference model of the aircraft from the acceleration of the aircraft measured by actual simulation or experiment.

[0047] Furthermore, the collected flight data is collected by two methods: theoretical simulation method and experimental method. The theoretical simulation method is to collect the simulated flight state data set of the aircraft under different wind fields when tracking the preset trajectory under the conventional PID controller on the computer; the experimental method includes wind tunnel tests, multi-gear industrial fan blowing tests or human-in-the-loop test methods to collect the actual flight state data set of the rotorcraft when tracking the preset trajectory under the conventional PID controller.

[0048] Furthermore, it also includes proving the stable convergence of the error between the unmodeled dynamic term and the adaptive term of the system under the influence of wind disturbance through the Lyapunov stability theorem, specifically including

[0049] According to the fact that neural networks have been proven to have the ability to approximate any function with arbitrary precision, it is assumed that through training, the deep neural network has accurately learned the shared aerodynamic feature representation basis function At the same time, let the ideal value of the adaptive parameter λ be λ * , then the residual force error term y is expressed as Let be the estimation error of the adaptive parameter, let the estimation error of the unmodeled dynamic term and the predicted residual force error term under the influence of wind disturbance be ε, and then the adaptive parameter update rate is expressed as Select the Lyapunov function as follows:

[0050]

[0051] where, R -1 is the inverse of matrix R, is the matrix of the trace;

[0052] According to the fact that the product of matrix R and its inverse matrix R -1 is the identity matrix I, that is, RR -1 = I, then taking the derivative of both sides at the same time, we can know:

[0053]

[0054] The derivative of the system trajectory along the Lyapunov function is:

[0055]

[0056] where, ||ε|| represents the norm of the estimation error ε and its value is greater than or equal to zero; according to the Lyapunov stability theorem, has a finite limit and the system is stable; at the same time, according to the Barbalat lemma commonly used in the stability proof of nonlinear adaptive control systems, when the second derivative of the Lyapunov function When it is a bounded value, the first derivative of the Lyapunov function is uniformly continuous. When time approaches infinity, i.e., t → ∞, the derivative of the Lyapunov function approaches zero, i.e., that is, the estimation error ε converges to zero, meeting the requirement that the error ε between the unmodeled dynamic term and the adaptive term of the rotorcraft system under the influence of wind disturbance converges stably.

[0057] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored program, wherein the program, when running, executes the above method.

[0058] The present invention also provides an electronic device, including a memory and a processor, characterized in that a computer program is stored in the memory, and the processor is configured to execute the above method through the computer program.

[0059] Compared with the prior art, the above technical solution conceived by the present invention can achieve the following beneficial effects:

[0060] First, the deep neural network is offline learned through pre-collected flight data, and the basis function of the unmodeled aerodynamic term generated in the wind disturbance environment is approximately fitted and obtained; then, in the adaptive control, the adaptive parameters are updated by using information such as the reference model prediction error term; finally, combined with the idea of feedback control, in the designed reference flight trajectory tracking controller, the wind disturbance aerodynamic term that cannot be accurately modeled in advance, that is, the unmodeled wind disturbance aerodynamic term, is compensated in real time.

[0061] The method designed by the present invention introduces the idea of adversarial transfer learning in the deep learning algorithm. Under the condition of few-sample data, an effective learning of the deep neural network representation of aerodynamics independent of the wind field is carried out. Through the characteristic that the deep neural network can fit any function, the shared aerodynamic model basis function under different wind conditions is obtained, and an adversarial training step is introduced to eliminate the influence brought by the changes of different wind conditions by the classification network, improving the adaptability of the system to different wind fields. At the same time, the technical solution of the present invention changes the traditional development process of the aircraft, breaks through the dependence on the accurate aerodynamic model, reduces the expensive wind tunnel test cost, and shortens the development cycle.

[0062] In the adaptive control method designed by the present invention, a linear coefficient adaptive estimation strategy based on normalization is introduced, which improves the anti-interference robustness of the system while proving the stable convergence of the error between the unmodeled dynamic term and the adaptive term of the rotorcraft system under the influence of wind disturbance. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a block diagram of the flight adaptive control algorithm integrating the AI deep learning method of the present invention;

[0064] Figure 2 This is the block diagram of the offline learning stage of the deep neural network of the present invention;

[0065] Figure 3 This is the block diagram of the stage of combining the online adaptive control and the learning and training results of the deep neural network of the present invention. Detailed implementation manners

[0066] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0067] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners. The following embodiments are merely illustrative and do not constitute a limitation to the present invention.

[0068] In order to solve the problem of accurate flight trajectory tracking control of a rotorcraft in an environment with different wind speed disturbances, the unmodeled wind disturbance aerodynamic terms are identified in the control system and compensated for this disturbance in real-time control. Based on the above objectives, the present invention designs an algorithm combining adversarial transfer learning deep neural network and adaptive control by establishing a theoretical dynamic model of the rotorcraft, such as Figure 1 shown, designs an algorithm combining adversarial transfer learning deep neural network and adaptive control. By simulating on a computer or actually flying a test flight to collect a flight state data set of the aircraft tracking a preset trajectory under a conventional PID controller under different wind speed disturbance conditions, the proposed adversarial transfer learning deep neural network is offline trained using the data set, and finally an adaptive control algorithm is used for online wind disturbance identification and compensation. The specific implementation is divided into the following four steps:

[0069] (1) Design an algorithm framework combining AI deep learning and adaptive control

[0070] Since the rotorcraft will generate different unmodeled wind disturbance aerodynamic terms in the face of external environmental disturbances, such as different wind speed profiles, this part has an important impact on the accurate trajectory tracking of the aircraft under the influence of the wind, and with the different flight dynamic environments, this unmodeled aerodynamic term cannot be accurately modeled and measured in advance. The dynamic equation of the rotorcraft considering the unmodeled wind disturbance aerodynamic term is as follows:

[0071]

[0072] Among them, the meanings of each state are as follows:

[0073] Position state vector in the geodetic horizontal coordinate system: P fs = [X, Y, Z] T , where the superscript T represents matrix transpose, X represents the north coordinate, Y represents the east coordinate, Z represents the geodetic coordinate, and the origin is set as the take-off point;

[0074] Attitude Euler angle state: Θ fs = [φ, θ, ψ], roll angle φ, yaw angle ψ, pitch angle θ, roll angular velocity Yaw angular velocity Pitch angular velocity

[0075] Moment of inertia: Moment of inertia about the roll axis Moment of inertia about the yaw axis Moment of inertia about the pitch axis

[0076] Mass of the aircraft: m;

[0077] Gravitational acceleration: g;

[0078] Aircraft control command item: U = [U1 U2 U3 U4] T , where U1 is the control quantity of the rotor thrust, U2 is the control quantity of the rotor roll moment, U3 is the control quantity of the rotor pitch moment, and U4 is the control quantity of the rotor yaw moment;

[0079] Unmodeled wind disturbance aerodynamic term f(x, μ) caused by different external wind conditions: f(x, μ) = [f X f Y f Z T are the components of the wind disturbance aerodynamic term in the three directions of the north X-axis, east Y-axis, and geodetic Z-axis of the geodetic horizontal coordinate system. This unmodeled aerodynamic term is related to the state vector x of the aircraft and the wind field vector μ. The state vector is selected as the combination of the three-dimensional velocity of the aircraft (three-dimensional), the pitch angle, roll angle, and yaw angle of the aircraft attitude represented by Euler angles (three-dimensional), and the four motor speed control commands of the rotor represented by pulse width modulation signals (PWM) (four-dimensional). The four-dimensional is the speed control signal of the four motors of the rotorcraft, the speed signal PWM1 of the No. 1 motor, the speed signal PWM2 of the No. 2 motor, the speed signal PWM3 of the No. 3 motor, and the speed signal PWM4 of the No. 4 motor; Assume that n different types of wind fields of wind speeds are represented as μ1, μ2,..., μ n , where a specific k-th type of wind speed is set as μ k ;

[0080] ​Neural networks have been proven to have the ability to approximate any function with arbitrary precision. In particular, in recent years, deep learning artificial intelligence (AI) algorithms based on deep neural networks (DNNs) have achieved remarkable success in various industrial applications. With the development of artificial intelligence technology, intelligent design technology based on machine learning has been proven to be an effective technical approach for the design of high-performance control systems. Therefore, the present invention designs an algorithm that combines AI deep learning and adaptive control, and its algorithm block diagram is as shown in Figure 1 shown;

[0081] The algorithm that combines AI deep learning and adaptive control includes an offline learning stage and an online adaptive control stage for real-time online learning. For the offline learning stage, an adversarial transfer learning algorithm is adopted to perform an effective learning of the deep neural network (DNN) representation of aerodynamics that is independent of the wind field under the condition of few-sample data. The output of the deep neural network is regarded as a set of basis functions representing the aerodynamic effects of wind disturbances. In the online adaptive control stage, a set of linear coefficients λ(μ) is updated. The linear coefficients λ(μ) represent a set of constant vectors that are different for different wind disturbance vectors μ under different wind disturbances and can be identified online through an adaptive control algorithm. The purpose of the algorithm is to approximate the unmodeled wind disturbance aerodynamic term f(x, μ) by combining the AI deep learning and the adaptive control algorithm, and it has the following expression:

[0082]

[0083] where, is the representation basis function shared by all different wind conditions learned through the deep neural network, used to learn and capture the common dependence of the unmodeled dynamics on the UAV state. This quantity is independent of the magnitude of the wind speed. λ(μ) is a set of linear coefficients, which are online identified or recognized through the adaptive algorithm formula (2), simply denoted as λ, and have different values under different wind conditions, and are online identified through the adaptive control algorithm. Multiplying by λ(μ) can approximately fit the unmodeled wind disturbance aerodynamic term f(x, μ). When the rotor UAV is under online control, the estimated f(x, μ) is added to the control force to compensate for the influence of the unmodeled dynamics in real time, enabling the aircraft to accurately track the reference trajectory;

[0084] (2) Design an adversarial transfer learning algorithm based on a deep neural network

[0085] During actual flight, an unmanned aerial vehicle is constantly in different dynamic environments due to different wind conditions. The model learned by a deep neural network needs to have the ability to handle different tasks. The most popular learning scheme in dynamic environments currently is meta-learning, that is, "learning to learn", whose purpose is to learn an effective model from data collected from different tasks or environments. The learned model, usually represented as a DNN, should ideally be able to quickly adapt to a new task or an unseen environment when the given training data set is limited;

[0086] The adversarial transfer learning deep network adopted in the present invention adopts a fully connected form, and two networks are trained in total and ξ, and the two are two different DNNs. The output of network is an aerodynamic representation function shared by different wind conditions, that is is independent of wind conditions; the adaptive parameter λ(μ) is a linear coefficient related to wind conditions, and each different wind condition has a specific parameter λ(μ), such as the kth type of wind condition has a specific parameter λ(μ ). The purpose is to train a deep neural network k so that the product of it and λ(μ) can approximately represent the uncertain aerodynamic term f(x,μ) under different wind conditions. Since the training set is data of flight state quantities x such as the speed, attitude, and motor speed of the aircraft under different wind conditions, during actual flight, the state quantity x of the aircraft itself will change due to changes in wind speed. If a simple least square error loss function is directly used to train the obtained neural network will probably learn an aerodynamic model related to wind conditions. Therefore, in order to eliminate the influence brought by changes in wind speed, and truly enable the network to learn a shared aerodynamic model independent of wind conditions, a classification network ξ is introduced. The role of the network ξ is to be able to output the corresponding wind speed category k when inputting different wind conditions, and the idea of adversarial training is introduced. The classification network ξ is used as a discriminator, and the two DNN networks and ξ are trained simultaneously. Adversarial training means that in each round of iteration, first, the classification network ξ fixes the network parameters and updates the network and then the network fixes the network parameters and updates the network ξ, and when the two reach the Nash equilibrium, the training is completed; After that, the network fixes the network parameters and updates the network ξ, and when the two reach the Nash equilibrium, the training is completed;

[0087] The least squares and stochastic gradient descent methods are used to train the deep learning neural network. The input state quantity of the neural network is the three-dimensional speed of the aircraft (three-dimensional), the pitch angle, roll angle, and yaw angle of the aircraft attitude represented by Euler angles (three-dimensional), and the control commands of the four rotor motors represented by pulse width modulation signals (PWM) (four-dimensional); the output state of the neural network is a representation basis function shared by different wind conditions (five-dimensional row vector). Combined with the adaptive control algorithm, the linear coefficient λ(μ) is updated online = [λ X (μ)λ Y (μ)λ Z (μ)] T (fifteen dimensions, adaptive parameter vector λ for each axis X (μ), λ Y (μ) and λ Z (μ) are five-dimensional column vectors), the five-dimensional output of the neural network and the first five dimensions λ of the fifteen-dimensional linear coefficients updated by adaptive control X (μ), the fifth dimension λ Y (μ) and the last five dimensions λ Z (μ), multiply them respectively, and we get three numbers. These three numbers are used to approximate the three components of the wind field in the geodetic coordinate system, namely the north X axis, the east Y axis, and the ground Z axis, which represent the three-axis unmodeled aerodynamic term of the aircraft in the geodetic coordinate system f(x,μ)=[f X f Y f Z ] T (three-dimensional) is:

[0088]

[0089] The loss function of the adversarial transfer learning deep network is as follows:

[0090]

[0091] Where k represents the single data collected under n different wind conditions, N K represents the total number of sample points collected under the k-th wind condition experiment, i represents a sample point in the total number of sample points collected under the k-th wind condition experiment, represents the aircraft state variables collected under the k-th wind condition and wind speed experiment, The residual force error term represents the difference between the aircraft acceleration predicted by the nominal reference model of the aircraft collected under the k-th wind condition and wind speed experiment and the aircraft acceleration measured by the actual simulation or flight test, that is, the unmodeled aerodynamic term. is a typical minimum square error loss function, Is the loss function of the K classification network, usually the cross entropy loss function. The above loss function introduces classification error in training At the same time, the classification model can accurately distinguish the current wind condition category k, ensuring the training Able to learn a shared aerodynamic model that is independent of wind conditions and speed;

[0092] The deep network training process of adversarial transfer learning is shown in Table 1 below

[0093] Table 1 Training Process

[0094]

[0095]

[0096] (3) Offline / Online Machine Training with Data Augmentation

[0097] The dataset collection strategy is divided into two types: theoretical simulation method and experimental method. The theoretical simulation method is to collect the simulation flight state dataset of the aircraft under different wind fields when tracking the preset trajectory under the conventional PID controller on the computer; the experimental method collects the actual flight state dataset of the rotorcraft when tracking the preset trajectory under the conventional PID controller by designing static and dynamic test experimental environments indoors and outdoors and through methods such as wind tunnel tests, multi-gear industrial fan blowing tests or human-in-the-loop tests;

[0098] In the offline learning stage, the goal is to learn a model from the data collected in different environments. Such a model contains the shared knowledge or features in all environments, and learns the aerodynamic feature basis functions shared in all wind conditions The block diagram of the offline learning stage is specifically as Figure 2 shown;

[0099] For the online adaptive control stage, the goal is to adapt the model learned offline, given limited online data from the training environment and training tasks, new environments or new tasks, such as fine-tuning the aerodynamic features under specific wind conditions. The block diagram of the combination stage of online adaptive control and deep neural network learning training is specifically as Figure 3 shown:

[0100] (4) Design an Adaptive Control Algorithm Based on Normalized Least Squares

[0101] Through deep neural network training, the shared aerodynamic feature representation basis function can be found to minimize the error of each batch of data prediction. For the online control stage, to minimize the tracking error of the rotorcraft for the desired trajectory, the key lies in real-time identifying accurate wind disturbances and compensating for them; at the same time, due to the shared aerodynamic feature representation basis function obtained by deep neural network training may have a relatively large amplitude, and the actual system will inevitably introduce measurement noise into the adaptive update process, which will cause the adaptive parameters obtained by the traditional design method to update too fast, ultimately leading to the divergence of the adaptive term and the deterioration of the robustness of the system.

[0102] For the above purposes, the present invention proposes a linear coefficient adaptive estimation method based on normalized least squares, which uses the normalization technique to make the adaptive term The amplitude is reduced to improve the robustness of the system, and it is combined with traditional PID control technology. This method can automatically adjust the gain online and can quickly estimate the parameter λ(μ) with great uncertainty related to the wind condition and wind speed. The online adaptive control algorithm proposed in the present invention includes a reference trajectory feedforward term, a basic PID control feedback term, an adaptive parameter update rate, and a time-varying matrix covariance update rate based on the normalized least squares method. The Lyapunov stability theorem, which is commonly used to prove the stability of the control system, is used to prove the stable convergence of the error between the unmodeled dynamic term and the adaptive term of the rotorcraft system under the influence of wind interference. The control block diagram is as shown in Figure 1 shown below:

[0103]

[0104] where u F is the rotorcraft control rate, λ is the adaptive linear parameter updated online, R is the time-varying covariance matrix, e is the reference trajectory tracking error term, PID(e) is the PID controller designed according to the reference trajectory tracking error term e, the superscript T is the transpose symbol, r is the position reference trajectory command in the earth coordinate system that the aircraft needs to track and is preset, k r is the reference trajectory feedforward control coefficient designed according to the tracking reference trajectory, the value range of k r is between 0 and 1. As an example, it takes 0 or 0.5 or 1. In this embodiment, it takes 0.5 and is adjusted according to the actual reference trajectory tracking situation. y is the residual force error term obtained by subtracting the acceleration of the aircraft predicted by the aircraft nominal reference model from the acceleration of the aircraft measured by actual simulation or experiment; the reference trajectory is the trajectory that needs to be tracked and is the reference trajectory of the position information that the aircraft is planned to fly and is pre-designed; the aircraft nominal reference model is the model described by formula (1).

[0105] The stable convergence of the error between the unmodeled dynamic term and the adaptive term of the system under the influence of wind interference is proved by the Lyapunov stability theorem. Since the neural network has been proved to have the ability to fit any function with arbitrary accuracy, it is assumed that through training, the deep neural network has accurately learned the shared aerodynamic feature representation basis function At the same time, let the ideal value of the adaptive parameter λ be λ * , then the residual force error term y can be expressed as Let be the estimation error of the adaptive parameter. Let the estimation error of the unmodeled dynamic term and the predicted residual force error term under the influence of wind interference be ε, and then the adaptive parameter update rate can be expressed as Select the Lyapunov function as follows:

[0106]

[0107] Among them, R -1 is the inverse of matrix R, is the matrix trace.

[0108] According to the product of matrix R and its inverse matrix R -1 being the identity matrix I, that is, RR -1 = I, then taking the derivative on both sides, we can know that:

[0109]

[0110] The derivative of the system trajectory along the Lyapunov function is:

[0111]

[0112] Among them, ||ε|| represents the norm of the estimation error ε and its value is greater than or equal to zero. According to the Lyapunov stability theorem, has a finite limit and the system is stable. At the same time, according to the Barbalat lemma commonly used in the stability proof of nonlinear adaptive control systems, when the second derivative of the Lyapunov function is a bounded value, the first derivative of the Lyapunov function is uniformly continuous. When time approaches infinity, that is, t → ∞, the derivative of the Lyapunov function approaches zero, that is that is, the estimation error ε converges to zero, meeting the requirement that the error ε between the unmodeled dynamic term and the adaptive term of the rotorcraft system under the influence of wind disturbance is stably convergent.

[0113] The present invention also provides a computer-readable storage medium. The computer-readable storage medium includes a stored program. Among them, the program, when running, executes the foregoing method.

[0114] The present invention also provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to execute the foregoing method when the computer program runs.

[0115] In summary, the present invention provides a rotorcraft trajectory tracking adaptive control method that integrates AI deep learning.

Claims

1. An adaptive control method for trajectory tracking of a rotorcraft integrating AI deep learning, characterized in that Using an algorithm that combines AI deep learning and adaptive control, an aerodynamic term of wind interference that is not accurately modeled in a wind interference environment is obtained, and trajectory tracking adaptive control is performed on a rotorcraft in different wind speed interference environments, including the following steps: Offline learning of the deep neural network is performed using the collected flight data, and a basis function for the aerodynamic term of wind interference that is not accurately modeled in a wind interference environment is obtained through approximate fitting during training; The adaptive parameters are updated using the reference model prediction error term information in adaptive control; Combining the basis function and the adaptive parameters, an aerodynamic term of wind interference that is not accurately modeled in a wind interference environment is obtained; Combined with feedback control, in the designed flight trajectory tracking controller, the aerodynamic terms of wind interference that are not accurately modeled and identified online adaptively are compensated in real time; The updating of the adaptive parameters using the reference model prediction error term information in adaptive control includes In the online adaptive control stage, a set of linear coefficients λ(μ) is updated. The linear coefficients λ(μ) represent a set of constant vectors that are different for different wind interferences and different wind field vectors μ, and are identified online through an adaptive control algorithm; The linear coefficients λ(μ) are obtained using an adaptive estimation method based on the normalized least squares method. The adaptive estimation method includes a reference trajectory feedforward term, a basic PID control feedback term, an adaptive parameter update rate, and a time-varying matrix covariance update rate based on the normalized least squares method, Specifically as follows: where, u F is the control rate of the rotary-wing aircraft, λ is the adaptive linear parameter updated online, R is the time-varying covariance matrix, e is the reference trajectory tracking error term, PID(e) is the PID controller designed according to the reference trajectory tracking error term e, the superscript T is the transpose symbol, r is the position reference trajectory command in the earth coordinate system that the aircraft needs to track preset in advance, k r is the reference trajectory feedforward control coefficient designed according to the tracking reference curve, k r ranges from 0 to 1 and is adjusted according to the actual reference trajectory tracking situation. y is the residual force error term obtained by subtracting the body acceleration predicted by the aircraft nominal reference model from the body acceleration measured by actual simulation or experiment.

2. The adaptive control method for trajectory tracking of a rotary-wing aircraft according to claim 1, characterized in that Using an algorithm that combines AI deep learning and adaptive control to obtain an aerodynamic term of wind interference that is not accurately modeled in a wind interference environment Including The dynamic equation of the aerodynamic term of wind interference that is not accurately modeled is: Among them, the meanings of each state are as follows: Position state vector in the geodetic horizontal coordinate system: P fs = [X, Y, Z] T , where the superscript T represents matrix transpose, X represents the north coordinate, Y represents the east coordinate, Z represents the vertical coordinate, and the origin is set as the take-off point; Attitude Euler angle state: Θ fs = [φ, θ, ψ], roll angle φ, yaw angle ψ, pitch angle θ, roll angular velocity Yaw angular velocity Pitch angular velocity Moment of inertia: Moment of inertia about the roll axis Moment of inertia about the yaw axis Moment of inertia about the pitch axis Aircraft mass: m; Acceleration due to gravity: g; Flight vehicle control instruction item: U = [U1 U2 U3 U4] T , where U1 is the control quantity of the rotor thrust, U2 is the control quantity of the rotor roll moment, U3 is the control quantity of the rotor pitch moment, and U4 is the control quantity of the rotor yaw moment; The unmodeled wind disturbance aerodynamic term f(x, μ) caused by different external wind conditions: f(x, μ) = [f X f Y f Z T are the components of the wind disturbance aerodynamic term in the three directions of the northward X-axis, eastward Y-axis, and downward Z-axis of the earth horizontal coordinate system. This unmodeled aerodynamic term is related to the state vector x of the aircraft and the wind field vector μ. The state vector is selected as the three-dimensional velocity of the aircraft, the pitch angle, roll angle, and yaw angle of the aircraft attitude represented by Euler angles, and the four-dimensional combination of the rotor four-motor speed control commands represented by pulse width modulation signals. The four dimensions are the speed control signals of the four motors of the rotary-wing aircraft. Assume that n different types of wind fields of wind speeds are represented as μ1, μ2,..., μ n , and a specific k-th type of wind speed is set as μ k ;​ By fusing the AI deep learning and adaptive control algorithms to fit the aerodynamic term f(x, μ) of wind interference that is not accurately modeled, it has the following expression: Among them, is a representation basis function shared by all different wind conditions learned by a deep neural network, that is, used to learn and capture the common dependencies of the unmodeled dynamics on the UAV state. This quantity is independent of the magnitude of the wind speed. λ(μ) is a set of linear coefficients, simply denoted as λ. λ(μ) has different values under different wind conditions and is online identified through an adaptive control algorithm. is multiplied by λ(μ) to approximately fit the aerodynamic terms of the wind disturbance f(x, μ) that are not accurately modeled; When the rotor UAV is under online control, the approximately fitted aerodynamic term f(x, μ) of wind interference that is not accurately modeled is added to the control force for real-time compensation to eliminate the dynamic influence of the inaccurate modeling and enable the aircraft to accurately track the reference trajectory.

3. The adaptive control method for trajectory tracking of a rotary-wing aircraft according to claim 2, characterized in that The offline learning of the deep neural network is performed using the collected flight data, and an approximate fit is trained to obtain a representation basis function shared by all different wind conditions in the inaccurately modeled wind interference aerodynamic terms under wind interference environments. Including Introduce the adversarial transfer learning deep network in the deep neural network learning algorithm. Under the condition of few-sample data, perform effective learning of the deep neural network representation of aerodynamics independent of the wind field. Through the characteristic of the deep neural network to fit any function, the output of the deep neural network is regarded as a set of basis functions representing the aerodynamic effects of wind interference, that is, the shared aerodynamic model basis functions under different wind conditions. And introduce an adversarial training step to eliminate the influence brought by the changes in different wind conditions with a classification network. Specifically, The adversarial transfer learning deep network adopts a fully connected form, and two networks are trained in total. and ξ, which are two different DNNs; the output of network is an aerodynamic representation function shared by different wind conditions, that is, is independent of the wind conditions; ​ The adaptive parameter λ(μ) is a linear coefficient related to wind conditions. Different wind conditions have a specific parameter λ(μ). The k-th wind condition has a specific parameter λ(μ k ); Train a deep neural network such that the product of it and λ(μ) can approximately represent the uncertain wind interference aerodynamic term f(x, μ) under different wind conditions; Introduce a classification network ξ. The role of network ξ is to input the and be able to output the corresponding wind condition wind speed category k. And introduce adversarial training, use the classification network ξ as the discriminator, and train and ξ two DNN networks at the same time. Adversarial training means that in each iteration, first fix the network parameters of the classification network ξ and update the network and then fix the network parameters of network and update network ξ. When the two reach the Nash equilibrium, the training is completed.

4. The rotorcraft trajectory tracking adaptive control method according to claim 3, characterized in that The specific adversarial training algorithm adopted by the adversarial transfer learning deep network specifically includes Train a deep neural network, namely an adversarial transfer learning deep network, using the least squares and stochastic gradient descent methods. The input state variables of the neural network are the three-dimensional velocity of the aircraft, the pitch angle, roll angle, and yaw angle of the aircraft attitude represented by Euler angles, and the control commands for the rotational speeds of the four motors of the rotor represented by pulse width modulation signals. The output state of the neural network is the representation basis function shared by different wind conditions The linear coefficient λ(μ)=[λ X (μ)λ Y (μ)λ Z (μ)] T , which is updated online by combining with the adaptive control algorithm. The 5-dimensional output of the neural network and the first 5 dimensions λ X (μ) of the 15-dimensional linear coefficient updated by the adaptive control, the middle 5 dimensions λ Y (μ) and the last 5 dimensions λ Z (μ) are multiplied respectively to exactly obtain three numbers. These three numbers are used to approximate the three components of the wind field on the three axes in the earth coordinate system, namely the northward X-axis, the eastward Y-axis, and the downward Z-axis, representing the three-axis unmodeled aerodynamic terms of the aircraft in the earth coordinate system f(x,μ)=[f X f Y f Z T as follows:​ The loss function of the adversarial transfer learning deep network is as follows: where k represents the single data collected under n different wind condition experiments, and N K represents the total number of sample points collected under the k-th wind condition experiment, and i represents a sample point among the total number of sample points collected under the k-th wind condition experiment, represents the aircraft state variables collected under the k-th wind speed experiment, represents the residual force error term between the aircraft body acceleration predicted by the nominal reference model of the aircraft and the aircraft body acceleration measured by actual simulation or flight test under the k-th wind speed experiment, that is, the unmodeled wind interference aerodynamic term; is a typical least square error loss function, is the loss function of the K-classification network, usually the cross-entropy loss function; the loss function introduces classification errors and enables the classification model to accurately distinguish the current wind condition category k during training while ensuring that the training can learn a shared aerodynamic model independent of wind speed.

5. The adaptive control method for trajectory tracking of a rotary-wing aircraft according to claim 1, characterized in that [[ID= 6. The adaptive control method for trajectory tracking of a rotary-wing aircraft according to any one of claims 1-5, characterized in that ​ According to the fact that neural networks have been proven to have the ability to approximate any function with arbitrary precision, it is assumed that through training, the deep neural network has accurately learned the shared aerodynamic feature representation basis function At the same time, let the ideal value of the adaptive parameter λ be λ * , then the residual force error term y is expressed as Let be the estimation error of the adaptive parameter, and let the estimation error of the unmodeled dynamic term under the influence of wind disturbance and the predicted residual force error term be ε, and then the adaptive parameter update rate is expressed as Select the Lyapunov function as follows: wherein, R -1 is the inverse of matrix R, is the trace of the matrix; According to the product of matrix R and its inverse matrix R -1 being the identity matrix I, that is, RR -1 = I, it can be known by taking the derivative of both sides that: ​ where, ||ε|| represents the norm of the estimation error ε and its value is greater than or equal to zero; according to the Lyapunov stability theorem, has a finite limit and the system is stable; meanwhile, according to the Barbalat's lemma commonly used in the stability proof of nonlinear adaptive control systems, when the second derivative of the Lyapunov function is a bounded value, the first derivative of the Lyapunov function is uniformly continuous, and when time approaches infinity, i.e., t → ∞, the derivative of the Lyapunov function approaches zero, i.e., that is, the estimation error ε converges to zero, meeting the requirement that the error ε between the unmodeled dynamic term and the adaptive term of the rotorcraft system under the influence of wind disturbance converges stably.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when running, executes the method described in any one of claims 1 to 6 above.

8. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is configured to execute the method described in any one of claims 1 to 6 through the computer program.

Images