Three-degree-of-freedom air floating platform control method based on improved neural network PID (Proportion Integration Differentiation)

Abstract

The invention relates to an improved neural network PID (Proportion Integration Differentiation)-based three-degree-of-freedom air floating platform control method, which is suitable for a high-precision motion control scene in a microgravity environment. The control method is based on an improved neural network PID control algorithm, adopts a single-position loop structure to realize feedback from global position expectation to local thrust control, improves the network learning rate and robustness by introducing Momentum and RMSprop optimization strategies, and realizes online adaptive adjustment of control parameters. According to the invention, the three-degree-of-freedom air floating platform has high stability, response speed and control precision in a complex disturbance environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a control method for the motion of a three-degree-of-freedom air-floating platform. Background Art

[0002] Air-floating platforms are devices commonly used in ground-based microgravity simulation experiments, and are widely applied in verification tasks such as on-orbit servicing, attitude control, and docking operations for spacecraft. Their working principle involves suspending the platform on a smooth plane using air cushion technology, enabling two-dimensional translation and one-dimensional rotation, thereby simulating three-degree-of-freedom (3-DOF) motion in space. However, these platforms suffer from problems such as dynamic model uncertainty, strong coupling, and sensitivity to external disturbances, making it difficult for traditional proportional-integral-derivative (PID) control methods to achieve good dynamic response and stability, especially under disturbed and nonlinear systems.

[0003] Therefore, in recent years, neural networks (NNs) have been gradually introduced into the field of intelligent control as adaptive regulators for PID parameters. These NNs are combined with dynamic real-time optimization of controller weights based on system errors, thereby enhancing the robustness and generalization ability of the control system, especially suitable for complex nonlinear systems. Although various intelligent algorithms have been applied to similar platforms, most still rely on offline training or neglect real-time error response characteristics. Summary of the Invention

[0004] The purpose of this invention is to solve the motion control problem of a three-degree-of-freedom air-bearing platform. It proposes a control method based on an improved neural network PID controller using the Momentum and RMSprop methods. Compared with existing technologies, this invention possesses stronger anti-interference capabilities, faster response speed, and higher control accuracy, making it particularly suitable for complex, nonlinear, and real-time-critical microgravity experimental platforms. The three degrees of freedom include translational motion along the platform's x-axis and y-axis, and rotational motion around the z-axis.

[0005] The technical solution adopted by the present invention to achieve the above objectives is: a three-degree-of-freedom air-float platform control method, comprising the following steps:

[0006] 1) Establish a dynamic model of the air-floating platform that considers platform mass, inertia, ducted fan thrust and torque, and air damping; based on the dynamic model of the air-floating platform, construct a neural network PID controller with an inner and outer loop nested structure;

[0007] 2) The platform's two-dimensional position and attitude angle are collected in real time and fed back to the execution controller;

[0008] 3) The actuator uses a neural network PID controller to perform closed-loop control of the ducted fan: the outer loop neural network PID controller generates speed reference commands based on the two-dimensional position and attitude error, and the inner loop neural network PID controller outputs thrust and torque control signals based on the speed error to drive the ducted fan.

[0009] The dynamic model of the air-float platform is as follows:

[0010]

[0011] in, These are the linear accelerations of the platform center in the x and y directions, respectively; w z Let be the angular velocity of the platform center around the z-axis; These represent the thrust of the ducted fan in the x-axis direction, the thrust in the y-axis direction, and the torque about the z-axis, respectively; k x k y k z , where are the air damping coefficients in the x, y, and z directions, respectively, and m is the platform mass. These represent the velocities of the platform center in the x and y directions, respectively; r represents the distance from the center of the ducted fan to the platform center. This indicates the thrust generated by the rotation of the ducted fan; This represents the Coriolis torque experienced by the center of the platform in the direction of rotation about the z-axis. This represents the average disturbance torque at the platform center along the z-axis. I represents the angular acceleration of the platform's center of rotation about the z-axis, used to obtain the attitude angle. z This represents the moment of inertia of the platform center about the Z-axis.

[0012] There are four ducted fans, which are labeled as first fan, second fan, third fan, and fourth fan in a clockwise direction around the center of the platform, starting with a certain ducted fan. The first and fourth fans are used to control the positive movement of the x-axis, and the second and third fans are used to control the negative movement of the x-axis. The third and fourth fans are used to control the positive movement of the y-axis, and the first and second fans are used to control the negative movement of the y-axis. The second and fourth fans are used to control the counterclockwise rotation of the platform, and the first and third fans are used to control the clockwise rotation of the platform.

[0013] The neural network PID controller constructed based on the air flotation platform dynamic model includes the following steps:

[0014] A neural network PID controller is used to control the position and attitude of a three-degree-of-freedom air-float platform;

[0015] The desired platform pose, including the desired two-dimensional position and attitude angle around the z-axis, and the actual platform pose, including the x and y-axis position information measured by ultrasonic sensors and the rotation angle around the z-axis measured by a six-axis gyroscope, are respectively used as inputs to two neurons in the input layer and fed into the neural network PID controller. The neural network structure processes these two sets of data iteratively to generate the output control quantity; the control quantity includes the control force of the platform in the x-axis direction. Control force in the y-axis direction and the control torque around the z-axis The control quantities are substituted into the corresponding control equations in the dynamic model of the air-floating platform to drive the four ducted fans, thereby achieving control and motion of the platform in three degrees of freedom.

[0016] The neural network PID controller is a three-layer feedforward network, as detailed below:

[0017] The input layer is used to obtain pose error information based on the received platform desired pose and platform actual pose. The platform desired pose includes the desired two-dimensional position and the attitude angle around the z-axis. The platform actual pose includes the x and y axis position information measured by the ultrasonic sensor and the rotation angle around the z-axis measured by the six-axis gyroscope.

[0018] A hidden layer is used to calculate and output PID parameters based on the pose error information;

[0019] The output layer is used to generate corresponding three-degree-of-freedom control force and torque control quantities based on the PID parameters, and distribute them to each ducted fan for execution.

[0020] The input to the input layer is:

[0021] v1(n)=z(n), v2(n)=y(n), x i (n) = NET[v i [n], i = 1, 2;

[0022] Among them, v i z(n) represents the input value of the i-th neuron in the input layer, z(n) is the expected output of the n-th sample, used to represent the expected position of the platform in the x and y directions and the expected angle around the z-axis, y(n) is the actual output of the n-th sample, used to represent the actual position of the platform in the x and y directions and the actual angle around the z-axis, NET(n) is the activation function, x i (n) represents the activation value of the neuron;

[0023] The hidden layer yields: h = 1, 2, 3;

[0024] Where, v′ h(n) represents the weighted sum of inputs to the h-th neuron in the hidden layer, w ih The weights are from the input layer to the hidden layer; the three neurons in the hidden layer perform proportional, integral, and derivative functions respectively, and after discretization, we get:

[0025] Proportional function: x′1(n) = v′1(n)

[0026] Integral function: x′2(n)=x′2(n-1)+v′2(n)

[0027] Differential function: x′3(n)=v′3(n)-v′3(n-1)

[0028] Where x′1(n) is used for the output of the proportional term, x′2(n) is used for the output of the integral term, and x′3(n) is used for the output of the differential term, serving as P, I, and D parameters respectively;

[0029] The output layer is: o = 1; x″ o (n)=v″ o (n)

[0030] Among them, W ho Let v″1(n) be the weights from the hidden layer to the output layer, and v″1(n) be the weighted inputs to the neurons in the output layer. h (n) represents the activation values ​​of the three neurons in the hidden layer, o represents the number of neurons in the output layer, and v″ o (n) represents the activation value of the output layer neuron, x″ o (n) represents the final control output of the controller, used to substitute the control force in the x-axis direction in the dynamic model of the air-floating platform. Control force in the y-axis direction and the control torque around the z-axis

[0031] In the neural network PID controller, the weights W from the hidden layer to the output layer ho The gradient descent method was used to obtain:

[0032]

[0033] Where W(n) is the weight, used to substitute into W ho η is the learning rate, and β is the parameter for weighted summation of historical values. 0 Let β be the power of 0, ΔW(n) be the current gradient calculated by the neural network in the nth iteration, and α be the parameter for weighted summation of historical values. 0 α is the power of 0, α n-1 Let α be the (n-1)th power of α, and ε be a positive number.

[0034] A three-degree-of-freedom air-float platform control system includes:

[0035] The controller construction unit is used to establish a dynamic model of the air-floating platform that considers the platform mass, inertia, ducted fan thrust and torque, and air damping; based on the dynamic model of the air-floating platform, a neural network PID controller with an inner and outer loop nested structure is constructed.

[0036] The data acquisition unit is used to acquire the two-dimensional position and attitude angle of the real-time acquisition platform and feed it back to the execution controller;

[0037] The actuator controller is used to perform closed-loop control of the ducted fan using a neural network PID controller: the outer loop neural network PID controller generates speed reference commands based on two-dimensional position and attitude errors, and the inner loop neural network PID controller outputs thrust and torque control signals based on speed errors to drive the ducted fan.

[0038] The present invention has the following beneficial effects and advantages:

[0039] 1. Compared with traditional control processes, this invention introduces an error-driven online adaptive learning mechanism, which supports real-time updates of PID parameters during operation, avoiding the lag problem of offline training and improving the control strategy's ability to respond quickly to environmental disturbances and nonlinear dynamics.

[0040] 2. This invention employs an improved neural network PID algorithm based on Momentum and RMSprop, which features faster convergence speed and stronger robustness. The Momentum optimization strategy accelerates the error descent process and suppresses gradient oscillations; the RMSprop algorithm dynamically adjusts the learning rate to avoid getting trapped in local optima, thereby improving the network's stability and generalization ability. 3. The overall control system possesses excellent anti-interference capabilities and control precision, maintaining the attitude stability and precise trajectory tracking of the three-degree-of-freedom air-bearing platform in complex microgravity environments, making it particularly suitable for nonlinear, highly coupled, and highly disturbed experimental scenarios.

[0041] 3. This invention is based on an improved neural network PID control algorithm. It employs a nested inner and outer loop structure to achieve hierarchical feedback from global position expectation to local thrust control. Furthermore, it introduces Momentum and RMSprop optimization strategies to enhance the network's learning rate and robustness, enabling online adaptive adjustment of control parameters. This method allows a three-degree-of-freedom air-bearing platform to possess strong stability, response speed, and control accuracy under complex disturbance environments. Attached Figure Description

[0042] Figure 1a This is a front view of the three-degree-of-freedom air-float platform structure described in this invention.

[0043] Figure 1bThis is a side view of the three-degree-of-freedom air-float platform structure described in this invention.

[0044] Figure 1c This is a top view of the three-degree-of-freedom air-float platform structure described in this invention.

[0045] Figure 2 This is a top view of the three-degree-of-freedom air-float platform described in this invention, with the coordinate system labeled.

[0046] Figure 3 This is a schematic diagram of the neural network PID controller described in this invention.

[0047] Figure 4 This is a flowchart of the control system described in this invention.

[0048] Figure 5 This is a schematic diagram of the inner and outer loop nested control principle described in this invention. Detailed Implementation

[0049] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0050] A control method for improving neural network PID based on Momentum and RMSprop methods includes the following steps:

[0051] S1: Establish the dynamic model of the three-degree-of-freedom air-float platform.

[0052] A schematic diagram of the three-degree-of-freedom air flotation platform model is shown below. Figures 1a to 1c As shown. The three-degree-of-freedom air-float platform structure is existing technology. Assuming the three-degree-of-freedom air-float platform is a rigid body, determine OX. n Y n Z n Let OX be an inertial coordinate system. b Y b Z b Using a geometric coordinate system, four ducted fans are distributed at the four corners of the air-float platform, and the four ducted fans are paired to control two forward degrees of freedom and one rotational degree of freedom of the air-float platform, as follows: Figure 2 As shown, the three-degree-of-freedom air-float platform structure is existing technology. In the top view of the platform, four identical ducted fans are positioned at the four corners (top left, bottom left, top right, bottom right). The axes of two ducted fans located on a single diagonal line coincide, and the intersection of the axes of the four ducted fans coincides with the geometric center of the platform. Driven by the four ducted fans, the three-degree-of-freedom air-float platform achieves translation along the x and y axes and rotation around the z-axis. Combined with ultrasonic sensors and a six-axis gyroscope, the platform's position and attitude information are acquired in real time. The position is obtained from the ultrasonic sensors, and the attitude angles are obtained from the six-axis gyroscope.

[0053] A dynamic model of a three-degree-of-freedom air-floating platform is established using the Newton-Euler equations.

[0054]

[0055] Where m represents the platform quality, These are the velocities in the x and y directions. r represents the distance from the center of the ducted fan to the center of the platform. w x This represents the angular velocity about the z-axis. g represents the center of mass. Used to represent gravity, buoyancy, and other external forces acting on an air-floating platform. It is used to represent the thrust generated by the rotation of a ducted fan. It is used to represent all external torques acting on the air-float platform, excluding the torque generated by the ducted fan. This represents the torque acting on the air-float platform generated by the rotation of the ducted fan. z k represents the moment of inertia about the Z-axis. x k y k z These represent the average coefficients of air resistance torque on the x, y, and z axes, respectively.

[0056] S2: A position controller is designed for a three-degree-of-freedom air-floating platform. The control structure adopts a single-position loop to achieve closed-loop regulation from the desired global position to the control force output. The controller receives the platform's desired position and current actual position data, calculates the position error and attitude error, and the neural network PID controller directly outputs the control force and control torque based on this error information, including the control force in the x and y directions and the control torque around the z-axis. The output results are applied to the platform's dynamic model to drive the propulsion device for adjustment. This control structure achieves direct feedback control from attitude to force, simplifies the system structure, and ensures good control accuracy and robustness. Figure 5 As shown.

[0057] A neural network PID controller is used to control the position and attitude of a three-degree-of-freedom air-floating platform. In this control structure, the desired pose of the platform (including the desired two-dimensional position and attitude angle around the z-axis), the x and y-axis position information measured by ultrasonic sensors, and the rotation angle around the z-axis measured by a six-axis gyroscope are respectively used as inputs to two neurons in the input layer and fed into the neural network PID controller. The controller processes these two sets of data through the neural network structure, iteratively, and then generates output control commands. The control output includes the control force of the platform in the x-axis direction. Control force in the y-axis direction and the control torque around the z-axis These control variables are substituted into the corresponding control equations in the dynamic model of the air-floating platform to drive the four ducted fans, thereby achieving precise control and stable motion of the platform in three degrees of freedom.

[0058] S3: Design a neural network PID controller improved with Momentum and RMSprop.

[0059] This neural network PID controller is a three-layer feedforward network, consisting of an input layer, a hidden layer, and an output layer, containing two, three, and one neurons respectively. Figure 3 As shown. The input consists of two parts: the desired input and the real-time output of the system. After each iteration, the neural network calculates new P, I, and D values ​​as the three parameters of the neural network PID controller. After being processed by the dynamic model, these values ​​are used as new inputs to continue iterating. The system tends to stabilize after the number of samplings N approaches a large value.

[0060] In the forward algorithm, when iterating to the nth time, the input and output of the input layer neurons are as follows:

[0061] v1(n) = z(n)

[0062] v2(n)=y(n)

[0063] x i (n) = NET[v i [n], i = 1, 2;

[0064] z(n) is the expected output of the nth sample, y(n) is the actual output of the nth sample, and NET(n) is the activation function, set to the identity activation function, i.e.:

[0065] NET[v i [n]=v i (n)

[0066] The hidden layer can be obtained as follows:

[0067] h = 1, 2, 3;

[0068] v i (n) represents the input value of the i-th neuron in the input layer, w ih As the weights from the input layer to the hidden layer, considering that this neural network serves a PID system, the three neurons here perform their proportional, integral, and derivative functions respectively. After discretizing the continuous system, we get:

[0069] Scale function:

[0070] x′1(n)=v′1(n)

[0071] Points system:

[0072] x′2(n)=x′2(n-1)+v′2(n)

[0073] Differential function:

[0074] x′3(n)=v′3(n)-v′3(n-1)

[0075] x′1(n) is used for the output of the proportional term, x′2(n) is used for the output of the integral term, and x′3(n) represents the output of the hidden layer for the differential term. Therefore, the output layer can be derived as follows:

[0076] o=1

[0077] x″ o (n)=v″ o (n)

[0078] W ho The weights from the hidden layer to the output layer, v″1(n) is the weighted input of the output layer neuron, x″ o (n) represents the final output of the controller.

[0079] The output layer generates corresponding three-degree-of-freedom force and torque control quantities based on the PID control parameters and distributes them to each ducted fan for execution. The distribution is as follows: There are four ducted fans, which are sequentially labeled as the first fan, second fan, third fan, and fourth fan in a clockwise direction around the center of the platform, starting from a certain ducted fan. The first and fourth fans are used to control positive movement along the x-axis, and the second and third fans are used to control negative movement along the x-axis. The third and fourth fans are used to control positive movement along the y-axis, and the first and second fans are used to control negative movement along the y-axis. The second and fourth fans are used to control the platform to rotate counterclockwise, and the first and third fans are used to control the platform to rotate clockwise.

[0080] The online learning method in a neural network PID controller employs an error backpropagation algorithm, which achieves learning and memorization effects by modifying the neural network weights (W). After each sampling, the network dynamically updates the PID weights based on error backpropagation, thus achieving real-time adaptive adjustment. First, the system error (ideal value minus actual value) is calculated, and then the weights of each layer of the neural network are updated using a general gradient descent method. During the (n+1)th sampling, the error... Let the learning step size be η. After the nth learning iteration, the weight values ​​between the hidden layer and the output layer change as follows:

[0081]

[0082] Where x″1 represents the activation value of the output layer neuron.

[0083] The weight values ​​between the input layer and the hidden layer are changed as follows:

[0084]

[0085] W ih x represents the weights between the input layer and the hidden layer. h ′ represents x1′, x2′, x3′.

[0086] Each sampled data point, after one learning iteration, will generate the latest neural network weights, specifically the weights W between hidden layer neurons and output layer neurons. ho The output x′1(n), x′2(n), and x′3(n) are the real-time P, I, and D coefficients.

[0087] Subsequently, the Momentum and RMSprop methods were introduced. The Momentum method changed the weight update rule in the standard gradient descent method by introducing an intermediate quantity:

[0088] V(n)=βV(n-1)+(1-β)ΔW(n)

[0089] After replacing ΔW(n) in the original expression with V(n), we get:

[0090] W(n)=W(n-1)-ηV(n)=W(n-1)-η[βV(n-1)+(1-β)ΔW(n)]

[0091] W(n)=W(n-1)-η(1-β)[ΔW(n)+βΔW(n-1)+β 2 ΔW(n-2)+…+β n-1 ΔW(1)]

[0092] β represents the parameter for weighted summation of historical values; a value around 0.9 generally yields good algorithm performance. ΔW(n) represents the current gradient calculated by the neural network in the nth iteration, which is equivalent to performing a weighted summation on the historically processed data. Historical data from further back than the nth iteration has a relatively smaller impact on weight updates, while weights updated in the more recent iterations have a greater impact.

[0093] Adagrad's adaptive learning rate adjustment method transforms η (learning rate) in W(n) = W(n-1) - ηΔW(n) into... To prevent system errors when the denominator is zero, the parameter ε is set to a very small number.

[0094] S(n) = S(n-1) + ΔW(n)·ΔW(n)

[0095] By combining the formula for S(n) and the formula for W(n) in Adagrad's learning rate, we can derive:

[0096]

[0097] If the weights are modified too much, the learning rate will decrease significantly. During gradient descent training, there's an initial period of rapid change, followed by a plateau where the learning rate changes very slowly. When it enters a rapid change phase again after the plateau, the descent is still slow and unable to catch up, as all historical weight data is considered, leading to problems. Therefore, the RMSprop method introduces momentum into the learning rate, making the formula:

[0098] S(n)=αS(n-1)+(1-α)ΔW(n)·ΔW(n)

[0099] S(n) represents the weighted sum of the squared gradients in each iteration, and α and β have the same function, representing the parameters for weighted summation of historical values. From the above derivation, we can deduce that the gradient descent method in the neural network-based PID is ultimately optimized as:

[0100]

[0101] α and β have the same effect, so we take α = β = 0.9. ε is a very small positive number set to prevent division by zero.

[0102] In this embodiment, the three-degree-of-freedom air-bearing platform generates thrust through four ducted fans, enabling translational motion along its x and y axes and rotational motion around the z axis. The platform suspends on a smooth working surface, and its motion is controlled by adjusting the thrust, achieving high-precision three-degree-of-freedom control. To obtain the platform's motion state in real time, it is equipped with an ultrasonic position sensor to measure two-dimensional position coordinates; simultaneously, a six-axis gyroscope is installed at the platform's center to acquire the platform's attitude angle information in real time, ensuring accurate attitude feedback.

[0103] This invention establishes a three-degree-of-freedom air-floating platform dynamic model based on the Newton-Euler equations, taking into account factors such as platform mass, inertia, thrust and torque generated by the ducted fan, and air resistance, providing a theoretical basis for the design of a neural network PID controller.

[0104] The overall structure of the control system is as follows Figure 4 As shown, the system comprises two main parts: a power system and a sensing system. The power system consists of four ducted fans, whose rotation direction and speed are directly adjusted by the actuator controller to generate the required thrust and torque, driving the platform to complete three-degree-of-freedom motion. The sensing system consists of position sensors and attitude sensors, which collect the platform's position and attitude data in real time and feed the collected data back to the actuator controller through the I / O interface.

[0105] The actuator controller receives real-time feedback signals from the position and attitude sensors and calculates the position and attitude errors. For example... Figure 5 As shown, through a neural network PID control algorithm with nested inner and outer loops, the outer loop generates a speed reference command based on the position error, and the inner loop outputs thrust and torque control signals based on the speed error, driving the ducted fan to adjust the magnitude and direction of the thrust, thereby achieving precise control of the platform's motion.

[0106] The neural network PID controller employs a three-layer feedforward structure. The input layer receives the desired and actual outputs, the hidden layers handle proportional, integral, and derivative functions, and the output layer generates the final control input. An online error backpropagation algorithm dynamically adjusts the weights, enabling real-time adaptive optimization of the PID parameters. Momentum and RMSprop algorithms are introduced to further improve learning stability and convergence speed.

[0107] Furthermore, mathematical analysis of the control system is performed based on Lyapunov stability theory, proving that the error converges under appropriate learning rate conditions, thus ensuring the closed-loop stability of the control system.

[0108] By integrating the power system and the sensing system, and combining the aforementioned neural network PID control algorithm and real-time online learning mechanism, efficient and stable control of the three-degree-of-freedom air-float platform was achieved. Experimental verification shows that this system possesses excellent response speed and control accuracy under complex disturbance environments, meeting the stringent motion control requirements of microgravity simulation experiments.

[0109] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An improved neural network PID control method for a three-degree-of-freedom air-floating platform, characterized in that, The following steps are involved: 1) Establish a dynamic model of the air-floating platform that considers platform mass, inertia, ducted fan thrust and torque, and air damping; based on the dynamic model of the air-floating platform, construct a neural network PID controller with an inner and outer loop nested structure; 2) The platform's two-dimensional position and attitude angle are collected in real time and fed back to the execution controller; 3) The actuator uses a neural network PID controller to perform closed-loop control of the ducted fan: the outer loop neural network PID controller generates speed reference commands based on the two-dimensional position and attitude error, and the inner loop neural network PID controller outputs thrust and torque control signals based on the speed error to drive the ducted fan.

2. The improved neural network PID control method for a three-degree-of-freedom air-floating platform according to claim 1, characterized in that, The dynamic model of the air-float platform is as follows: in, These are the linear accelerations of the platform center in the x and y directions, respectively; w z Let be the angular velocity of the platform center around the z-axis; These represent the thrust of the ducted fan in the x-axis direction, the thrust in the y-axis direction, and the torque about the z-axis, respectively; k x k y k z , where are the air damping coefficients in the x, y, and z directions, respectively, and m is the platform mass. These represent the velocities of the platform center in the x and y directions, respectively; r represents the distance from the center of the ducted fan to the platform center. This indicates the thrust generated by the rotation of the ducted fan; This represents the Coriolis torque experienced by the center of the platform in the direction of rotation about the z-axis. This represents the average disturbance torque at the platform center along the z-axis. I represents the angular acceleration of the platform's center of rotation about the z-axis, used to obtain the attitude angle. z This represents the moment of inertia of the platform center about the Z-axis.

3. The improved neural network PID control method for a three-degree-of-freedom air-floating platform according to claim 1, characterized in that, There are four ducted fans, which are labeled as first fan, second fan, third fan, and fourth fan in a clockwise direction around the center of the platform, starting with a certain ducted fan. The first and fourth fans are used to control the positive movement of the x-axis, and the second and third fans are used to control the negative movement of the x-axis. The third and fourth fans are used to control the positive movement of the y-axis, and the first and second fans are used to control the negative movement of the y-axis. The second and fourth fans are used to control the counterclockwise rotation of the platform, and the first and third fans are used to control the clockwise rotation of the platform.

4. The improved neural network PID control method for a three-degree-of-freedom air-floating platform according to claim 1, characterized in that, The neural network PID controller constructed based on the air flotation platform dynamic model includes the following steps: A neural network PID controller is used to control the position and attitude of a three-degree-of-freedom air-float platform; The desired platform pose, including the desired two-dimensional position and attitude angle around the z-axis, and the actual platform pose, including the x and y-axis position information measured by ultrasonic sensors and the rotation angle around the z-axis measured by a six-axis gyroscope, are respectively used as inputs to two neurons in the input layer and fed into the neural network PID controller. The neural network structure processes these two sets of data iteratively to generate the output control quantity; the control quantity includes the control force of the platform in the x-axis direction. Control force in the y-axis direction and the control torque around the z-axis The control quantities are substituted into the corresponding control equations in the dynamic model of the air-floating platform to drive the four ducted fans, thereby achieving control and motion of the platform in three degrees of freedom.

5. A three-degree-of-freedom air-floating platform control method based on an improved neural network PID according to claim 1 or 4, characterized in that, The neural network PID controller is a three-layer feedforward network, as detailed below: The input layer is used to obtain pose error information based on the received platform desired pose and platform actual pose. The platform desired pose includes the desired two-dimensional position and the attitude angle around the z-axis. The platform actual pose includes the x and y axis position information measured by the ultrasonic sensor and the rotation angle around the z-axis measured by the six-axis gyroscope. A hidden layer is used to calculate and output PID parameters based on the pose error information; The output layer is used to generate corresponding three-degree-of-freedom control force and torque control quantities based on the PID parameters, and distribute them to each ducted fan for execution.

6. The improved neural network PID control method for a three-degree-of-freedom air-floating platform according to claim 5, characterized in that, The input to the input layer is: v1(n)=z(n),v2(n)=y(n),x i (n)=NET[v i (n)],i=1,2; Among them, v i z(n) represents the input value of the i-th neuron in the input layer, z(n) is the expected output of the n-th sample, used to represent the expected position of the platform in the x and y directions and the expected angle around the z-axis, y(n) is the actual output of the n-th sample, used to represent the actual position of the platform in the x and y directions and the actual angle around the z-axis, NET(n) is the activation function, x i (n) represents the activation value of the neuron; The hidden layer yields: Where, v′ h (n) represents the weighted sum of inputs to the h-th neuron in the hidden layer, w ih The weights are from the input layer to the hidden layer; the three neurons in the hidden layer perform proportional, integral, and derivative functions respectively, and after discretization, we get: Proportional function: x′1(n) = v′1(n) Integral function: x′2(n)=x′2(n-1)+v′2(n) Differential function: x′3(n)=v′3(n)-v′3(n-1) Where x′1(n) is used for the output of the proportional term, x′2(n) is used for the output of the integral term, and x′3(n) is used for the output of the differential term, serving as P, I, and D parameters respectively; The output layer is: x″ o (n)=v″ o (n) Among them, W ho Let v″1(n) be the weights from the hidden layer to the output layer, and v″1(n) be the weighted inputs to the neurons in the output layer. h (n) represents the activation values ​​of the three neurons in the hidden layer, o represents the number of neurons in the output layer, and v″ o (n) represents the activation value of the output layer neuron, x″ o (n) represents the final control output of the controller, used to substitute the control force in the x-axis direction in the dynamic model of the air-floating platform. Control force in the y-axis direction and the control torque around the z-axis 7. The improved neural network PID control method for a three-degree-of-freedom air-floating platform according to claim 6, characterized in that, In the neural network PID controller, the weights W from the hidden layer to the output layer ho The gradient descent method was used to obtain: Where W(n) is the weight, used to substitute into W ho η is the learning rate, and β is the parameter for weighted summation of historical values. 0 Let β be the power of 0, ΔW(n) be the current gradient calculated by the neural network in the nth iteration, and α be the parameter for weighted summation of historical values. 0 α is the power of 0, α n-1 Let α be the (n-1)th power of α, and ε be a positive number.

8. An improved neural network PID three-degree-of-freedom air-float platform control system, characterized in that, include: The controller building unit is used to establish a dynamic model of the air-floating platform that takes into account the platform mass, inertia, ducted fan thrust and torque, and air damping. A neural network PID controller with an inner and outer loop nested structure was constructed based on the dynamic model of the air flotation platform. The data acquisition unit is used to acquire the two-dimensional position and attitude angle of the real-time acquisition platform and feed it back to the execution controller; The actuator controller is used to perform closed-loop control of the ducted fan using a neural network PID controller: the outer loop neural network PID controller generates speed reference commands based on two-dimensional position and attitude errors, and the inner loop neural network PID controller outputs thrust and torque control signals based on speed errors to drive the ducted fan.

Images