Control method of automotive axial-phase split-magnetic levitation flywheel rotor system based on inverse extended neural network structure

Abstract

The present invention is a control method for an axially split-phase magnetic levitation flywheel rotor system for an automobile based on a neural network inverse extension structure. The method comprises the following steps: Step 1: constructing a flywheel rotor dynamics model and obtaining a composite flywheel rotor system through linear amplification; Step 2: constructing a dynamic neural network inverse system; Step 3: linearizing and decoupling the combined pseudo-linear composite flywheel rotor system, and then performing closed-loop control of the system; Step 4: introducing a disturbance estimate into the model of the dynamic neural network inverse system to form an extension structure; Step 5: replacing the dynamic neural network inverse system with the extension structure, ultimately forming a complete magnetic levitation flywheel rotor auto-disturbance rejection control system based on the neural network inverse extension structure. The present invention enhances the robustness of the flywheel rotor under vehicle-borne disturbances, simplifies the model, improves the anti-interference performance of the control system, and makes rotor system control simpler, more robust, and more anti-disturbance.

Description

Technical Field

[0001] The present invention belongs to the technical field of control of magnetic levitation motors, and in particular relates to a control method for an axially split-phase magnetic levitation flywheel rotor system for a vehicle based on an inverse extension structure of a neural network. Background Art

[0002] In recent years, electric vehicles have become a new type of transportation. They are powered by an on-board power supply and use motors to drive the wheels. There is a complex electromechanical-magnetic coupling between the motor and the support system inside the on-board flywheel battery and the various degrees of freedom of the rotor. Due to different vehicle driving states such as starting, acceleration, deceleration, steering and emergency stops, as well as changes in uphill, downhill, curved and uneven road conditions, they all become complex vehicle-road-axle unmodeled dynamic factors.

[0003] Since the bending critical speed of the high-speed flywheel rotor system is much higher than its rated operating speed, the flywheel rotor can be approximately regarded as a rigid rotor system. When the flywheel rotor rotates, assuming that the on-board magnetic levitation rotor is mainly subjected to vertical up and down vibrations during operation, the air gap between the rotor and the electromagnetic coil will change, causing the rotor to become unstable or even lose its suspended support state, making the flywheel battery unable to work normally.

[0004] At present, a large number of studies have been carried out at home and abroad on the stable suspension control of magnetic levitation flywheel rotors, and a variety of methods have been proposed. They have been applied relatively maturely to a certain extent, but there are still deficiencies in accuracy, robustness and real-time performance. For example, the most widely used decentralized PID control ignores the coupling between the degrees of freedom and has low control accuracy under high-speed rotor conditions; the cross-feedback control algorithm mostly uses the Taylor linearization method to linearize the system model at the equilibrium point to complete feedback control. The disadvantage is that it is not robust enough to changes in the motor air gap; the adaptive feedback control method has good control accuracy and robustness, but the algorithm has a large amount of calculation and low real-time performance: the inverse system linearization decoupling algorithm has a clear physical concept and is easy to implement, but it is easily affected by changes in models and parameters. Therefore, in practical applications, it is necessary to further design robust controllers in the control system, such as sliding mode control, H∞ control, μ synthesis, LQR control, neural networks, fuzzy control, etc., which makes the algorithm very complex.

[0005] At the same time, the above control scheme does not fully consider the impact of vehicle operating conditions. For on-board flywheel batteries, changes in vehicle driving status and road conditions will affect the dynamic characteristics of the flywheel battery. This is an "unmodeled dynamic" outside the mechanism model, so reliable control performance under vehicle operating conditions cannot be guaranteed.

[0006] Furthermore, building an accurate model of the vehicle's support system is difficult. Complex motor operation modeling errors, nonlinear uncertainties, and environmental influences make precise suspended rotor control difficult to achieve. For flywheel batteries in vehicle-mounted situations, anti-disturbance capabilities are crucial. Traditional flywheel rotors use mechanical bearings as support structures. When the vehicle decelerates and accelerates, inertial forces act directly on the mechanical bearings. At this point, the high-speed flywheel rotor collides with the mechanical bearings, significantly reducing the life of the support system. Magnetic bearings, due to the lack of mechanical contact, can achieve frictionless support in a vacuum environment, making them ideal bearings for flywheels. However, electromagnetic bearings require controlling the winding current to achieve rotor suspension support. Complex changes in vehicle driving conditions and road conditions can alter the forces acting on the rotor, making the self-disturbance rejection capability of the suspended support system even more important.

[0007] Therefore, it is necessary to study the impact mechanism of changes in electric vehicle driving conditions such as starting acceleration, cruising constant speed, braking deceleration and driving conditions such as turning, uphill, downhill, and uneven road surface on the flywheel battery support system, and to study the gyroscopic effect suppression and robust control methods of the flywheel rotor under complex on-board disturbances, so as to achieve stable operation of the flywheel battery magnetic levitation rotor under on-board working conditions. Summary of the Invention

[0008] In order to solve the above problems, the present invention provides a control method for an automotive axial-split magnetic levitation flywheel rotor system based on a neural network inverse extension structure. From the perspective of flywheel rotor dynamics, the neural network inverse system can be used to convert the control of the complex system of the axial-split magnetic levitation flywheel motor into the control of five second-order linear subsystems, and then use an anti-disturbance control controller to perform closed-loop control, derive the disturbance value in the state observer, and construct a neural network inverse system extension structure to further realize stable suspension control of the magnetic levitation flywheel rotor under vehicle-mounted conditions, which can solve the existing robustness and anti-interference performance.

[0009] In order to achieve the above object, the present invention is achieved through the following technical solutions:

[0010] The present invention is a control method for an axially split-phase magnetic levitation flywheel rotor system for a vehicle based on an inverse extended neural network structure, comprising the following steps:

[0011] Step 1: Considering the interference of both the vehicle's driving state and the road conditions, the on-board flywheel battery rotor dynamics analysis results are formed, and a flywheel rotor dynamics model is constructed. After linear amplification by a bipolar power amplifier, a composite flywheel rotor system is obtained;

[0012] Step 2: Use static neural network and integrator S -1 Construct dynamic neural network inverse system;

[0013] Step 3: The dynamic neural network inverse system constructed in step 2 is placed before the composite flywheel rotor system to form a pseudo-linear system. The composite flywheel rotor system obtained in step 1 is linearized and decoupled. The decoupled pseudo-linear system is then closed-loop controlled using an active disturbance rejection controller.

[0014] Step 4: Introduce the disturbance estimation value generated by the extended state observer in the ADRC into the model of the neural network inverse system to form an extended structure of the neural network inverse system to increase the adaptability and anti-interference ability of the composite flywheel rotor system to disturbance changes;

[0015] Step 5: Replace the dynamic neural network inverse system in the previous step with the neural network inverse system expansion structure. By increasing the number of input nodes of the neural network and utilizing the interference estimation signal of the controlled flywheel rotor system, a complete magnetic levitation flywheel rotor self-disturbance rejection control system based on the neural network inverse expansion structure is finally formed.

[0016] The specific process of step 1 is:

[0017] Step 1-1: Since the bending critical speed of the high-speed flywheel rotor system is much higher than its rated operating speed, the flywheel rotor can be approximated as a rigid rotor system. A three-dimensional coordinate system is established with the rigid rotor mass center G as the origin. The external damping and gravity of the system are ignored. According to the Lagrange equation method in the dynamics theory of multi-rigid body systems, the dynamic equation of the axially split-phase magnetic levitation flywheel rotor system can be obtained:

[0018]

[0019] Where: m is the mass of the flywheel rotor, x, y, z are the translational displacements of the rotor in the x, y, and z axes under the coordinates of the center of mass, α and β are the rotation angles of the rotor around the x and y axes without considering the bending deformation of the rotor, -α is a positive value, are their second-order derivatives, f x and p x is the electromagnetic force and torque in the x direction at the center of mass, f y and p y is the electromagnetic force and torque in the y direction at the center of mass, f z is the electromagnetic force in the z direction at the center of mass, Δf and Δp are the external interference force and interference torque, f ax and f bx are the electromagnetic forces in the x-axis direction in the motor A phase and motor B phase coordinate systems, respectively, and f ay and f by are the electromagnetic forces in the y-axis direction of the motor phase A and phase B coordinate systems;

[0020] Step 1-2: Express the dynamic equation of the axial split-phase magnetic levitation flywheel rotor system in matrix form:

[0021]

[0022] Right now:

[0023]

[0024] Where: The rotor mass matrix is Coordinate vector of the flywheel rotor's center of mass Gyro torque coefficient matrix Rotor arm coefficient matrix Magnetic bearing electromagnetic force

[0025] Step 1-3: Connect the bipolar power amplifier that generates the actual control signal from the driving control signal in series before the magnetic levitation flywheel rotor system to form a composite flywheel rotor system.

[0026] The specific process of constructing the dynamic neural network inverse system in step 2 is:

[0027] Step 2-1: Use a static neural network and 10 integrators S -1 Construct a neural network inverse system, use a static neural network to approximate the nonlinear mapping, and use the integrator to reflect the dynamic characteristics of the inverse system. As the expected output of the neural network, the output of the flywheel rotor system, y = [y1, y2, y3, y4, y5] T =[x a ,y a ,z,x b ,y b ] T As the input of the neural network, when selecting samples, first select the appropriate excitation current signal, that is, multiple groups of different random square waves to excite the controlled object. The controlled object outputs a displacement signal and uses a high-order digital filter to filter out the high-frequency noise in the sampled data, thereby obtaining a high-precision original data sample {u1, u2, u3, u4, u5, y1, y2, y3, y4, y5}, and samples the measurable internal state x. A state observer is designed to observe the internal state of the system that is difficult to measure directly. Then, a high-order numerical differentiation method, that is, the five-point derivation method, is used to more accurately calculate the first-order and second-order derivatives of y. The five-point derivation formula is as follows:

[0028]

[0029] Step 2-2: Normalize the input and output {u, y} signals and normalize all data with a large span to -1~1. The normalized data can eventually constitute the input sample set and expected output sample set for training the neural network. In the simulation system, the sampled data is directly stored in the "To Workspace" module, the simulation time of the system is set to 30 seconds, the sampling interval is 0.01 seconds, and 3000 sets of data are obtained. Based on the data samples, a neural network is used as the identification model of the inverse system, with 15 input nodes, 4 output nodes, and the number of hidden layer nodes determined by the experiment. Further, 2500 sets of data are selected from the training samples for training, and another 500 sets of data are used for detection and verification. Offline learning is performed on the composite controlled object to determine the weight coefficients of the static neural network layer, and finally the accuracy requirements are met. After the static neural network is completed, 10 integrators S -1 Before being connected in series to the static neural network, since the input is the derivative of displacement and angle, x is obtained after the integration link. a ,y a ,z,x b ,y b The first-order and second-order derivatives and original values ​​are input into the static neural network, and finally form a dynamic neural network structure.

[0030] The specific process of step 3 is as follows:

[0031] Step 3-1: The dynamic neural network inverse system constructed in step 2 is placed before the composite flywheel rotor system to form a pseudo-linear system, which is equivalent to five second-order linear integral subsystems, that is, it is linearized and decoupled into five independent integral linear subsystems.

[0032] Step 3-2: Transform the rotor displacement signal detected by the displacement sensor to the rotor mass center. Assume that the rotor radial translation displacement detected by the motor A phase displacement sensor is x a and y a The motor B phase displacement sensor detects the rotor radial translation displacement as x b and y b , then the translational displacement at the rotor mass center O and the rotation angle of the rotor around the x-axis and y-axis are:

[0033]

[0034] The translational and rotational signals (x, y, z, α, β) at the center of mass of the rotor are obtained. The displacement in the Z direction does not pass through the sensor, so it does not need to be converted.

[0035] Step 3-3: When using a second-order linear active disturbance rejection controller, consider the model of the controlled plant as:

[0036]

[0037] Where f(t) is the generalized disturbance, including the disturbance d caused by the vehicle's own driving state and the road conditions that the flywheel rotor system encounters when it is on board. In order to estimate the generalized disturbance, it is regarded as a new state variable, and z1=y, z3=f, y only represents the output, and then the given value (x * ,y * ,z * ,α * ,β * ) is input to the ADRC, and the tracking differentiator arranges the transition process for the input signal. * ,y * ,z * ,α * ,β * ) and the value of the processed input signal itself is recorded as Its derivative is denoted as Waiting to be transferred to the next step;

[0038] Step 3-4: Input the actual displacement and rotation signals (x, y, z, α, β) generated by the pseudo-linear composite system feedback after adding the generalized disturbance in step 3-3 and the input u into the second-order extended state observer. Observe and estimate the states of each order of the control model, the sum of the internal and external disturbances acting on the model, and the unmodeled dynamics of the system to obtain the actual signal estimate z1 and the disturbance estimate z3. Then, perform differentiation on the actual signal estimate z1 to obtain z2.

[0039] Step 3-5: Get the The initial control quantity u0 is calculated by the linear feedback controller after being subtracted from z1 and z2 obtained in steps 3-4. The disturbance estimate z3 obtained by the second-order extended state observer is then used to compensate for the initial control quantity u0 to obtain the final input quantity u.

[0040] The step 4 specifically includes the following steps:

[0041] Step 4-1: The actual disturbance estimate z3 generated by the ADRC is used as a new state variable of the neural network inverse system in the vehicle-mounted state. When the vehicle starts and accelerates, the acceleration is set to remain constant, consistent with the acceleration transmitted to the magnetic levitation flywheel battery. At this time, the flywheel rotor is initially stationary, and the center of mass of the flywheel rotor shaft will lag behind the axis in the forward direction. When the set acceleration value suddenly increases, the relative offset of the rotor will change, causing a change in the electromagnetic force. At this time, the ADRC estimates the displacement and angular operating state of the flywheel rotor under different operating conditions.

[0042] Step 4-2: Establish a closed-loop simulation model of a pseudo-linear system based on an ADRC. Since the neural network inverse system and the flywheel rotor system form a second-order linear system, the second-order pseudo-linear system can be used to replace the decoupled flywheel rotor system. This helps to ignore the interference in the neural network inverse system and the influence of the integral link module of the Simulink software itself in the rotor system, and then form a closed-loop simulation model of the pseudo-linear system based on the ADRC.

[0043] Step 4-3: Use different types of disturbance signals to simulate different vehicle operating conditions, and introduce various disturbance signals into the pseudo-linear system closed-loop simulation model formed in step 4-2. Adjust the parameters of the active disturbance rejection controller. After the adjustment, the closed-loop simulation can achieve good control results.

[0044] Step 4-4: After simulation, the value of the disturbance estimate z3 is extracted and stored in the "To Workspace" module. After normalizing the signal, the input sample set and the expected output sample set for training the neural network can be formed. Since 3000 sets of data are obtained when building the neural network inverse system, z3 here also adjusts the sampling time, sets the time to 30 seconds, and the sampling time interval to 0.01 seconds, obtaining 3000 sets of data for data processing;

[0045] Step 4-5: The disturbance estimation value is included in the construction of the neural network inverse system. When using the neural network fitting function, the extended structure of the neural network inverse system is fitted according to the corresponding relationship between the input value and the output value. The extended structure after introducing the disturbance estimation value considers the unmodeled dynamics on the original basis to solve the disturbance problem encountered by the magnetic levitation flywheel battery rotor during operation, making subsequent control more precise.

[0046] The step 5 is specifically as follows: using the extended structure of the neural network inverse system obtained in step 4 to replace the neural network inverse system, and replacing the reference value of the displacement and rotation signal (x * ,y * ,z * ,α * ,β * ) and the five active disturbance rejection controllers of the actual values ​​(x, y, z, α, β) are combined to form a complete closed-loop control system containing an active disturbance rejection controller, a neural network inverse system expansion structure, a PWM amplifier and an axial-split magnetic levitation flywheel rotor system, and a complete magnetic levitation flywheel rotor active disturbance rejection control system based on the neural network inverse expansion structure is constructed. Finally, on the basis of achieving decoupling, stable suspension control of the automotive axial-split magnetic levitation flywheel rotor system under different working conditions is achieved.

[0047] The beneficial effects of the present invention are:

[0048] 1. Based on the neural network inverse system method, the present invention estimates the complex disturbances under vehicle-mounted conditions using the extended state observer algorithm in the active disturbance rejection controller. This is then introduced into the neural network to form an inverse system expansion structure, thereby enhancing the robustness of the flywheel rotor under vehicle-mounted disturbances.

[0049] 2. The present invention uses the neural network inverse method to achieve high-performance decoupling of the high coupling and gyroscopic effect of the axially split-phase magnetic levitation flywheel rotor system, further simplifying the model;

[0050] 3. The present invention uses an active disturbance rejection controller to perform closed-loop control on the axial-phase split magnetic levitation flywheel rotor system, further improving the anti-interference performance of the control system.

[0051] The present invention converts the motion system control of the axial-phase split-magnetic levitation flywheel rotor into the control of the second-order linear subsystem of the rotor position by utilizing the algorithm of the neural network inverse system. Then, the self-disturbance rejection controller is used to estimate the internal and external disturbances of the position model in real time, and the disturbances are introduced into the neural network to form an extended structure of the neural network inverse system, thereby enhancing the adaptability of the neural network inverse system, that is, making it approach the unmodeled dynamics, making the rotor system control simpler, more robust and more resistant to disturbances. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the rotor coordinate system of the axially split-phase magnetic levitation flywheel motor in the embodiment.

[0053] Figure 2 Schematic diagram of the structure of the composite controlled object in the embodiment.

[0054] Figure 3 Schematic diagram of the RBF neural network inverse system structure in the embodiment.

[0055] Figure 4 Schematic diagram of the structure of the dynamic neural network system in the embodiment.

[0056] Figure 5 Schematic diagram of the pseudo-linear system structure in the embodiment.

[0057] Figure 6 Schematic diagram of the structure of the second-order active disturbance rejection controller in the embodiment.

[0058] Figure 7 3 is a schematic diagram of sampling of the pseudo-linear system disturbance estimation value in the rotor x-axis direction in the embodiment.

[0059] Figure 8 It is a schematic diagram of the expanded structure of the neural network inverse system in the rotor x-axis direction in the embodiment.

[0060] Figure 9 This is a principle block diagram of the self-disturbance rejection control of the magnetic levitation flywheel rotor based on the inverse extension structure of the neural network. DETAILED DESCRIPTION

[0061] The following diagrams illustrate embodiments of the present invention. For clarity, many practical details are included in the following description. However, it should be understood that these practical details are not intended to limit the present invention. In other words, in some embodiments of the present invention, these practical details are not essential.

[0062] like Figure 9 As shown, the present invention is a control method for an axially split-phase magnetic levitation flywheel rotor system for a vehicle based on an inverse extended neural network structure, and the control method comprises the following steps:

[0063] Step 1: Considering the interference of both the vehicle's driving state and the road conditions, the on-board flywheel battery rotor dynamics analysis results are formed, and a flywheel rotor dynamics model is constructed. After linear amplification by a bipolar power amplifier, a composite flywheel rotor system is obtained;

[0064] Step 2: Use static neural network and integrator S -1 Construct dynamic neural network inverse system;

[0065] Step 3: The dynamic neural network inverse system constructed in step 2 is placed before the composite flywheel rotor system to form a pseudo-linear system. The composite flywheel rotor system obtained in step 1 is linearized and decoupled. The decoupled pseudo-linear system is then closed-loop controlled using an active disturbance rejection controller.

[0066] Step 4: Introduce the disturbance estimation value generated by the extended state observer in the active disturbance rejection controller into the model of the dynamic neural network inverse system in step 2, forming an extended structure of the neural network inverse system to increase the adaptability and anti-interference ability of the composite flywheel rotor system to disturbance changes;

[0067] Step 5: Replace the dynamic neural network inverse system in step 4 with the neural network inverse system expansion structure in step 4. By increasing the number of input nodes of the neural network and utilizing the interference estimation signal of the controlled flywheel rotor system, a complete magnetic levitation flywheel rotor self-disturbance rejection control system based on the neural network inverse expansion structure is finally formed.

[0068] The control method of the present invention is specifically described below with reference to embodiments.

[0069] The control method of the present invention comprises the following steps:

[0070] Step S1: construct a dynamic model in Simulink based on the dynamic equation of the axial-phase split magnetic levitation flywheel rotor system, and obtain a mathematical model of the composite controlled object through linear amplification by a bipolar power amplifier.

[0071] In step S11, an axial phase split magnetic suspension flywheel rotor motion coordinate system is constructed, such as Figure 1 As shown, where: a three-dimensional coordinate system is established with the rigid rotor mass center O as the origin, f ax and f bx are the radial x-direction and y-direction suspension forces acting on the rotor of the A-phase axial split-phase magnetic levitation flywheel motor, respectively, and f z is the axial suspension force applied in the z direction at the center of mass, l is the distance between the stator centers of phases A and B of the motor, l a and l b are the distances from the stator center of motor phase A and phase B to the center of mass O, l sa and l sa are the distances from the A-phase displacement sensor and the B-phase displacement sensor to the center of mass O, l s is the distance between the centers of the motor phase A and phase B displacement sensors.

[0072] Assume that the motor A phase displacement sensor detects the rotor radial translation displacement is x a and y a The motor B phase displacement sensor detects the rotor radial translation displacement as x b and y b , then the translational displacement at the rotor mass center O and the rotation angle of the rotor around the x-axis and y-axis are:

[0073]

[0074] In step S12, a dynamic model of the magnetic levitation rotor is constructed without considering the gyroscopic effect. According to Newtonian mechanics, the dynamic model of the magnetic levitation flywheel rotor can be obtained as follows:

[0075]

[0076] Where: m is the mass of the flywheel rotor; x, y, z are the translational displacements of the rotor in the x, y, and z axes under the coordinates of the center of mass; α and β are the rotation angles of the rotor around the x and y axes without considering the bending deformation of the rotor, and -α is a positive value. are their second-order derivatives respectively; f x and p x is the electromagnetic force and torque in the x direction at the center of mass; f y and p y is the electromagnetic force and torque in the y direction at the center of mass; f z is the electromagnetic force in the z direction at the center of mass; Δf and Δp are the external interference force and interference torque; fax and f bx are the electromagnetic forces in the x-axis direction in the motor A phase and motor B phase coordinate systems, respectively, and f ay and f by They are the electromagnetic forces in the y-axis direction of the motor phase A and phase B coordinate systems.

[0077] The motion equation of the rotor of the magnetic levitation flywheel system is expressed in matrix form:

[0078]

[0079] Right now:

[0080]

[0081] Where: The rotor mass matrix is Coordinate vector of the flywheel rotor's center of mass Gyro torque coefficient matrix Rotor arm coefficient matrix Magnetic bearing electromagnetic force The suspension force f of motor phase A and phase B in equation 2 is ax 、f bx 、f ay and f by Available current stiffness k at the equilibrium position i and displacement stiffness k r express:

[0082]

[0083] Where: x a 、y a 、x b and y b are the translational displacement of the rotor in the A-phase and B-phase coordinate systems of the motor respectively; z is the axial displacement of the motor; i ax 、i ay 、i bx and i by are the control currents of the suspension windings in the motor phase A and phase B coordinate systems respectively; i z is the control current of the motor axial suspension winding;

[0084] In step S13, the bipolar power amplifier that generates the actual control signal from the driving control signal is connected in series before the magnetic suspension flywheel rotor system. Figure 2 As shown, the module outputs the signal (x a 、y a 、x b and y b) is the rotor radial displacement signal detected by the sensor at the electromagnetic bearing, which needs to be converted to the displacement signal at the center of mass (x, y, α and β) for use in the control system. The motor axial displacement z does not require coordinate transformation. The coordinate transformation process is constructed as follows:

[0085]

[0086] Step S2: Using static neural network and integrator S -1 Construct a dynamic neural network inverse system of a composite control object.

[0087] Step S21: In this application, RBF neural network is used to perform inverse system mapping, such as Figure 3 As shown in the figure, a static neural network and 10 integrators S -1 Construct a neural network inverse system, use a static neural network to approximate the nonlinear mapping, and use the integrator to reflect the dynamic characteristics of the inverse system. As the expected output of the neural network, the output of the flywheel rotor system, y = [y1, y2, y3, y4, y5] T =[x a ,y a ,z,x b ,y b ] T As the input of the neural network, when selecting samples, first select the appropriate excitation current signal (multiple groups of different random square waves) to excite the controlled object. The controlled object outputs a displacement signal and uses a high-order digital filter to filter out the high-frequency noise in the sampled data, thereby obtaining a high-precision original data sample {u1, u2, u3, u4, u5, y1, y2, y3, y4, y5}, sampling the measurable internal state x, and designing a state observer to observe the internal state of the system that is difficult to measure directly. Then, the first-order and second-order derivatives of y are calculated using a high-order numerical differentiation method (five-point derivation method). The five-point derivation formula is as follows:

[0088]

[0089] In step S21, the input and output {u, y} signals are normalized to normalize the data with a large span to -1~1. The normalized data can eventually constitute the input sample set and the expected output sample set for training the neural network. In the simulation system, the sampled data is directly stored in the "To Workspace" module, the simulation time of the system is set to 30 seconds, the sampling interval is 0.01 seconds, and 3000 sets of data are obtained. Based on the data samples, the RBF neural network is used as the identification model of the inverse system. The RBF neural network structure is as follows: Figure 3As shown in the figure, the input nodes are set to 15, the output nodes are set to 4, and the number of hidden layer nodes is determined by experiments. 2500 sets of data are further selected from the training samples for training, and another 500 sets of data are used for detection and verification. Offline learning is performed on the composite controlled object to determine the weight coefficients of the static neural network layer and finally meet the accuracy requirements. After the static neural network is completed, 10 integrators S -1 Before being connected in series to the static neural network, since the input is the derivative of displacement and angle, x is obtained after the integration link. a ,y a ,z,x b ,y b The first-order and second-order derivatives and original values ​​are input into the static neural network, and the dynamic neural network structure is finally formed as follows Figure 4 shown.

[0090] Step S3: Place the constructed dynamic neural network inverse system before the composite flywheel rotor system to form a pseudo-linear system to linearize and decouple the composite flywheel rotor system obtained in step 1, and use the active disturbance rejection controller to perform closed-loop control of the decoupled pseudo-linear system.

[0091] Step S31 places the constructed dynamic neural network inverse system before the composite flywheel rotor system to form a pseudo linear system, such as Figure 5 It is equivalent to five second-order linear integral subsystems, that is, it is linearized and decoupled into five independent integral linear subsystems;

[0092] Step S32 transforms the rotor displacement signal detected by the displacement sensor to the rotor mass center. Assume that the rotor radial translation displacement detected by the motor A phase displacement sensor is x a and y a The motor B phase displacement sensor detects the rotor radial translation displacement as x b and y b , then the translational displacement at the rotor mass center O and the rotation angle of the rotor around the x-axis and y-axis are:

[0093]

[0094] The translational and rotational signals (x, y, z, α, β) at the center of mass of the rotor are obtained. The displacement in the Z direction does not pass through the sensor, so it does not need to be converted.

[0095] When a second-order linear active disturbance rejection controller is used in step S33, the model of the controlled object is regarded as:

[0096]

[0097] Where f(t) is the generalized disturbance, including the disturbance d caused by the vehicle's own driving state and the road conditions that the flywheel rotor system encounters when it is on board. In order to estimate the generalized disturbance, it is regarded as a new state variable, and z1=y, z3=f, y only represents the output, and then the given value (x * ,y * ,z * ,α * ,β * ) is input to the ADRC, and the tracking differentiator arranges the transition process for the input signal. * ,y * ,z * ,α * ,β * ) is processed, for example, for a given rotor translation displacement signal x * Perform smooth noise reduction processing and record the value of the processed input signal itself as Its derivative is denoted as Waiting to be transferred to the next step;

[0098] In step S34, the observer adopts a second-order extended state observer, such as Figure 6 , the actual displacement and rotation signals (x, y, z, α, β) generated by the feedback of the pseudo-linear composite system after adding the generalized disturbance and the input u are input into the second-order extended state observer LESO, and the states of each order of the control model, as well as the sum of the internal and external disturbances acting on the model and the unmodeled dynamics of the system are observed and estimated to obtain the actual signal estimate z1 and the disturbance estimate z3, and the actual signal estimate z1 is differentiated to obtain z2;

[0099] Step S35, the step S33 obtained The initial control quantity u0 is calculated by the linear feedback controller after being subtracted from z1 and z2 obtained in step S34, and then the disturbance estimation value z3 obtained by the observer is used to compensate for the initial control quantity u0 to obtain the final input quantity u.

[0100] Step S4: Introduce the disturbance estimation value generated by the extended state observer in the self-disturbance rejection controller into the model of the dynamic neural network inverse system in step 2, forming an extended structure of the neural network inverse system to increase the adaptability and anti-interference ability to the interference changes of the composite flywheel rotor system.

[0101] In step S41, the actual disturbance estimate z3 generated by the active disturbance rejection controller is used as a new state variable of the inverse system of the neural network in the vehicle-mounted state. For example, when the vehicle starts and accelerates, assuming that the acceleration remains constant and is consistent with the acceleration transmitted to the magnetic levitation flywheel battery, the flywheel rotor is initially stationary, and the center of mass of the flywheel rotor shaft will lag behind the axis in the forward direction. When a disturbance with a sudden increase in the acceleration value is set, the relative offset of the rotor will change, causing a change in the electromagnetic force. At this time, the state observer in the active disturbance rejection controller estimates this disturbance.

[0102] In step S42, a pseudo linear system closed-loop simulation model based on the active disturbance rejection controller is established, such as Figure 7 In the figure, a second-order pseudo-linear system is used to replace the decoupled flywheel rotor system. For example, the pseudo-linear system in the rotor x-axis direction is used for closed-loop control. This helps to ignore the interference in the neural network inverse system and the influence of the integral link module of the Simulink software itself in the rotor system. Then, a pseudo-linear system closed-loop simulation model based on the active disturbance rejection controller is formed to pave the way for the subsequent sampling of the disturbance estimation value.

[0103] In step S43, based on this, the influence of the unmodeled dynamics on the system is simulated using 10 different types of disturbance signals to simulate different operating conditions of the vehicle, such as starting acceleration, cruising at a constant speed, braking deceleration, climbing, turning, up and down vibration, lateral, longitudinal, yaw, and pitch. Various different disturbance signals are introduced into the closed-loop pseudo-linear system formed in the previous step to adjust the parameters of the active disturbance rejection controller. The corresponding second-order system state equation is Where z1 and z2 are the estimated values ​​of y and its changing speed respectively, z3 is the estimated value of the system generalized disturbance, β1, β2, β3 are observer parameters, b is an object parameter, u represents the control rate, and B is the control gain of the controller.

[0104] When the ESO can achieve z3≈f, the control system will be converted into two integral series links, and the expected equation in the form of transfer function is: The parameter that needs to be adjusted is the controller gain B, k p , k d and observer parameters β1, β2, β3.

[0105] In the actual control process, the low- and medium-frequency coefficients (β2 and β3) are significantly larger than the high-frequency coefficient β1, so after simplifying the formula, the disturbance estimate and the Laplace change of the system are obtained: in The larger the k value, the faster the ESO observation speed. Let β1=3ω0, k d =2ω c , ω0=4ω c ω c is the controller bandwidth, ω0 is the observer bandwidth. The six parameter adjustments are converted into three, namely the controller bandwidth ω c , the controller's control gain B and dynamic adjustment coefficient K. ω c When K and B remain unchanged, the dynamic characteristics of the system improve. c When K and B are constant, the dynamic characteristics of the system deteriorate as B increases. The three factors affect each other. Increase K and reduce B, then adjust ω. c Finally, observe the simulation diagram of the closed-loop control system. The control is at a higher dynamic characteristic and can obtain better control effect, which proves that the parameter adjustment is completed.

[0106] In step S44, after simulation, the value of the disturbance estimate z3 is extracted and stored in the "To Workspace" module. For example, the sampling of the pseudo-linear system disturbance estimate in the x-axis direction of the rotor is performed. After the signal is normalized, the input sample set and the expected output sample set for training the neural network can be formed. Since 3000 sets of data are obtained when building the neural network inverse system, z3 here also adjusts the sampling time, sets the time to 30 seconds, and the sampling time interval to 0.01 seconds, obtaining 3000 sets of data for data processing;

[0107] In step S45, the disturbance estimation value is used to construct the neural network inverse system. When the neural network fitting function is used, the neural network inverse system extension structure is fitted according to the corresponding relationship between the input value and the output value, such as Figure 8 For example, the expanded structure of the neural network inverse system in the x-axis direction of the rotor is taken into account. After introducing the disturbance estimation value, the expanded structure considers the unmodeled dynamics on the original basis, solves the disturbance problem encountered by the magnetic levitation flywheel battery rotor during operation, and makes the subsequent control more precise.

[0108] S5. Replace the dynamic neural network inverse system in step 4 with the neural network inverse system extension structure in step 4. By increasing the number of input nodes of the neural network and using the interference estimation signal of the controlled flywheel rotor system, a complete magnetic levitation flywheel rotor self-disturbance rejection control system based on the neural network inverse extension structure is finally formed. Specifically, the neural network inverse system extension structure obtained in the previous step is used to replace the neural network inverse system, and the reference value of the displacement and rotation signal (x * ,y * ,z * ,α * ,β *) and the actual values ​​(x, y, z, α, β) are combined to form a complete closed-loop control system containing an active disturbance rejection controller, a neural network inverse system expansion structure, a PWM amplifier, and an axial-phase split-phase magnetic levitation flywheel rotor system, and a complete magnetic levitation flywheel rotor active disturbance rejection control system based on a neural network inverse expansion structure is constructed, such as Figure 9 Finally, on the basis of achieving decoupling, stable suspension control of the automotive axial split-phase magnetic levitation flywheel rotor system is achieved under different working conditions.

[0109] The control method of the present invention first constructs a dynamic switching model of a magnetic levitation rotor based on the rotor dynamics equation of an axially split-phase magnetic levitation flywheel motor, and then utilizes a static neural network plus an integrator to construct a dynamic neural network, thereby realizing nonlinear dynamic decoupling between levitation forces and adopting an active disturbance rejection controller for closed-loop control. On this basis, the present invention considers two aspects of interference that the rotor used in electric vehicles may encounter during operation: the vehicle's own driving state (starting acceleration, braking deceleration, turning and climbing) and the road conditions (longitudinal vibration, lateral vibration and pitch vibration of the vehicle caused by uneven road surface). The extended state observer in the active disturbance rejection controller is used to estimate a relatively accurate value of the disturbance, which is then placed into the construction of the neural network inverse system to form an extended structure of the inverse system. This enables the neural network to have stronger adaptability to changes in the flywheel rotor system after considering unmodeled dynamics, and can further realize strong robust control of the axially split-phase magnetic levitation flywheel rotor used in electric vehicles under different operating conditions.

[0110] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. It will be apparent to those skilled in the art that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A control method for an automotive axially split-phase magnetically suspended flywheel rotor system based on an inverse extended neural network structure, characterized by: The control method comprises the following steps: Step 1: Considering the interference of both the vehicle's driving state and the road conditions, the on-board flywheel battery rotor dynamics analysis results are formed, and a flywheel rotor dynamics model is constructed. After linear amplification by a bipolar power amplifier, a composite flywheel rotor system is obtained; Step 2: Use a static neural network and 10 integrators S -1 Construct dynamic neural network inverse system; Step 3: The dynamic neural network inverse system constructed in step 2 is placed before the composite flywheel rotor system to form a pseudo-linear system. The composite flywheel rotor system obtained in step 1 is linearized and decoupled. The decoupled pseudo-linear system is then closed-loop controlled using an active disturbance rejection controller. Step 4: Introduce the disturbance estimation value generated by the extended state observer in the active disturbance rejection controller into the model of the dynamic neural network inverse system in step 2, forming an extended structure of the neural network inverse system to increase the adaptability and anti-interference ability of the composite flywheel rotor system to disturbance changes; Step 5: Replace the dynamic neural network inverse system in step 4 with the neural network inverse system extension structure in step 4. By increasing the number of input nodes of the neural network and using the interference estimation signal of the controlled flywheel rotor system, a complete magnetic levitation flywheel rotor active disturbance rejection control system based on the neural network inverse extension structure is finally formed. The step 4 specifically includes the following steps: Step 4-1: The actual disturbance estimate z3 generated by the ADRC is used as a new state variable of the neural network inverse system in the vehicle-mounted state. When the vehicle starts and accelerates, the acceleration is set to remain constant, consistent with the acceleration transmitted to the magnetic levitation flywheel battery. At this time, the flywheel rotor is initially stationary, and the center of mass of the flywheel rotor shaft will lag behind the axis in the forward direction. When the set acceleration value suddenly increases, the relative offset of the rotor will change, causing a change in the electromagnetic force. At this time, the ADRC estimates the displacement and angular operating state of the flywheel rotor under different operating conditions. Step 4-2: Establish a pseudo-linear system closed-loop simulation model based on the active disturbance rejection controller. The pseudo-linear system replaces the flywheel rotor system to obtain a second-order closed-loop active disturbance rejection control system. Step 4-3: Use different types of disturbance signals to simulate different vehicle operating conditions, and introduce various disturbance signals into the pseudo-linear system closed-loop simulation model formed in step 4-2. Adjust the parameters of the active disturbance rejection controller. After the adjustment, the closed-loop simulation can achieve good control results. Step 4-4: After simulation, extract the value of the disturbance estimate z3 and store it in the "To Workspace" module. After normalizing the signal, the input sample set and the expected output sample set for training the neural network can be formed. Step 4-5: The disturbance estimation value is included in the construction of the neural network inverse system. When using the neural network fitting function, the extended structure of the neural network inverse system is fitted according to the corresponding relationship between the input value and the output value.

2. The control method for an axially split-phase magnetic levitation flywheel rotor system for a vehicle based on an inverse extended neural network structure according to claim 1 is characterized by: The specific process of constructing the dynamic neural network inverse system in step 2 is: Step 2-1: Use a static neural network and 10 integrators S -1 Construct a neural network inverse system and transform the input of the flywheel rotor system into As the expected output of the neural network, the output of the flywheel rotor system y = [y1, y2, y3, y4, y5] T =[x a ,y a ,z,x b ,y b ] T As the input of the neural network, when selecting samples, first select a suitable excitation current signal to excite the controlled object. The controlled object outputs a displacement signal and uses a high-order digital filter to filter out high-frequency noise in the sampled data, thereby obtaining a high-precision original data sample {u1, u2, u3, u4, u5, y1, y2, y3, y4, y5}. The measurable internal state x is sampled, and a state observer is designed to observe the internal state of the system that is difficult to measure directly. Then, the five-point derivation method is used to calculate the first-order and second-order derivatives of y. The five-point derivation formula is as follows: Step 2-2: Normalize the input and output {u, y} signals to form the input sample set and expected output sample set for training the neural network. Step 2-3: In the simulation system, store the input sample set and expected output sample set in step 2-3 directly into the "ToWorkspace" module, set the simulation time of the system to obtain training data, and use the neural network as the identification model of the inverse system based on the data samples; Step 2-4: Further select data from the training data samples for training, testing and verification, and conduct offline learning on the composite controlled object to determine the weight coefficients of the static neural network layer and finally meet the accuracy requirements. After the static neural network is completed, the 10 integrators S -1 Before being connected in series to the static neural network, since the input is the derivative of displacement and angle, x is obtained after the integration link. a ,y a ,z,x b ,y b The first-order and second-order derivatives and original values ​​are input into the static neural network, and finally form a dynamic neural network structure.

3. The control method for an axially split-phase magnetic levitation flywheel rotor system for a vehicle based on an inverse extended neural network structure according to claim 2 is characterized by: The specific process of step 3 is as follows: Step 3-1: The dynamic neural network inverse system constructed in step 2 is placed before the composite flywheel rotor system to form a pseudo-linear system; Step 3-2: Transform the rotor displacement signal detected by the displacement sensor to the rotor mass center. Assume that the rotor radial translation displacement detected by the motor A phase displacement sensor is x a and y a The motor B phase displacement sensor detects the rotor radial translation displacement as x b and y b , then the translational displacement at the rotor mass center O and the rotation angle of the rotor around the x-axis and y-axis are: The translational and rotational signals (x, y, z, α, β) at the center of mass of the rotor are obtained. The displacement in the Z direction does not pass through the sensor, so it does not need to be converted. Step 3-3: When using a second-order linear active disturbance rejection controller, consider the model of the controlled plant as: Where f(t) is the generalized disturbance, including the disturbance d caused by the vehicle's own driving state and the road conditions that the flywheel rotor system encounters when it is on board. In order to estimate the generalized disturbance, it is regarded as a new state variable, and z1=y, z3=f, y only represents the output, and then the given value (x * ,y * ,z * ,α * ,β * ) is input to the ADRC, and the tracking differentiator arranges the transition process for the input signal. * ,y * ,z * ,α * ,β * ) and the value of the processed input signal itself is recorded as Its derivative is denoted as Waiting to be transferred to the next step; Step 3-4: Input the actual displacement and rotation signals (x, y, z, α, β) generated by the pseudo-linear composite system feedback after adding the generalized disturbance in step 3-3 and the input u into the second-order extended state observer. Observe and estimate the states of each order of the control model, the sum of the internal and external disturbances acting on the model, and the unmodeled dynamics of the system to obtain the actual signal estimate z1 and the disturbance estimate z3. Then, perform differentiation on the actual signal estimate z1 to obtain z2. Step 3-5: Get the The initial control quantity u0 is calculated by the linear feedback controller after being subtracted from z1 and z2 obtained in steps 3-4. The disturbance estimate z3 obtained by the second-order extended state observer is then used to compensate for the initial control quantity u0 to obtain the final input quantity u.

4. The control method of the automotive axial-phase split-magnetic levitation flywheel rotor system based on the inverse extended neural network structure according to claim 1 is characterized by: The step 5 is specifically as follows: using the extended structure of the neural network inverse system obtained in step 4 to replace the neural network inverse system, and replacing the reference value of the displacement and rotation signal (x * ,y * ,z * ,α * ,β * ) and the five active disturbance rejection controllers of the actual values ​​(x, y, z, α, β) are combined to form a complete closed-loop control system containing an active disturbance rejection controller, a neural network inverse system expansion structure, a PWM amplifier and an axial-split magnetic levitation flywheel rotor system switching model, and a complete magnetic levitation flywheel rotor active disturbance rejection control system based on the neural network inverse expansion structure is constructed. Finally, on the basis of achieving decoupling, stable suspension control of the automotive axial-split magnetic levitation flywheel rotor system under different working conditions is achieved.

5. The control method of the automotive axial-phase split-magnetic levitation flywheel rotor system based on the inverse extended neural network structure according to claim 1 is characterized by: The specific process of step 1 includes the following steps: Step 1-1: Design the dynamic equations of the axially split-phase magnetically suspended flywheel rotor system: Where: m is the mass of the flywheel rotor, x, y, z are the translational displacements of the rotor in the x, y, and z axes under the coordinates of the center of mass, α and β are the rotation angles of the rotor around the x and y axes without considering the bending deformation of the rotor, -α is a positive value, are their second-order derivatives, f x and p x is the electromagnetic force and torque in the x direction at the center of mass, f y and p y is the electromagnetic force and torque in the y direction at the center of mass, f z is the electromagnetic force in the z direction at the center of mass, Δf and Δp are the external interference force and interference torque, f ax and f bx are the electromagnetic forces in the x-axis direction in the motor A phase and motor B phase coordinate systems, respectively, and f ay and f by are the electromagnetic forces in the y-axis direction of the motor phase A and phase B coordinate systems; Step 1-2: Express the dynamic equation of the axially split-phase magnetic levitation flywheel rotor system in step 1-1 in matrix form: Right now: Where: The rotor mass matrix is Coordinate vector of the flywheel rotor's center of mass Gyro torque coefficient matrix Rotor arm coefficient matrix Magnetic bearing electromagnetic force Step 1-3: Connect the bipolar power amplifier that generates the actual control signal from the driving control signal in series before the magnetic levitation flywheel rotor to form a composite flywheel rotor system.

Images