CN104767449A - RBF Neural Network Adaptive Inverse Decoupling Control and Parameter Identification Method for Bearingless Induction Motor - Google Patents

Abstract

The invention discloses a bearing-free asynchronous motor RBF neural network self-adaptive inverse decoupling control and parameter identification method. An SVPWM module, a voltage inverter, a bearing-free asynchronous motor and a load of the bearing-free asynchronous motor form a whole serving as a composite controlled object. Two radial basis function neural networks are adopted to achieve inverse control and parameter identification conducted on the composite controlled object. A self-adaptive inverse controller is formed by using an RBF neural network through learning, and is serially connected in front of the composite controlled object, errors of a feedback signal and a given signal are input into an inverse controller, and accordingly closed-loop control is formed, then a self-adaptive parameter identifier is formed by using one RBF neural network through learning and identifies output quantity speed and displacement of the composite controlled object, speed-less and displacement-free sensor control is achieved, online learning of an estimation signal is aided by means of a learning algorithm, and non-linear dynamic decoupling control of the bearing-free asynchronous motor is achieved. The bearing-free asynchronous motor RBF neural network self-adaptive inverse decoupling control and parameter identification method is high in control speed and higher in identification accuracy, and a control system is excellent.

Description

RBF Neural Network Adaptive Inverse Decoupling Control and Parameter Identification Method for Bearingless Induction Motor

technical field

The invention is a multi-variable nonlinear bearingless asynchronous motor RBF neural network adaptive inverse control and parameter identification method, which is suitable for high-performance control of the bearingless asynchronous motor. Bearingless asynchronous motors inherit the advantages of magnetic bearing motors, and have the characteristics of no friction, no wear, no lubrication and sealing, sterility, no pollution, and long life. They are very suitable for injection into high-speed precision CNC machine tools, high-pressure sealed pumps, and special robots. , High-speed gyroscope, satellite flywheel, high-speed aircraft and control device, high-speed centrifuge, high-speed flywheel energy storage and other high-speed drive high-tech fields have broad application prospects and belong to the technical field of electric drive control equipment.

Background technique

The bearingless asynchronous motor has complex electromagnetic relations inside, so it is a kind of multi-variable, nonlinear, and strongly coupled controlled object, and it is very difficult to achieve accurate control of its radial displacement and speed. In order to realize the stable levitation of the rotor of the bearingless asynchronous motor and the operation following the given speed, it is necessary to decouple the torque force and levitation force of the motor.

However, the control strategy of dynamic decoupling control has always been a difficult point in realizing the stable operation of bearingless asynchronous motors. Common controls include air-gap field-oriented control and rotor field-oriented control. Experiments have proved that these two methods can achieve relatively stable control of bearingless asynchronous motors. Although the air-gap field-oriented control method can achieve decoupling control between electromagnetic torque and radial levitation force, this algorithm is greatly affected by motor parameters (such as rotor resistance, rotor leakage inductance, etc.), and there is a problem of stable rotation torque, the scope of application is limited; the rotor field oriented control method can achieve the decoupling between the torque current and the excitation current, but only when the rotor flux linkage is stable and remains constant can the decoupling of the electromagnetic torque and the rotor flux linkage be realized , belonging to steady-state decoupling and cannot achieve dynamic decoupling. BP neural network is applied to the control of bearingless asynchronous motors and achieves good control results. However, BP neural network has shortcomings such as slow convergence speed and easy to fall into local minimum in function approximation, and it does not match the biological background theoretically.

In order to further improve the dynamic performance of bearingless asynchronous motors, it is necessary to consider the combination of dynamic decoupling and multivariable coordinated control of bearingless asynchronous motors, and then a decoupling controller for bearingless asynchronous motors with more compact structure and better performance is required.

Related domestic patent applications: 1) Patent application number CN20061038711.3, titled: Control method of bearingless AC asynchronous motor neural network inverse decoupling controller, this invention patent is designed for bearingless AC asynchronous motor neural network inverse solution Coupling controller; 2) Patent application number CN200510038099.5, titled: Bearingless Switched Reluctance Motor Radial Neural Network Inverse Decoupling Controller and Construction Method, this invention designs radial neural network inverse solution for magnetic levitation switched reluctance motor Coupling controller; 3) Patent application number CN200510040065.X, based on neural network inverse five-degree-of-freedom bearingless permanent magnet synchronous motor control system and control method, this invention is a control method designed for five-degree-of-freedom bearingless permanent magnet synchronous motor; 4 ) Article number 0258-8013 (2004) 07-0117-05 RBF neural network-based ultrasonic motor parameter identification is a method for ultrasonic motor parameter identification; the idea of neural network inverse controller used in the above three inventions to control the motor is somewhat related to this patent However, what the neural network used in this paper is the RBF neural network, which is different from the BP network they adopt; compared with the article 4, the present invention has essential differences in the structure, mathematical model, control method, control difficulty and requirements of the motor. The design of the RBF neural network self-adaptive inverse control system for the bearing asynchronous motor has no related patents and documents at present.

Contents of the invention

The purpose of the present invention is to use RBF neural network self-adaptive inverse controller to carry out nonlinear dynamic decoupling control on levitation force, torque force and rotor flux linkage for the nonlinear and strongly coupled complex system of bearingless asynchronous motor, and to provide a It can make the bearingless asynchronous motor have excellent dynamic and static performance, and has strong robustness against motor parameter changes and load disturbance, and can effectively improve various control performance indicators of the bearingless asynchronous motor; in addition, the RBF neural network is adopted The network adaptive identifier realizes on-line identification of the output of the bearingless asynchronous motor, such as radial displacement, rotational speed and flux linkage.

The technical solution of the present invention is: a bearingless asynchronous motor RBF neural network adaptive inverse decoupling control and parameter identification method, including steps:

Step 1. Use commonly used sensors to detect voltage, current, and rotational speed signals. The signals are sent to the flux linkage observation model after 3s/2r coordinate transformation to obtain flux linkage information required for flux linkage closed-loop control and neural network training;

Step 2, connect SVPWM algorithm module 1 and voltage inverter module 1 in series to form extended SVPWM voltage inverter module 1, connect SVPWM algorithm module 2 and voltage inverter module 2 in series to form extended SVPWM voltage type inverter module two;

Step 3, construct the bearingless asynchronous motor and its load model, and combine the extended SVPWM voltage-type inverter module 1, the extended SVPWM voltage-type inverter module 2, and the bearingless asynchronous motor and its load model as a whole to form a composite control object;

Step 4: Construct the inverse controller of the compound controlled object through the RBF neural network RBFNNC, use the method of combining offline and online to train and obtain the structure and parameters of the RBF neural network RBFNNC, and place the trained RBF neural network RBFNNC in the compound controlled object The linear control system is formed before the control object, so as to realize the decoupling control of the bearingless asynchronous motor;

Step 5: Construct the identifier of the compound controlled object through the RBF neural network RBFNNI, use the method of combining offline and online to train and obtain the structure and parameters of the RBF neural network RBFNNI, and replace the sensor with the identification signal after the identification accuracy meets the design requirements The detected signal realizes sensorless control.

Further, the 3s/2r coordinate transformation in the step 1 can be divided into the first coordinate transformation and the second coordinate transformation, the first coordinate transformation is to convert the phase current of the stator winding phase current of the bearingless asynchronous motor detected by the Hall current sensor to i _1a , i _1b , i _1c obtain the current i _1d , i _1q in the rotating coordinate system through Clark transformation and Park transformation; the second coordinate transformation is to transform the bearingless asynchronous motor stator winding phase voltage U _1a , phase voltage detected by the Hall voltage sensor U _1b and U _1c obtain voltages U _1d and U _1q in the rotating coordinate system through Clark transformation and Park transformation.

Further, the flux observation model in the step 1 includes a stator flux observation model and a rotor flux observation model;

The stator flux linkage observation model is to transform the current i _1d , i _1q and voltage U _1d , U _1q in the rotating coordinate system through function transformation to obtain the stator flux linkage components ψ _1d , ψ _1q in the rotating coordinate system:

ψ _1d ＝∫(U _1d -Ri _1d )dt-L ₁ i _1d

ψ _1q ＝∫(U _1q -Ri _1q )dt-L ₁ i _1q

The rotor flux observation model is to transform the current i _1d , i _1q in the rotating coordinate system and the stator flux components ψ _1d , ψ _1q in the rotating coordinate system through function transformation to obtain the rotor flux components ψ _dr , ψ _qr in the rotating coordinate system :

$\begin{array}{c}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; dr</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; d</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; d\&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; dr</span>}}}{\mathrm{<span\; class="notranslate">\; dt</span>}}<span\; class="notranslate">\; )</span>\\ {\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; qr</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; q</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; d</span>}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; qr</span>}}}{\mathrm{<span\; class="notranslate">\; dt</span>}}<span\; class="notranslate">\; )</span>\end{array}<span\; class="notranslate">\; .</span>$

Further, in the step 2, the SVPWM algorithm module 1 converts the given voltage signals U _α1s *, U _β1s * into duty ratio signals S _a1s , S _b1s , S _c1s , and outputs the duty ratio signals to the voltage type Inverter 1 generates voltage signals U _a1s , U _b1s , U _c1s to control the torque winding system;

The SVPWM algorithm module 2 converts the given voltage signals U _α2s *, U _β2s * into duty ratio signals S _a2s , S _b2s , S _c2s , and outputs the duty ratio signals to the voltage-type inverter 2 to generate a voltage signal U _a2s , U _b2s , U _c2s to control the suspension winding system.

Further, the mathematical model of the torque winding system of the bearingless asynchronous motor and its load model in step 3 is a common cage-type asynchronous motor mathematical model; the suspension force mathematical equation of the bearingless asynchronous motor and its load model of the suspension winding system is as follows:

F _x ＝M(-i _d1s i _d2s +i _q1s i _q2s )

F _y ＝M(i _q1s i _d2s +i _d1s i _q2s )

In the formula, M is the mutual inductance coefficient between the torque winding and the suspension winding;

The state equation of the suspension winding system of the bearingless asynchronous motor and its load model is as follows:

$\begin{array}{c}\mathrm{<span\; class="notranslate">\; m</span>}\stackrel{<span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; sx</span>}}\\ \mathrm{<span\; class="notranslate">\; m</span>}\stackrel{<span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&CenterDot;</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; sy</span>}}\end{array}$

In the formula, m is the mass of the rotor; F _sx , F _sy inherent Maxwell force, its expression is:

F _sx =k _s x

F _sy = k _s y

In the formula r is the radius of the rotor; l is the length of the rotor shaft; μ ₀ is the air permeability; δ is the length of the air gap; k is the attenuation factor, generally 0.3.

Further, in the step 4-5, the off-line training method of RBF neural network RBFNNC and RBF neural network RBFNNI structure and parameter determination is:

Determine the number of hidden layer nodes and their center and width through offline training, and calculate the initial value of the connection weight between the hidden layer and the output layer. The input samples of RBFNNC are displacement X, Y, rotational speed ω _r , flux linkage ψ The error between the given value and the actual value of _r and the output value {e ₁ , e ₂ , e ₃ , e ₄ , e _c1 , e _c2 , e _c3 , e _c4 } of the error signal through the e _c function module, the output sample is After coordinate transformation {U _1α , U _1β , U _2α , U _2β }; the input samples of RBFNNI are the input and output of the delayed compound controlled object {U _1α , U _1β , U _2α , U _2β , X , Y, ω _r , ψ _r }, the output sample is the identified compound plant output {X′, Y′, ω _r ′, ψ _r ′};

In the training process, the number of nodes is adaptively increased according to certain rules, and the hidden layer units that have too little effect on the output signal are deleted according to the rules, and the system nonlinear mapping is effectively realized with the least hidden layer units.

Further, in the step 4-5, the online training method for determining the structure and parameters of the RBF neural network RBFNNC and RBF neural network RBFNNI adopts the recursive least square method learning rule.

The advantages of the present invention are:

1. Bearingless asynchronous motor not only inherits the advantages of magnetic bearing support motor, but also solves the disadvantages of traditional support devices such as complex structure, large volume, high cost, low efficiency and high failure rate, and is more reasonable and practical than magnetic bearing motor. 1) The axial space is greatly shortened, the utilization rate of the axial direction is improved, and the limitation of high power and ultra-high speed can be broken through; 2) The power amplifier circuit in the radial suspension control system of the bearingless asynchronous motor adopts a three-phase motor based on the SVPWM algorithm. The power inverter circuit makes the motor control power consumption low, the torque ripple is small, the sine degree of the phase current is increased, and the THD of the current is reduced.

2. Through the RBF neural network adaptive inversion, the dynamic decoupling between the torque force and the radial levitation force is performed on the bearingless asynchronous motor to realize the control of the position system, rotor speed and flux linkage. Neural network inverse has its unique characteristics: RBF neural network has good biological background and function approximation ability, not only has simple structure, fast convergence speed, strong generalization ability, but also has global optimal and best approximation properties, which further simplifies control device network structure.

3. The prominent feature in the present invention is to adopt the RBF neural network self-adaptive identifier to identify the radial displacement, rotating speed and flux linkage information of the bearingless asynchronous motor without knowing the precise parameters of the motor by the easily detected voltage and current signals, and The recognition accuracy is high, so as to realize sensorless control, reduce system cost and improve system reliability, especially suitable for harsh environments and occasions with high system requirements.

4. In the training process, the present invention adaptively increases the number of nodes according to certain rules, and deletes hidden layer units that have too little effect on the output signal according to the rules, so as to ensure that the network structure is simple and compact, and is effective with the least hidden layer units Realize the nonlinear mapping of the system. Learning by recursive least squares method has the advantage of adjusting the network connection weight with supervision according to recursive least squares method through online training, and enhancing the robustness of the inverse system.

The RBF neural network adaptive inverse control system of the bearingless asynchronous motor based on the RBF neural network adaptive inverse structure of the present invention improves the control performance of the bearingless asynchronous motor, and is also applicable to other bearingless motor control systems and various types of magnetic bearing support motor control system. Therefore, the application prospect of this control method is very broad, and it also has important application value for other bearingless motors.

Description of drawings

Fig. 1 is a schematic diagram of a rotor flux observer of a bearingless asynchronous motor;

Figure 2 is a schematic diagram of a compound controlled object composed of the expanded SVPWM voltage-type inverter module 1, the expanded SVPWM voltage-type inverter module 2, and the bearingless asynchronous motor and its load model as a whole;

Fig. 3 is an internal schematic diagram and an equivalent diagram of the RBF neural network;

Fig. 4 is a schematic diagram of composite input and output learning samples composed of the RBF neural network inverse network and the RBF neural network parameter identification network;

Fig. 5 is a block diagram of a bearingless asynchronous motor RBF neural network adaptive inverse decoupling control system;

Figure 6 is a block diagram of the RBF neural network adaptive inverse control and parameter identification system for a bearingless asynchronous motor.

Detailed ways

Embodiments of the present invention are as follows: first, a flux observer composed of a commonly used current, voltage, speed, flux observation model and Park transformation and Clark transformation is used to estimate the rotor flux of the bearingless asynchronous motor required for the flux linkage closed loop. chain information. The two SVPWM and voltage-type inverter modules, as well as the bearingless asynchronous motor and its load model are taken as a whole to form a composite controlled object. The controlled variables of the composite controlled object are the rotor radial displacement and rotational speed of the bearingless asynchronous motor and flux linkage; a RBF neural network is used to construct the inverse controller of the compound controlled object, and the input of the inverse controller is the error signal of the given signal and the feedback signal to form a closed loop; in addition, an RBF neural network RBFNNI is used to realize the speed control of the controlled object and displacement signal identification; both RBF neural networks use a three-layer feed-forward network, including an input layer (8 nodes), a hidden layer and an output layer (4 nodes). Combined with online learning to realize network structure initialization and weight optimization; finally, the RBF neural network adaptive inverse system and adaptive parameter identification, two SVPWM and voltage inverter modules together constitute the RBF neural network adaptive inverse The self-adaptive inverse control system of the bearing asynchronous motor realizes the independent control of the torque force and the radial suspension force of the bearingless asynchronous motor, so as to realize the dynamic decoupling of the motor and the parameter identification of the controlled object.

The specific implementation manner of the present invention will be further described below in conjunction with the accompanying drawings.

As shown in Figure 1, the flux observer of the bearingless asynchronous motor is constructed: it is composed of two coordinate transformations 11 and 12, the stator flux observation model 13 and the rotor flux identification model 14; one of the coordinate transformations is to convert the bearingless asynchronous motor Stator winding phase current i _1a , i _1b , i _1c collect the winding phase current i _1d , i _1q of the bearingless asynchronous motor through 3s/2r transformation 11; another 3s/2r transformation 12 coordinate transformation is to transform the bearingless asynchronous motor stator The winding phase voltages U _1a , U _1b , U _1c are collected through the 3s/2r transformation 12 to collect the winding phase voltages U _1d , U _1q of the bearingless asynchronous motor; then the required flux linkage value is obtained through the corresponding flux linkage identification module. The stator flux linkage observation model 13 obtains the stator flux linkage components ψ _1d and ψ _1q in the rotating coordinate system by transforming the current i _1d , i _1q and the voltages U _1d , U _1q through functions:

ψ _1d ＝∫(U _1d -Ri _1d )dt-L ₁ i _1d

ψ _1q ＝∫(U _1q -Ri _1q )dt-L ₁ i _1q

The rotor flux observation model 14 is to transform the current i _1d , i _1q in the rotating coordinate system and the stator flux components ψ _1d , ψ _1q in the rotating coordinate system through function transformation to obtain the rotor flux components ψ _dr , ψ in the rotating coordinate system _qr :

$\begin{array}{c}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; dr</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; d</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; +</span>\frac{{\mathrm{<span\; class="notranslate">\; d\&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; dr</span>}}}{\mathrm{<span\; class="notranslate">\; dt</span>}}<span\; class="notranslate">\; )</span>\\ {\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; qr</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; q</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; d</span>}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; qr</span>}}}{\mathrm{<span\; class="notranslate">\; dt</span>}}<span\; class="notranslate">\; )</span>\end{array}<span\; class="notranslate">\; .</span>$

In Fig. 2, SVPWM algorithm module 1 21 and SVPWM algorithm module 2 31 are common SVPWM algorithms; voltage-type inverter 1 22 and voltage-type inverter 2 32 are voltage-type power inverter IPM; given voltage signal U _α1s *, U _β1s * Obtain duty cycle signals S _a1s , S _b1s , S _c1s through SVPWM algorithm module 121, and output the duty cycle signals to voltage type inverter 122 to generate voltage signals U _a1s , U _b1s , U _c1s to control the torque system; the given voltage signals U _α2s *, U _β2s * are obtained by the SVPWM algorithm module 2 31 to obtain the duty cycle signals S _a2s , S _b2s , S _c2s , and output the duty cycle signals to the voltage source inverter 2 32 Generate voltage signals U _a2s , U _b2s , U _c2s to control the suspension system.

As shown in Figure 2, the two SVPWM21, 31 and voltage-type inverter modules 22, 32, as well as the bearingless asynchronous motor and its load model 1 form a composite controlled object 4 as a whole, and the controlled object of the composite controlled object The quantity input is {U _1α , U _1β , U _2α , U _2β } voltage signal, and the rotor radial displacement, speed and flux linkage of the bearingless asynchronous motor are output.

The core design of the present invention is the design and learning method of the RBF neural network. First establish the RBF neural network RBFNN 75 as shown in Figure 3. It can be seen from the figure that the input layer 71 has 8 nodes, and the hidden layer 72 is a node composed of a radial basis function. The radial basis function adopts a Gaussian kernel function. The output of the layer unit is:

${\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; -</span>\frac{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>$

In the formula, h _i is the output of the i-th hidden layer node; x is the input vector; C _i is the center of the i-th hidden layer node; b _i is the width of the hidden layer; ||*|| is the Euclidean norm 4 output nodes 74 are obtained through addition; RBF neural network RBFNNC 51 and RBF neural network RBFNNI 52 are established according to the above method.

The RBF neural network is trained by combining offline learning and online learning. The training samples are shown in Figure 4: the input samples of RBFNNC 51 are displacement X, Y, rotational speed ω _r , flux linkage ψ _r given value and feedback value error signal And the error signal through the e _c function:

e _ci (t)＝e _i (t)-e _i (t-1) (i=1, 2, 3, 4)

The output value {e ₁ , e ₂ , e ₃ , e ₄ , e _c1 , e _c2 , e _c3 , e _c4 }, the output sample is {U _1α , U _1β , U _2α , U _2β } after coordinate transformation; The input samples of RBFNNI52 are the input and output of the delayed composite plant {U _1α , U _1β , U _2α , U _2β , X, Y, ω _r , ψ _r }, and the output samples are the output of the identified composite plant {X′, Y′, _ωr ′, _ψr ′}. Offline training is performed first, the initial value of the hidden layer unit of the RBF neural network is set to zero, and the hidden layer nodes are adaptively added according to the "novelty" condition, and a deletion strategy is adopted. As the learning continues, those contributions to the output Delete nodes that are no longer active to a certain extent to ensure a simple and compact network structure, effectively realize the nonlinear mapping of the system with the least hidden layer units, and perform offline training to determine the network structure and initialize the hidden layer center and output weights ; After online learning, the parameters of the network are corrected online through the gradient descent method, so that the network can adapt to environmental changes.

The specific algorithm for effectively realizing the nonlinear mapping of the system with the least number of hidden layer units is as follows:

For the i-th learning sample (x(i), Z(i)),

Step 7.1, initialize c _j to any real number, and calculate the output of each hidden layer unit of the RBF neural network respectively and the network outputs y(i):

Step 7.2, calculate the error ||E(i)||=||Z(i)-y(i)||, where Z(i) is the target output, that is, the record of the system after sampling and conditioning, y(i ) is the actual output of the network, and calculate the distance between the sample and the existing hidden layer:

d _j =||x(i)-c _j ||, j=1, 2,..., m

In the formula, m is the number of existing hidden layer units;

Let d _min =min(d _j )

Step 7.3, if ||E _i ||＞ε, d _min ＞λ(i), then:

λ(i)=max(λ _max γ ⁱ , λ _min )

In the formula, ε is the expected accuracy of the network; λ(i) is the fitting accuracy of the network at the i-th input, and λ(i) decreases from λ _max to λ _min as the learning progresses; γ is the attenuation factor, 0<γ<1, then add a hidden layer unit, its parameters:

c _k = x _i

${\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; p</span>}}{<span\; class="notranslate">\; (</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

In the formula, c _j is the center of the p hidden layer units closest to the input sample, here p=2;

Step 7.4, otherwise, adjust the network connection weight according to the recursive least squares method;

Step 7.5, if all n samples of contact input satisfy:

In the formula, σ is a predefined constant. If the condition in the formula is satisfied at the same time when i=1 and 2, when the jth hidden layer node is deleted;

In step 7.6, input the i+1th group of samples, and repeat the above process.

In said steps 4-5, the online training method of RBF neural network RBFNNC51 and RBF neural network RBFNNI52 structure and parameter determination adopts the recursive least squares method learning rule:

Step 8.1, for the kth group of inputs, re-establish the network output equation:

In the formula, ω(k), u(k) are the weight vector and the radial basis function vector respectively, and H represents the conjugate transpose;

Step 8.2, let P(0)=δ ^-1 I, ω(0)=0;

Step 8.3, Calculate $v\left(k\right)=\frac{{\mathrm{\&eta;}}^{-1}P(k-1)u\left(k\right)}{1+{\mathrm{\&eta;}}^{-1}{u}^h\left(k\right)P(k-1)u\left(k\right)}$

ζ(k)=y(k)-ω ^H (k-1)u(k)

ω(k)=ω(k-1)+v(k)ζ ^* (k)

P(k)=η ^-1 P(k-1)-η ^-1 v(k)uH(k)P(k-1)

In the formula, δ is a small positive constant; η is the forgetting factor, 0≤η≤1; * indicates complex conjugate.

As shown in Figure 5, the established RBF neural network inverse 50 is combined with two expanded SVPWM voltage-type inverters 2 and 3 to form an RBF neural network adaptive inverse controller 6, which is placed before the bearingless asynchronous motor to realize the solution coupling control.

The control system is constituted as shown in Figure 6: RBF neural network RBFNNC51 and RBF neural network RBFNNI52, two SVPWM 21, 31 and voltage-type inverter modules 22, 32 together form RBF neural network adaptive inverse bearingless asynchronous motor adaptive inverse control And parameter identification system7.

The present invention can be realized according to the above-mentioned drawings and related steps.

The technical principle of the present invention is further summarized below.

The principle of the invention is to change the traditional bearingless asynchronous motor using rotor field orientation and air gap magnetic field orientation decoupling control strategy, design and invent a non-linear dynamic decoupling control for bearingless asynchronous motor using RBF neural network self-adaptive inverse control system .

The RBF neural network adaptive inverse system control and parameter identification method of the present invention uses the RBF neural network adaptive inverse controller to replace the team response inverse system model in the existing decoupling control method, making up for such problems as air gap magnetic field orientation and rotor magnetic field orientation. Compared with the inverse decoupling of BP neural network, this method better realizes the dynamic decoupling of torque force and suspension force, has high online identification accuracy, reduces the use of sensors and reduces costs, and enhances the reliability of the speed control system. The bearing asynchronous motor speed control system has stronger anti-interference and robustness.

The control method of the RBF neural network adaptive inverse control and parameter identification system of the bearingless asynchronous motor is as follows: firstly, the commonly used current, voltage, speed, flux linkage observation model and a flux linkage observer composed of Park transform and Clark transform are used to Rotor flux information of a bearingless asynchronous motor required for estimating flux closure. The two SVPWM and voltage-type inverter modules, as well as the bearingless asynchronous motor and its load model are taken as a whole to form a composite controlled object. The controlled variables of the composite controlled object are the rotor radial displacement and rotational speed of the bearingless asynchronous motor and flux linkage; a RBF neural network is used to construct the inverse system of the compound controlled object, and the input of the inverse system is the error signal of the given signal and the feedback signal to form a closed loop to realize the decoupling control between the torque force and the radial suspension force; in addition A RBF neural network is used to realize the sensorless control of speed and displacement, and realize the nonlinear dynamic decoupling control of the bearingless asynchronous motor; finally, the RBF neural network adaptive inverse controller and adaptive parameter identifier, two SVPWM and voltage type The inverter modules together constitute the RBF neural network adaptive inverse bearingless asynchronous motor adaptive inverse control system to realize the independent control of the torque force and radial suspension force of the bearingless asynchronous motor, so as to realize the stable suspension and operation of the motor rotor.

Among them, the flux observer mentioned above is composed of two coordinate transformations, stator flux observation model and rotor flux identification model; one of the coordinate transformations is to pass the phase currents i _1a , i _1b , i _1c of the stator windings of the bearingless asynchronous motor through Clark transformation and Park transformation to collect the winding phase current i _1d , i _1q of the bearingless asynchronous motor; another coordinate transformation is to use the bearingless asynchronous motor stator winding phase voltage U _1a , U _1b , U _1c through Clark transformation and Park transformation to obtain Collect the winding phase voltage U _1d , U _1q of the bearingless asynchronous motor; then obtain the required flux linkage value through the corresponding flux linkage identification module.

It should be understood that the above-mentioned embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, those skilled in the art all fall into the appended claims of the present application to the amendments of various equivalent forms of the present invention limited range.

Claims (7)

1. The RBF neural network adaptive inverse decoupling control and parameter identification method of a bearingless asynchronous motor is characterized in that it comprises steps: Step 1. Use commonly used sensors to detect voltage, current, and rotational speed signals. The signals are sent to the flux linkage observation model after 3s/2r coordinate transformation to obtain flux linkage information required for flux linkage closed-loop control and neural network training; Step 2, the SVPWM algorithm module one (21) and the voltage type inverter module one (22) are connected in series to form an expanded SVPWM voltage type inverter module one (2), and the SVPWM algorithm module two (31) and the voltage type inverter module two (31) are connected in series Inverter module two (32) are connected in series to form extended SVPWM voltage type inverter module two (3); Step 3, constructing the bearingless asynchronous motor and its load model (1), the extended SVPWM voltage-type inverter module one (2), the extended SVPWM voltage-type inverter module two (3) and the bearingless asynchronous motor and Its load model (1) forms a compound controlled object (4) as a whole; Step 4, construct the inverse controller of the compound controlled object (4) through the RBF neural network RBFNNC (51), use the method of combining offline and online to train and obtain the structure and parameters of the RBF neural network RBFNNC (51), and train the The RBF neural network RBFNNC (51) is placed before the compound controlled object (4) to form a linear control system, thereby realizing the decoupling control of the bearingless asynchronous motor; Step 5, construct the identifier of the compound plant (4) through the RBF neural network RBFNNI (52), use the method of combining offline and online to train and obtain the structure and parameters of the RBF neural network RBFNNI (52), when the identification accuracy reaches After the design requirements, the identification signal is used to replace the signal detected by the sensor to realize sensorless control. 2. bearingless asynchronous motor RBF neural network adaptive inverse decoupling control and parameter identification method according to claim 1, is characterized in that, the 3s/2r coordinate transformation in the described step 1 can be divided into the first coordinate transformation ( 11) and the second coordinate transformation (12), the first coordinate transformation (11) is to pass the phase current i 1a , i 1b , i 1c of the stator winding phase current i _1a , i _1b , i _1c of the bearingless asynchronous motor detected by the Hall current sensor through the Clark transformation and The current i _1d and i _1q in the rotating coordinate system are obtained by Park transformation; the second coordinate transformation (12) is to transform the phase voltage U _1a , U _1b , U _1c of the stator winding of the bearingless asynchronous motor detected by the Hall voltage sensor through the Clark transformation and The voltages U _1d and U _1q in the rotating coordinate system are obtained through Park transformation. 3. the bearingless asynchronous motor RBF neural network adaptive inverse decoupling control and parameter identification method according to claim 2, is characterized in that, the flux observation model in the described step 1 comprises stator flux observation model (13) and rotor flux observation model (14); The stator flux linkage observation model (13) is to transform the current i _1d , i _1q and voltages U 1d , _{U 1q in} the rotating coordinate system to obtain the stator flux linkage components ψ _1d , ψ _1q in the rotating coordinate system: ψ _1d ＝∫(U _1d -Ri _1d )dt-L ₁ i _1d ψ _1q ＝∫(U _1q -Ri _1q )dt-L ₁ i _1q The rotor flux observation model (14) is to transform the current i _1d , i _1q in the rotating coordinate system and the stator flux components ψ _1d , ψ _1q in the rotating coordinate system through function transformation to obtain the rotor flux component ψ _dr in the rotating coordinate system , ψ _qr : ${\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; dr</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; d</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&omega;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; d</span>}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; dr</span>}}}{\mathrm{<span\; class="notranslate">\; dt</span>}}<span\; class="notranslate">\; )</span>$ ${\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; qr</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; q</span>}}{\mathrm{<span\; class="notranslate">\; L</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; q</span>}}<span\; class="notranslate">\; +</span>\frac{\mathrm{<span\; class="notranslate">\; d</span>}{\mathrm{<span\; class="notranslate">\; \&psi;</span>}}_{\mathrm{<span\; class="notranslate">\; qr</span>}}}{\mathrm{<span\; class="notranslate">\; dt</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; .</span>$ 4. The bearingless asynchronous motor RBF neural network adaptive inverse decoupling control and parameter identification method according to claim 1, characterized in that, in the step 2, The SVPWM algorithm module one (21) converts the given voltage signals U _α1s *, U _β1s * into duty cycle signals S _a1s , S _b1s , S _c1s , and outputs the duty cycle signals to the voltage type inverter one ( 22) Generate voltage signals U _a1s , U _b1s , U _c1s to control the torque winding system; The SVPWM algorithm module two (31) converts the given voltage signals U _α2s *, U _β2s * into duty cycle signals S _a2s , S _b2s , S _c2s , and outputs the duty cycle signals to the voltage type inverter two ( 32) Generate voltage signals U _a2s , U _b2s , U _c2s to control the suspension winding system. 5. bearingless asynchronous motor RBF neural network self-adaptive inverse decoupling control and parameter identification method according to claim 1, is characterized in that, in the described step 3, the torque of bearingless asynchronous motor and its load model (1) The mathematical model of the winding system is a common cage-type asynchronous motor; the mathematical equation of the suspension force of the suspension winding system of the bearingless asynchronous motor and its load model (1) is as follows: F _x ＝M(-i _d1s i _d2s +i _q1s i _q2s ) F _y ＝M(i _q1s i _d2s +i _d1s i _q2s ) In the formula, M is the mutual inductance coefficient between the torque winding and the suspension winding; The state equation of the suspension winding system of the bearingless asynchronous motor and its load model (1) is as follows: $\mathrm{<span\; class="notranslate">\; m</span>}\stackrel{<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; sx</span>}}$ $\mathrm{<span\; class="notranslate">\; m</span>}\stackrel{<span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span>}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; sy</span>}}$ In the formula, m is the mass of the rotor; F _sx , F _sy inherent Maxwell force, its expression is: F _sx =k _s x F _sy = k _s y In the formula r is the radius of the rotor; l is the length of the rotor shaft; μ ₀ is the air permeability; δ is the length of the air gap; k is the attenuation factor, generally 0.3. 6. The bearingless asynchronous motor RBF neural network adaptive inverse decoupling control and parameter identification method according to claim 1, is characterized in that, in the described step 4-5, RBF neural network RBFNNC (51) and RBF neural network The off-line training method of RBFNNI(52) structure and parameter determination is: Determine the number of hidden layer nodes and their center and width through offline training, and calculate the initial value of the connection weight between the hidden layer and the output layer. The input samples of RBFNNC(51) are displacement X, Y, rotation speed ω _r , The error between the given value and the actual value of the flux linkage ψ _r and the output value {e ₁ , e ₂ , e ₃ , e ₄ , e _c1 , e _c2 , e _c3 , e _c4 } of the error signal through the e _c function module, The output sample is {U _1α , U _1β , U _2α , U _2β } after coordinate transformation; the input sample of RBFNNI(52) is the input and output of the compound controlled object after delay {U _1α , U _1β , U _2α , U _2β , X, Y, ω _r , ψ _r }, the output sample is the identified compound controlled object output {X′, Y′, ω _r ′, ψ _r ′}; In the training process, the number of nodes is adaptively increased according to certain rules, and the hidden layer units that have too little effect on the output signal are deleted according to the rules, and the system nonlinear mapping is effectively realized with the least hidden layer units. 7. The bearingless asynchronous motor RBF neural network adaptive inverse decoupling control and parameter identification method according to claim 1, is characterized in that, in the described step 4-5, RBF neural network RBFNNC (51) and RBF neural network The RBFNNI(52) structure and parameters determine the online training method using the recursive least squares method to learn the rules.