Rotor suspension control method of bearingless flux switching motor based on neural network

Abstract

The present invention proposes a bearingless flux switching motor rotor suspension control method based on a neural network. The control method utilizes an inverter output voltage vector to realize closed-loop direct control of the rotor suspension force, and utilizes the nonlinear mapping characteristics of a BP neural network to construct a rotor suspension force observer based on a BP neural network to improve the observation accuracy of the suspension force. In the suspension control, a weight adjustment rule of a parameter identifier is designed based on a simple mathematical model of the bearingless motor and a rotor radial displacement control error, and the weight coefficient of a neural network parameter identifier of a rotor radial displacement closed-loop PID controller based on a BP neural network is dynamically adjusted. The present invention can solve the problems of difficulty in observing the motor suspension force and controlling the radial displacement caused by the severe nonlinearity and multiple unmodeled characteristics of a dual-winding BFSPMM.

Description

Technical Field

[0001] The invention relates to the technical field of motors, and in particular to a bearingless flux switching motor rotor suspension control method based on a neural network. Background Art

[0002] As a stator-excited permanent magnet motor, the flux switching motor has no permanent magnets on its rotor, which is conducive to high-speed rotation of the rotor and convenient heat dissipation of the permanent magnets. In addition, compared with other stator permanent magnet motors, the flux switching motor has the advantages of bipolar permanent magnet flux linkage and highly sinusoidal back electromotive force. Therefore, the bearingless nature of the flux switching motor has a high research value. The closed-loop control strategies for the suspension channel of the bearingless motor mainly include current closed-loop control and suspension force closed-loop direct control strategies. Among them, the current closed-loop control is easy to implement, has no special requirements for hardware and software, and uses feedback and coordinate transformation to generate current control signals, but the coupling between its power winding and suspension winding cannot be avoided. The suspension force closed-loop direct control can achieve a fast response of the suspension force without the need to control the current, but its suspension force is difficult to accurately estimate. As a mature control method, traditional PID control is simple and reliable, with a wide range of applications. It has good control effects when the control object model is determined and the parameters are unchanged. However, the bearingless flux switching motor (BFSPMM) is a multivariable, nonlinear, and strongly coupled system, and it is difficult for traditional PID control to achieve satisfactory control effects.

[0003] Therefore, the present invention proposes an improved rotor suspension direct control method based on BP neural network PID and BP neural network suspension force observer for a dual-winding bearingless flux switching motor. Summary of the invention

[0004] The present invention proposes a bearingless flux switching motor rotor suspension control method based on a neural network, which can solve the problems of difficulty in observing the motor suspension force and controlling the radial displacement caused by the severe nonlinearity and multiple unmodeled characteristics of the dual-winding BFSPMM.

[0005] The present invention adopts the following technical solutions.

[0006] A bearingless flux switching motor rotor suspension control method based on neural network is proposed. The control method uses the inverter output voltage vector to realize closed-loop direct control of the rotor suspension force, and uses the nonlinear mapping characteristics of the BP neural network to construct a rotor suspension force observer based on the BP neural network to improve the suspension force observation accuracy. In the suspension control, the weight adjustment rule of the parameter identifier is designed based on the simple mathematical model of the bearingless motor and the rotor radial displacement control error, and the weight coefficient of the neural network parameter identifier of the rotor radial displacement closed-loop PID controller based on the BP neural network is dynamically adjusted.

[0007] In the control method, the finite element simulation model is used to obtain the x-direction radial displacement of the motor rotor, the y-direction radial displacement, the x-axis and y-axis suspension currents of the suspension plane, the d-axis and q-axis currents of the torque plane, and the rotor tangential position angle θ. r as input vector and the training data of suspension force in x and y directions as output vector; constructing BP neural network suspension force observer with the obtained training data;

[0008] In the control method, a rotor radial displacement closed-loop PID controller neural network parameter identifier is constructed, which takes the x and y radial displacements of the motor rotor, the x and y radial displacement givens of the rotor, the x and y radial displacement errors of the rotor, and the unit 1 as input vectors, and takes the proportional coefficient, integral coefficient, and differential coefficient in the rotor radial displacement closed-loop PID controller as output vectors; when the motor rotor is running, the suspension force given for the rotor radial displacement control is calculated according to the identified PID parameters, and the weighting coefficient adjustment rule of the parameter identifier is designed according to the rotor radial position error, and the weighting coefficient is dynamically adjusted;

[0009] In the control method, the suspension plane flux control error is calculated according to the suspension force control error in the radial x and y directions of the rotor, and the flux control error is accurately compensated by using the inverter output voltage vector to achieve direct control of the suspension force.

[0010] In the motor, the stator is composed of 12 U-shaped iron cores and 12 alternating tangentially magnetized permanent magnets, and the rotor is a ten-tooth winding-free structure. Each permanent magnet and the U-shaped iron cores on both sides form a stator tooth, and 12 power winding coils are wound on the 12 stator teeth. Four power winding coils with a mechanical angle difference of 90° form one phase, with a total of three phases mA, mB, and mC, which are star-connected;

[0011] The motor uses small-sized permanent magnets, and 12 suspended winding coils are placed in its permanent magnet slots. Four coils with symmetrical spatial positions form one phase, with a total of three phases sa, sb, and sc connected in a star shape. The spatially symmetrical air gap magnetic field is reversely modulated through the suspended winding coils to generate a suspension force, thereby realizing the magnetic suspension of the rotor.

[0012] The rotor radial displacement closed-loop PID controller and suspension force observer based on the BP neural network optimize the suspension control of the rotor. The radial displacement difference of the rotor in two directions is obtained by the neural network PID controller to obtain the suspension force given F. x * 、F y *; Suspension force given and suspension force F estimated by BP neural network observer x 、F y The difference in suspension force is obtained by subtracting ΔF x , ΔF y , and then calculate the change value of the suspension winding flux Δψsx , Δψ sy , and then calculate the precise voltage vector u applied to the suspension winding through the voltage equation sx * 、u sy * Finally, the switch tube of the suspended winding inverter is controlled by SVPWM.

[0013] The suspension force observer uses a three-layer BP neural network to observe the suspension force of the rotor, with 7 input nodes, 20 hidden layer nodes, and 2 output nodes; the neural network input is the power winding dq axis current i md with i mq , suspension winding xy axis current i sx with i sy , motor rotor eccentric position x and y, rotor electrical angle θ r ; The output of the neural network is the suspension force F x 、F y ;

[0014] The structure of the rotor radial displacement closed-loop PID controller parameter identifier based on BP neural network is as follows: 4 input nodes, 5 hidden layer nodes, and 3 output nodes; the input is the control target, i.e. the actual value of radial displacement x(k), the given value x * (k), error Δx(k), and bias term 1 to increase network stability, the output is the x-direction PID controller parameter x-direction proportional coefficient k px , x-direction integral coefficient k ix , x-direction differential coefficient k dx ; The hidden layer uses the Tanh activation function; The output layer uses the sigmoid activation function; When the PID controller is working, the radial displacement is given by x * (k); the actual radial displacement x(k) and the actual radial displacement x(k-1) at the previous moment; the radial displacement error Δx(k) and the radial displacement error Δx(k-1) at the previous moment are input into the neural network PID error back propagation weight training algorithm to update the weights of the neural network input and output layers; then the forward propagation and reverse training process are repeated until the optimal PID parameters are obtained for the current state.

[0015] The suspension control method comprises the following steps in sequence:

[0016] Step S1, x radial displacement given x * , y radial displacement given y * Differences are made with the x radial displacement x and the y radial displacement y respectively, and the outputs are the x radial displacement control error Δx and the y radial displacement control error Δy:

[0017]

[0018] Step S2: Set the x radial displacement to x * , x radial displacement x, x radial displacement control error Δx, and bias term 1 are sent to the x radial displacement BP neural network PID controller parameter identifier, and the x direction proportional coefficient k is output. px , x-direction integral coefficient k ix , x-direction differential coefficient k dx ; Give the y radial displacement to y * , y radial displacement y, y radial displacement control error Δy, and bias term 1 are sent to the parameter identifier of the y radial displacement BP neural network PID controller, and the y direction proportional coefficient k is output. py , y-direction integral coefficient k iy , y-direction differential coefficient k dy ; The parameter identifiers of the PID controllers of the BP neural network for the radial displacement in the x direction and the radial displacement in the y direction are exactly the same.

[0019] In the x direction, at time k, the input of the jth node of the network input layer is O j (1) (k) is:

[0020]

[0021] The input of the i-th node in the network hidden layer is net i (2) (k), the i-th node outputs O i (2) (k) is:

[0022]

[0023] In the formula, w ij (2) is the weight coefficient between the jth input and the ith hidden layer neuron; the superscripts (1), (2), and (3) represent the input layer, hidden layer, and output layer, respectively; the hidden layer activation function f(x) is the Tanh function, as shown below:

[0024]

[0025] The input net of the lth node of the network output layer l (3) (k), the lth node outputs O l (3) (k) is:

[0026]

[0027] Among them, w li (3)is the weighted coefficient between the lth output neuron and the i-th hidden layer neuron; the output layer activation function g(x) uses the sigmoid activation function, as shown in the following formula.

[0028]

[0029] The network output is:

[0030]

[0031] Step S3: x radial displacement control error Δx, x direction proportional coefficient k px , x-direction integral coefficient k ix , x-direction differential coefficient k dx Send it to the x radial displacement PID controller, and output the x radial suspension force given F x * ; The y radial displacement control error Δy, y direction proportional coefficient k py , y-direction integral coefficient k iy , y-direction differential coefficient k dy Sent to the y radial displacement PID controller, outputting the y radial suspension force given F y * , the formula is:

[0032]

[0033] Among them, k and k-1 represent the kth beat and k-1th beat respectively;

[0034] Step S4: x radial suspension force given F x * , y radial suspension force given F y * Respectively with the x radial suspension force F x , y radial suspension force F y Make a difference and output the control error of x and y radial suspension force ΔF x , ΔF y :

[0035]

[0036] Step S5: Control error ΔF according to x and y radial suspension forces x , ΔF y Calculate the x and y radial suspension flux change Δψ sx , Δψ sy :

[0037]

[0038] Among them, L s and ψ fseThey are the self-inductance of the suspension winding and the flux linkage of the suspension winding coupled permanent magnet;

[0039] Step S6: According to the three-phase suspension winding current i sa 、i sb 、i sc and the x and y radial suspension flux change Δψ sx , Δψ sy Calculate the x and y radial suspension winding voltage given

[0040]

[0041]

[0042] Among them, R s 、T s They are the suspension winding resistance and digital control period;

[0043] Step S7: Given the voltage of the radial suspension windings x and y The space voltage vector modulation (SVPWM) method is used to apply the same voltage to the ends of the three-phase suspension windings through the three-phase inverter power switches. The corresponding voltage is used to achieve the closed-loop inner-loop control of the x and y radial suspension force and the closed-loop outer-loop control of the x and y radial displacement.

[0044] In step S2, the weight coefficient w between the jth input and the i-th hidden layer neuron is ij (2) , the weight coefficient w between the lth output neuron and the i-th hidden layer neuron li (3) The setting method is as follows:

[0045] Step A1: Adjust the weighted coefficient between the kth output neuron and the ith hidden layer neuron Calculated as:

[0046]

[0047] in, The calculation is as follows:

[0048]

[0049] The calculations for each part are as follows:

[0050]

[0051] The calculation is as follows

[0052]

[0053] Step A2: k-th output layer weighting coefficient is calculated as:

[0054]

[0055] Step A3: Adjust the weighted coefficient between the jth input of the kth beat and the ith hidden layer neuron is calculated as:

[0056]

[0057] in, The calculation is as follows:

[0058]

[0059] Step A4: k-th hidden layer weight coefficient is calculated as:

[0060]

[0061] In step S4, the x radial suspension force F x , y radial suspension force F y The method of obtaining is as follows:

[0062] Step B1: Calculate the dq axis current i of different power windings using finite element analysis software or analytical model md with i mq , suspension winding current i sx with i sy , rotor eccentricity position x and y, rotor electrical angle θ r The x and y radial suspension forces F acting on the rotor are x 、F y , get i md 、i mq 、i sx 、i sy , x, y, θ r For input, F x 、F y is the output array M[1…N];

[0063] Step B2: Use the array M[1…N] to train the BP neural network suspension force observer to obtain a stable BP neural network weight coefficient, thereby obtaining a BP neural network suspension force observer BPNN. The BPNN uses a three-layer BP neural network, with 7 input nodes, 20 hidden layer nodes, and 2 output nodes. The hidden layer activation function f(x) is a Tanh function, which is expressed as follows:

[0064]

[0065] Step B3: Motor three-phase stator power winding current i mA 、i mB 、i mC The αβ axis current component i of the stator power winding current is obtained by transforming it into the αβ stationary coordinate system through the 3s / 2s transformation link. mα 、i mβ :

[0066]

[0067] Step B4: αβ-axis current component i of the stator power winding current mα 、i mβ The dq axis current component i of the stator power winding current is obtained by transforming it into the dq rotating coordinate system through the 2s / 2r transformation link. md 、i mq :

[0068]

[0069] Step B5: The power winding dq axis current i md with i mq , suspension winding current i sx with i sy , rotor eccentricity position x and y, rotor electrical angle θ r Sent to the BPNN suspension force observer, outputting the x and y radial suspension forces F x 、F y .

[0070] The control method is implemented by a control system, which includes an AC power supply, a rectifier circuit, a filter circuit, a DC bus part, a voltage acquisition circuit, a dual three-phase inverter, a dual-winding bearingless motor, a rotary encoder, a rotor radial displacement acquisition circuit, a phase winding current acquisition circuit, a controller, an isolation drive, a human-computer interaction part, and a rotor position angle detection circuit;

[0071] The DC bus voltage of the dual three-phase inverter is provided by a matching DC power supply; the power tube in the inverter adopts an IGBT or MOSFET with a diode in parallel; the controller adopts a DSP or a single-chip microcomputer;

[0072] The phase winding current acquisition circuit is composed of a Hall current sensor combined with an operational amplifier, or a winding series power resistor followed by a differential operational amplifier.

[0073] The voltage acquisition circuit is a DC bus voltage acquisition circuit, which is composed of a Hall voltage sensor combined with an operational amplifier, or a parallel resistor voltage divider followed by a voltage follower composed of an operational amplifier;

[0074] The rotor position angle detection circuit is composed of a rotary encoder followed by a level conversion circuit, or a rotary transformer followed by a decoding circuit;

[0075] The rotor radial displacement acquisition circuit is a rotor radial xy offset acquisition circuit, which is composed of an eddy current sensor followed by an operational amplifier, or a linear optical coupler followed by an operational amplifier.

[0076] The weak current signal output by the current detection circuit, voltage sampling circuit and rotor radial displacement acquisition circuit in the control system is sent to the A / D conversion module of the controller, and the pulse signal output by the position angle detection circuit is sent to the QEP module of the controller;

[0077] When the control system is running, the radial displacement control errors Δx and Δy in two directions are obtained through the neural network PID to obtain the suspension force given F x * 、F y * The suspension force is given by F x * 、F y * The suspension force F estimated by BP neural network observer x 、F y The difference in suspension force is obtained by subtracting ΔF x , ΔF y , and then calculate the change value of the suspension winding flux Δψ sx , Δψ sy , and then calculate the precise voltage vector u applied to the suspension winding using Formula 12 sx * 、u sy * Finally, the switching tube of the suspended winding inverter is controlled by SVPWM.

[0078] The suspension force of the suspension winding is provided by the suspension winding flux. The current expression of the suspension winding current that forms the suspension winding flux is:

[0079] The magnetic flux expression of the suspension winding suspension force is:

[0080]

[0081] In the motor, the electrical time constant is much smaller than the mechanical time constant, which is equivalent to the radial displacement x and y of the motor not changing within one control cycle. Therefore, the suspension force F at the next moment is x * 、F y * and its magnetic flux ψ fsx * , fsy * The relationship is:

[0082]

[0083] Subtracting Formula 2 from Formula 3 gives the change in suspension force ΔF within a control cycle. x , ΔF y The change of the magnetic flux of the suspension winding is Δψ sx , Δψ sy The relationship is as follows:

[0084]

[0085] Rewrite Formula 4 as the suspension winding flux difference Δψ sx , Δψ sy The relationship between the difference in suspension force and the difference in suspension force is:

[0086]

[0087] The voltage vector u applied to the suspension winding is calculated by the difference in the suspension winding flux. sx * 、u sy * :

[0088]

[0089] Where R s is the suspension winding resistance; the voltage vector generates a switching signal with a constant switching frequency through the SVPWM module to accurately compensate for the suspension force error and realize direct control of the suspension force.

[0090] The method of the present invention can be used for a bearingless flux switching motor, and compared with the prior art, has the following advantages:

[0091] (1) The inverter output voltage vector is used to realize closed-loop direct control of the suspension force, which achieves a fast response of the suspension force and reduces the pulsation of the rotor radial displacement control;

[0092] (2) By utilizing the nonlinear mapping characteristics of the BP neural network, a rotor suspension force observer based on the BP neural network was constructed, which effectively overcame the adverse effects of the serious nonlinearity of the magnetic circuit of the dual-winding bearingless flux switching motor and the mutual coupling between the power winding and the suspension winding on the rotor suspension force observation, thereby improving the accuracy of the suspension force observation.

[0093] (3) Based on the simple mathematical model of the bearingless motor and the rotor radial displacement control error, the weight adjustment rules of the design parameter identifier are used to dynamically adjust the weight coefficient of the neural network parameter identifier of the rotor radial displacement closed-loop PID controller based on the BP neural network, which effectively adapts to the adverse effects of load dynamics, speed dynamics, etc. on the high stiffness and stable operation of the rotor suspension subsystem. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0095] Attached Figure 1 It is a topological diagram of a bearingless flux switching motor;

[0096] Attached Figure 2 It is a schematic diagram of the control principle of the present invention;

[0097] Attached Figure 3 It is a schematic diagram of the suspension force observer based on BP neural network;

[0098] Attached Figure 4 It is a schematic diagram of the structure of the parameter identifier of PID controller based on BP neural network;

[0099] Attached Figure 5 It is a schematic diagram of the hardware principle of the control system on which the control method is based;

[0100] Attached Figure 6 It is a schematic diagram of the coordinate system definition of the rotor's suspension plane;

[0101] Attached Figure 7 It is a schematic diagram of the change of magnetic flux of the suspension winding during one control cycle. DETAILED DESCRIPTION

[0102] As shown in the figure, the present invention proposes an improved dual-winding BFSPMM rotor suspension direct control method based on BP neural network PID and BP neural network suspension force observer. The finite element simulation model is used to obtain the rotor x radial displacement and y radial displacement, suspension plane x-axis and y-axis suspension current, torque plane d-axis and q-axis current, rotor tangential position angle θ rThe training data is taken as the input vector and the suspension force in the x and y directions as the output vector; the BP neural network suspension force observer is constructed using the obtained training data. A neural network parameter identifier for a rotor radial displacement closed-loop PID controller is constructed with the rotor x and y radial displacement, the rotor x and y radial displacement given, the rotor x and y radial displacement error, and unit 1 as input vectors, and the proportional coefficient, integral coefficient, and differential coefficient in the rotor radial displacement closed-loop PID controller as output vectors. The suspension force given for rotor radial displacement control is calculated based on the identified PID parameters, and the weight adjustment rule of the parameter identifier is designed based on the rotor radial position error to dynamically adjust the weight. According to the suspension force control error in the radial x and y directions of the rotor, the suspension plane flux control error is calculated, and the inverter output voltage vector is used to accurately compensate for the flux control error to achieve direct control of the suspension force.

[0103] In this example, a bearingless flux switching motor rotor suspension control method based on a neural network is proposed. The control method utilizes the inverter output voltage vector to realize closed-loop direct control of the rotor suspension force, and utilizes the nonlinear mapping characteristics of the BP neural network to construct a rotor suspension force observer based on the BP neural network to improve the suspension force observation accuracy. In the suspension control, the weight adjustment rule of the parameter identifier is designed based on the simple mathematical model of the bearingless motor and the rotor radial displacement control error, and the weight coefficient of the neural network parameter identifier of the rotor radial displacement closed-loop PID controller based on the BP neural network is dynamically adjusted.

[0104] In the control method, the finite element simulation model is used to obtain the x-direction radial displacement of the motor rotor, the y-direction radial displacement, the x-axis and y-axis suspension currents of the suspension plane, the d-axis and q-axis currents of the torque plane, and the rotor tangential position angle θ. r as input vector and the training data of suspension force in x and y directions as output vector; constructing BP neural network suspension force observer with the obtained training data;

[0105] In the control method, a rotor radial displacement closed-loop PID controller neural network parameter identifier is constructed, which takes the x and y radial displacements of the motor rotor, the x and y radial displacement givens of the rotor, the x and y radial displacement errors of the rotor, and the unit 1 as input vectors, and takes the proportional coefficient, integral coefficient, and differential coefficient in the rotor radial displacement closed-loop PID controller as output vectors; when the motor rotor is running, the suspension force given for the rotor radial displacement control is calculated according to the identified PID parameters, and the weighting coefficient adjustment rule of the parameter identifier is designed according to the rotor radial position error, and the weighting coefficient is dynamically adjusted;

[0106] In the control method, the suspension plane flux control error is calculated according to the suspension force control error in the radial x and y directions of the rotor, and the flux control error is accurately compensated by using the inverter output voltage vector to achieve direct control of the suspension force.

[0107] In the motor, the stator is composed of 12 U-shaped iron cores and 12 alternating tangentially magnetized permanent magnets, and the rotor is a ten-tooth winding-free structure. Each permanent magnet and the U-shaped iron cores on both sides form a stator tooth, and 12 power winding coils are wound on the 12 stator teeth. Four power winding coils with a mechanical angle difference of 90° form one phase, with a total of three phases mA, mB, and mC, which are star-connected;

[0108] The motor uses small-sized permanent magnets, and 12 suspended winding coils are placed in its permanent magnet slots. Four coils with symmetrical spatial positions form one phase, with a total of three phases sa, sb, and sc connected in a star shape. The spatially symmetrical air gap magnetic field is reversely modulated through the suspended winding coils to generate a suspension force, thereby realizing the magnetic suspension of the rotor.

[0109] The rotor radial displacement closed-loop PID controller and suspension force observer based on the BP neural network optimize the suspension control of the rotor. The radial displacement difference of the rotor in two directions is obtained by the neural network PID controller to obtain the suspension force given F. x * 、F y *; Suspension force given and suspension force F estimated by BP neural network observer x 、F y The difference in suspension force is obtained by subtracting ΔF x , ΔF y , and then calculate the change value of the suspension winding flux Δψ sx , Δψ sy , and then calculate the precise voltage vector u applied to the suspension winding through the voltage equation sx * 、u sy * Finally, the switch tube of the suspended winding inverter is controlled by SVPWM.

[0110] The suspension force observer uses a three-layer BP neural network to observe the suspension force of the rotor, with 7 input nodes, 20 hidden layer nodes, and 2 output nodes; the neural network input is the power winding dq axis current i md with i mq , suspension winding xy axis current i sx with i sy , motor rotor eccentric position x and y, rotor electrical angle θ r ; The output of the neural network is the suspension force F x 、F y ;

[0111] The structure of the rotor radial displacement closed-loop PID controller parameter identifier based on BP neural network is as follows: 4 input nodes, 5 hidden layer nodes, and 3 output nodes; the input is the control target, i.e. the actual value of radial displacement x(k), the given value x * (k), error Δx(k), and bias term 1 to increase network stability, the output is the x-direction PID controller parameter x-direction proportional coefficient k px , x-direction integral coefficient k ix , x-direction differential coefficient k dx ; The hidden layer uses the Tanh activation function; The output layer uses the sigmoid activation function; When the PID controller is working, the radial displacement is given by x * (k); the actual radial displacement x(k) and the actual radial displacement x(k-1) at the previous moment; the radial displacement error Δx(k) and the radial displacement error Δx(k-1) at the previous moment are input into the neural network PID error back propagation weight training algorithm to update the weights of the neural network input and output layers; then the forward propagation and reverse training process are repeated until the optimal PID parameters are obtained for the current state.

[0112] The suspension control method comprises the following steps in sequence:

[0113] Step S1, x radial displacement given x * , y radial displacement given y * Differences are made with the x radial displacement x and the y radial displacement y respectively, and the outputs are the x radial displacement control error Δx and the y radial displacement control error Δy:

[0114]

[0115] Step S2: Set the x radial displacement to x * , x radial displacement x, x radial displacement control error Δx, and bias term 1 are sent to the x radial displacement BP neural network PID controller parameter identifier, and the x direction proportional coefficient k is output. px , x-direction integral coefficient k ix , x-direction differential coefficient k dx ; Give the y radial displacement to y * , y radial displacement y, y radial displacement control error Δy, and bias term 1 are sent to the parameter identifier of the y radial displacement BP neural network PID controller, and the y direction proportional coefficient k is output. py , y-direction integral coefficient k iy , y-direction differential coefficient k dy ; The parameter identifiers of the PID controllers of the BP neural network for the radial displacement in the x direction and the radial displacement in the y direction are exactly the same.

[0116] In the x direction, at time k, the input of the jth node of the network input layer is O j(1) (k) is:

[0117]

[0118] The input of the i-th node in the network hidden layer is net i (2) (k), the i-th node outputs O i (2) (k) is:

[0119]

[0120] In the formula, w ij (2) is the weight coefficient between the jth input and the ith hidden layer neuron; the superscripts (1), (2), and (3) represent the input layer, hidden layer, and output layer, respectively; the hidden layer activation function f(x) is the Tanh function, as shown below:

[0121]

[0122] The input net of the lth node of the network output layer l (3) (k), the lth node outputs O l (3) (k) is:

[0123]

[0124] Among them, w li (3) is the weighted coefficient between the lth output neuron and the i-th hidden layer neuron; the output layer activation function g(x) uses the sigmoid activation function, as shown in the following formula.

[0125]

[0126] The network output is:

[0127]

[0128] Step S3: x radial displacement control error Δx, x direction proportional coefficient k px , x-direction integral coefficient k ix , x-direction differential coefficient k dx Send it to the x radial displacement PID controller, and output the x radial suspension force given F x * ; The y radial displacement control error Δy, y direction proportional coefficient k py , y-direction integral coefficient k iy , y-direction differential coefficient k dySent to the y radial displacement PID controller, outputting the y radial suspension force given F y * , the formula is:

[0129]

[0130] Among them, k and k-1 represent the kth beat and k-1th beat respectively;

[0131] Step S4: x radial suspension force given F x * , y radial suspension force given F y * Respectively with the x radial suspension force F x , y radial suspension force F y Make a difference and output the control error of x and y radial suspension force ΔF x , ΔF y :

[0132]

[0133] Step S5: Control error ΔF according to x and y radial suspension forces x , ΔF y Calculate the x and y radial suspension flux change Δψ sx , Δψ sy :

[0134]

[0135] Among them, L s and ψ fse They are the self-inductance of the suspension winding and the flux linkage of the suspension winding coupled permanent magnet;

[0136] Step S6: According to the three-phase suspension winding current i sa 、i sb 、i sc and the x and y radial suspension flux change Δψ sx , Δψ sy Calculate the x and y radial suspension winding voltage given

[0137]

[0138]

[0139] Among them, R s 、T s They are the suspension winding resistance and digital control period;

[0140] Step S7: Given the voltage of the radial suspension windings x and y The space voltage vector modulation (SVPWM) method is used to apply the same voltage to the ends of the three-phase suspension windings through the three-phase inverter power switches. The corresponding voltage is used to achieve the closed-loop inner-loop control of the x and y radial suspension force and the closed-loop outer-loop control of the x and y radial displacement.

[0141] In step S2, the weight coefficient w between the jth input and the i-th hidden layer neuron is ij (2) , the weight coefficient w between the lth output neuron and the i-th hidden layer neuron li (3) The method of obtaining is as follows:

[0142] Step A1: Adjust the weighted coefficient between the kth output neuron and the ith hidden layer neuron is calculated as:

[0143]

[0144] in, The calculation is as follows:

[0145]

[0146] The calculations for each part are as follows:

[0147]

[0148] The calculation is as follows

[0149]

[0150] Step A2: k-th output layer weighting coefficient is calculated as:

[0151]

[0152] Step A3: Adjust the weighted coefficient between the jth input of the kth beat and the ith hidden layer neuron is calculated as:

[0153]

[0154] in, The calculation is as follows:

[0155]

[0156] Step A4: k-th hidden layer weight coefficient is calculated as:

[0157]

[0158] In step S4, the x radial suspension force F x , y radial suspension force F y The method of obtaining is as follows:

[0159] Step B1: Calculate the dq axis current i of different power windings using finite element analysis software or analytical model md with i mq , suspension winding current i sx with i sy , rotor eccentricity position x and y, rotor electrical angle θ r Input, the x and y radial suspension forces F acting on the rotor x 、F y , get i md 、i mq 、i sx 、i sy , x, y, θ r For input, F x 、F y is the output array M[1…N];

[0160] Step B2: Use the array M[1…N] to train the BP neural network suspension force observer to obtain a stable BP neural network weight coefficient, thereby obtaining a BP neural network suspension force observer BPNN. The BPNN uses a three-layer BP neural network, with 7 input nodes, 20 hidden layer nodes, and 2 output nodes. The hidden layer activation function f(x) is a Tanh function, which is expressed as follows:

[0161]

[0162] Step B3: Motor three-phase stator power winding current i mA 、i mB 、i mC The αβ axis current component i of the stator power winding current is obtained by transforming it into the αβ stationary coordinate system through the 3s / 2s transformation link. mα 、i mβ :

[0163]

[0164] Step B4: αβ-axis current component i of the stator power winding current mα 、i mβ The dq axis current component i of the stator power winding current is obtained by transforming it into the dq rotating coordinate system through the 2s / 2r transformation link. md 、i mq :

[0165]

[0166] Step B5: The power winding dq axis current i md with i mq , suspension winding current i sx with i sy , rotor eccentricity position x and y, rotor electrical angle θ r Sent to the BPNN suspension force observer, outputting the x and y radial suspension forces F x 、F y .

[0167] The control method is implemented by a control system, which includes an AC power supply, a rectifier circuit, a filter circuit, a DC bus part, a voltage acquisition circuit, a dual three-phase inverter, a dual-winding bearingless motor, a rotary encoder, a rotor radial displacement acquisition circuit, a phase winding current acquisition circuit, a controller, an isolation drive, a human-computer interaction part, and a rotor position angle detection circuit;

[0168] The DC bus voltage of the dual three-phase inverter is provided by a matching DC power supply; the power tube in the inverter adopts an IGBT or MOSFET with a diode in parallel; the controller adopts a DSP or a single-chip microcomputer;

[0169] The phase winding current acquisition circuit is composed of a Hall current sensor combined with an operational amplifier, or a winding series power resistor followed by a differential operational amplifier.

[0170] The voltage acquisition circuit is a DC bus voltage acquisition circuit, which is composed of a Hall voltage sensor combined with an operational amplifier, or a parallel resistor voltage divider followed by a voltage follower composed of an operational amplifier;

[0171] The rotor position angle detection circuit is composed of a rotary encoder followed by a level conversion circuit, or a rotary transformer followed by a decoding circuit;

[0172] The rotor radial displacement acquisition circuit is a rotor radial xy offset acquisition circuit, which is composed of an eddy current sensor followed by an operational amplifier, or a linear optical coupler followed by an operational amplifier.

[0173] The weak current signal output by the current detection circuit, voltage sampling circuit and rotor radial displacement acquisition circuit in the control system is sent to the A / D conversion module of the controller, and the pulse signal output by the position angle detection circuit is sent to the QEP module of the controller;

[0174] When the control system is running, the radial displacement control errors Δx and Δy in two directions are obtained through the neural network PID to obtain the suspension force given F x *、F y * The suspension force is given by F x * 、F y * The suspension force F estimated by BP neural network observer x 、F y The difference in suspension force is obtained by subtracting ΔF x , ΔF y , and then calculate the change value of the suspension winding flux Δψ sx , Δψ sy , and then calculate the precise voltage vector u applied to the suspension winding using Formula 12 sx * 、u sy * Finally, the switching tube of the suspended winding inverter is controlled by SVPWM.

[0175] Example:

[0176] In this case, the rotor's suspension plane is Figure 6 As shown, the sasbsc axis is a stationary three-phase coordinate system, the sαsβ axis is a stationary two-phase coordinate system, and the sxsy axis is a stationary two-phase coordinate system obtained by rotating the sαsβ axis clockwise by 30°;

[0177] The suspension force of the suspension winding is provided by the suspension winding flux. The current expression of the suspension winding current that forms the suspension winding flux is:

[0178] In this example, the control concept of direct torque is extended to suspension force control. The relationship between suspension winding flux and current and the relationship between suspension force and current are used to obtain the relationship between suspension winding flux and suspension force. Figure 7 shown.

[0179] The magnetic flux expression of the suspension winding suspension force is:

[0180]

[0181] In the motor, the electrical time constant is much smaller than the mechanical time constant, which is equivalent to the radial displacement x and y of the motor not changing within one control cycle. Therefore, the suspension force F at the next moment is x * 、F y * and its magnetic flux ψ sx * , sy * The relationship is:

[0182]

[0183] Subtracting Formula 2 from Formula 3 gives the change in suspension force ΔF within a control cycle. x , ΔF y The change of the magnetic flux of the suspension winding is Δψ sx , Δψ sy The relationship is as follows:

[0184]

[0185] Rewrite Formula 4 as the suspension winding flux difference Δψ sx , Δψ sy The relationship between the difference in suspension force and the difference in suspension force is:

[0186]

[0187] The voltage vector u applied to the suspension winding is calculated by the difference in the suspension winding flux. sx * 、u sy * :

[0188]

[0189] Where R s is the suspension winding resistance; the voltage vector generates a switching signal with a constant switching frequency through the SVPWM module to accurately compensate for the suspension force error and realize direct control of the suspension force.

[0190] In this example, the specific method of the suspension force observer based on the neural network is as follows:

[0191] According to the analysis results of formula (5) and (6), direct suspension force control does not require observation of suspension winding flux, so the accuracy of suspension force observation is more important for direct suspension force control. The mathematical model of suspension force established by virtual displacement method does not consider the influence of power winding current, rotor eccentric position, rotor electrical angle, etc. However, considering the above factors, it is impossible to establish an accurate mathematical model that can be directly applied to control. Therefore, the nonlinear mapping ability of BP neural network is used to observe the suspension force. Figure 3 As shown, the present invention typically adopts a three-layer BP neural network to observe the suspension force, wherein there are 7 input nodes, 20 hidden layer nodes, and 2 output nodes. The network input is the power winding current i md with i mq , suspension winding current i sx with i sy , motor rotor eccentric position x and y, rotor electrical angle θ r The network output is the suspension force F x 、F y .

[0192] The hidden layer activation function is Tanh, as shown in formula (7).

[0193]

[0194] The typical implementation method is: by inputting random excitation signals into the finite element simulation model, 50,000 sets of sample data are obtained. 40,000 sets of data are used for training and 10,000 sets of data are used for testing. The BP neural network is trained using the Bayesian regularization training algorithm. Finally, a BP neural network suspension force observer that can be used for control is obtained.

[0195] In this example, the specific method for radial displacement neural network PID controller parameter identification is as follows:

[0196] Taking the radial displacement control in the x direction as an example, the present invention adopts a position PID controller:

[0197]

[0198] Among them, F x * (k) is the controller output at the current moment, Δx(k) is the error at the current moment, Δx(k-1) is the error at the previous moment, and k px , k ix , k dx is the pid parameter. px , k ix , k dx When it is regarded as a parameter that can be updated in real time according to the system operation status, formula (8) can be described as:

[0199]

[0200] Among them, f[·] is related to k px , k ix , k dx , Δx(k), Δx(k-1) and other related nonlinear functions, a neural network can be used to establish an optimal PID control.

[0201] The bearingless motor suspension control system is nonlinear, so using only fixed PID parameters cannot achieve the desired control effect. The BP neural network has the ability to express any nonlinearity, and can adjust the PID parameters in real time through online learning to achieve better control effects.

[0202] The structure of the parameter identifier of PID controller based on BP neural network is as follows Figure 4 As shown in the figure, there are 4 input nodes, 5 hidden layer nodes, and 3 output nodes. The input is the actual value of the control target x(k), the given x *(k), error Δx(k), and bias term 1 to increase network stability, the output is pid parameter k px , k ix , k dx The hidden layer uses the Tanh activation function. The output layer uses the sigmoid activation function, as shown in formula (10).

[0203]

[0204] When the neural network is forward propagating,

[0205] The i-th node input of the network input layer at time k is O j (1) (k) is:

[0206]

[0207] The input of the i-th node in the network hidden layer is net i (2) (k), the i-th node outputs O i (2) (k) is:

[0208]

[0209] In the formula, w ij (2) is the weight coefficient between the j-th input and the i-th hidden layer neuron. The superscripts (1), (2), and (3) represent the input layer, hidden layer, and output layer, respectively.

[0210] The input net of the lth node of the network output layer l (3) (k), the lth node outputs O l (3) (k) is:

[0211]

[0212] Among them, w li (3) is the weight coefficient between the lth output neuron and the ith hidden layer neuron.

[0213] The network output is:

[0214]

[0215] The neural network dynamically adjusts the weights according to the position error. By designing reasonable adjustment rules, the neural network can calculate a set of weights to adapt to the displacement stiffness and current stiffness of the rotor, and achieve accurate tracking of the rotor position.

[0216] In this example, the neural network controller uses an online learning algorithm based on the delta adaptive law, specifically:

[0217] In this example,

[0218] The neural network controller is a multi-layer network, and the BFSPMM corresponds to the last layer of the network. Since it is impossible to obtain an accurate BFSPMM model, the structure of the last layer of the network is unknown, but the weights can be updated using the gradient descent method. This method is reversed based on the BP learning rule and implemented using the chain rule of partial derivatives. The adjustment amount in the algorithm is defined as the derivative of the objective function with respect to the network parameters, which must be related to the position error. The objective function E is selected as

[0219]

[0220] Back propagation starts from the last layer of the network, and the output layer weights are adjusted to:

[0221]

[0222] in:

[0223]

[0224] Since there is no accurate BFSPMM system model, it is difficult to obtain According to the mathematical model of the bearingless motor suspension channel, the value of σ(k) is approximated as:

[0225]

[0226] In formula (16) Through formula (8), we get: Through formula (10), we get: It is obtained by formula (13). See the following formula for details:

[0227]

[0228] Substituting equations (18) and (19) into equation (16), we can obtain the output layer weight adjustment as:

[0229]

[0230] According to the above calculation method, the hidden layer weight adjustment amount can be obtained as:

[0231]

Claims

1. A bearingless flux switching motor rotor suspension control method based on a neural network, characterized in that: The control method uses the inverter output voltage vector to realize the closed-loop direct control of the rotor suspension force, and uses the nonlinear mapping characteristics of the BP neural network to construct a rotor suspension force observer based on the BP neural network to improve the observation accuracy of the suspension force. In the suspension control, the weight adjustment rule of the parameter identifier is designed based on the simple mathematical model of the bearingless motor and the rotor radial displacement control error, and the weight coefficient of the neural network parameter identifier of the rotor radial displacement closed-loop PID controller based on the BP neural network is dynamically adjusted; In the control method, the finite element simulation model is used to obtain the x-direction radial displacement of the motor rotor, the y-direction radial displacement, the x-axis and y-axis suspension currents of the suspension plane, the d-axis and q-axis currents of the torque plane, and the rotor tangential position angle θ. r as input vector and the training data of suspension force in x and y directions as output vector; constructing BP neural network suspension force observer with the obtained training data; In the control method, a rotor radial displacement closed-loop PID controller neural network parameter identifier is constructed with the x and y radial displacements of the motor rotor, the given x and y radial displacements of the rotor, the x and y radial displacement errors of the rotor, and the units as input vectors, and with the proportional coefficient, integral coefficient, and differential coefficient in the rotor radial displacement closed-loop PID controller as output vectors; when the motor rotor is running, the given suspension force for the rotor radial displacement control is calculated according to the identified PID parameters, and the weighting coefficient adjustment rule of the parameter identifier is designed according to the rotor radial position error, and the weighting coefficient is dynamically adjusted; In the control method, the suspension plane flux control error is calculated according to the suspension force control error in the radial x and y directions of the rotor, and the flux control error is accurately compensated by using the inverter output voltage vector to achieve direct control of the suspension force; The control method is based on BP neural network; in the motor, the stator is composed of 12 U-shaped iron cores and 12 alternating tangentially magnetized permanent magnets, and the rotor is a ten-tooth winding-free structure; each permanent magnet and the U-shaped iron cores on both sides form a stator tooth, 12 power winding coils are wound on the 12 stator teeth, and four power winding coils with a mechanical angle difference of 90° form one phase, with a total of three phases mA, mA, and mC and a star connection; The motor uses small-sized permanent magnets, and 12 suspension winding coils are placed in the permanent magnet slots. Four coils with symmetrical spatial positions form one phase, with a total of three phases sa, sb, and sc connected in a star shape. The suspension winding coils reversely modulate the spatially symmetrical air gap magnetic field to generate suspension force, thereby realizing magnetic suspension of the rotor. The rotor radial displacement closed-loop PID controller and suspension force observer based on the BP neural network optimize the suspension control of the rotor. The radial displacement difference of the rotor in two directions is obtained by the neural network PID controller to obtain the suspension force given F. x * 、F y *; Suspension force given and suspension force F estimated by BP neural network observer x 、F y The difference in suspension force is obtained by subtracting ΔF x , ΔF y , and then calculate the change value of the suspension winding flux Δψ sx , Δψ sy , and then calculate the precise voltage vector u applied to the suspension winding through the voltage equation sx * 、u sy * , and finally the switch tube of the suspension winding inverter is controlled by SVPWM; The suspension force observer uses a three-layer BP neural network to observe the suspension force of the rotor, including 7 input nodes, 20 hidden layer nodes, and 2 output nodes. The neural network input is the power winding dq axis current i md with i mq , suspension winding xy axis current i sx with i sy , motor rotor eccentric position x and y, rotor electrical angle θ r ; The output of the neural network is the suspension force F x 、F y ; The structure of the rotor radial displacement closed-loop PID controller parameter identifier based on BP neural network is as follows: 4 input nodes, 5 hidden layer nodes, and 3 output nodes; the input is the control target, i.e. the actual value of radial displacement x(k), the given value x * (k), error Δx(k), and bias term 1 to increase network stability, the output is the x-direction PID controller parameter x-direction proportional coefficient k px , x-direction integral coefficient k ix , x-direction differential coefficient k dx ; The hidden layer uses the Tanh activation function; The output layer uses the sigmoid activation function; When the PID controller is working, the radial displacement is given by x * (k); actual radial displacement x(k) and actual radial displacement x(k-1) at the previous moment; radial displacement error Δx(k) and radial displacement error Δx(k-1) at the previous moment are input into the neural network PID error back propagation weight training algorithm to update the weights of the neural network input and output layers; then the forward propagation and reverse training process are repeated until the optimal PID parameters are obtained for the current state; The suspension control method comprises the following steps in sequence: Step S1, x radial displacement given x * , y radial displacement given y * Differences are made with the x radial displacement x and the y radial displacement y respectively, and the outputs are the x radial displacement control error Δx and the y radial displacement control error Δy: Step S2: Set the x radial displacement to x * , x radial displacement x, x radial displacement control error Δx, and bias term 1 are sent to the x radial displacement BP neural network PID controller parameter identifier, and the x direction proportional coefficient k is output px , x-direction integral coefficient k ix , x-direction differential coefficient k dx ; Give the y radial displacement to y * , y radial displacement y, y radial displacement control error Δy, and bias term 1 are sent to the parameter identifier of the y radial displacement BP neural network PID controller, and the y direction proportional coefficient k is output. py , y-direction integral coefficient k iy , differential coefficient k in the y direction dy ; The parameter identifiers of the PID controllers of the BP neural network for the radial displacement in the x direction and the radial displacement in the y direction are exactly the same. In the x direction, at time k, the input of the jth node of the network input layer is O j (1) (k) is: The input of the i-th node in the network hidden layer is net i (2) (k), the i-th node outputs O i (2) (k) is: In the formula, w ij (2) is the weight coefficient between the jth input and the ith hidden layer neuron; the superscripts (1), (2), and (3) represent the input layer, hidden layer, and output layer, respectively; the hidden layer activation function f(x) is the Tanh function, as shown below: The input net of the lth node of the network output layer l (3) (k), the lth node outputs O l (3) (k) is: Among them, w li (3) is the weight coefficient between the lth output neuron and the ith hidden layer neuron; the output layer activation function g(x) uses the sigmoid activation function, as shown in the following formula; The network output is: Step S3: x radial displacement control error Δx, x direction proportional coefficient k px , x-direction integral coefficient k ix , x-direction differential coefficient k dx Send it to the x radial displacement PID controller, and output the x radial suspension force given F x * ; The y radial displacement control error Δy, y direction proportional coefficient k py , y-direction integral coefficient k iy , differential coefficient k in the y direction dy Sent to the y radial displacement PID controller, outputting the y radial suspension force given F y * , the formula is: Among them, k and k-1 represent the kth beat and k-1th beat respectively; Step S4: x radial suspension force given F x * , y radial suspension force given F y * Respectively with the x radial suspension force F x , y radial suspension force F y Make a difference and output the control error of x and y radial suspension force ΔF x , ΔF y : Step S5: Control error ΔF according to x and y radial suspension forces x , ΔF y Calculate the x and y radial suspension flux change Δψ sx , Δψ sy : Among them, L s and ψ fse They are the self-inductance of the suspension winding and the flux linkage of the suspension winding coupled permanent magnet; Step S6: According to the three-phase suspension winding current i sa 、i sb 、i sc and the x and y radial suspension flux change Δψ sx , Δψ sy Calculate the x and y radial suspension winding voltage given Among them, R s , T s They are the suspension winding resistance and digital control period; Step S7: Given the voltage of the radial suspension windings x and y The space voltage vector modulation (SVPWM) method is used to apply the same voltage to the ends of the three-phase suspension windings through the three-phase inverter power switches. The corresponding voltage is used to achieve the closed-loop inner-loop control of the x and y radial suspension force and the closed-loop outer-loop control of the x and y radial displacement.

2. The method for controlling the rotor suspension of a bearingless flux switching motor based on a neural network according to claim 1, characterized in that: In step S2, the weight coefficient w between the jth input and the i-th hidden layer neuron is ij (2) , the weight coefficient w between the lth output neuron and the i-th hidden layer neuron li (3) The setting method is as follows: Step A1: Adjust the weighted coefficient between the kth output neuron and the ith hidden layer neuron Calculated as: in, The calculation is as follows: The calculations for each part are as follows: The calculation is as follows Step A2: k-th output layer weighting coefficient Calculated as: Step A3: Adjust the weighted coefficient between the jth input of the kth beat and the ith hidden layer neuron Calculated as: in, The calculation is as follows: Step A4: k-th hidden layer weighting coefficient Calculated as:

3. The method for controlling the rotor suspension of a bearingless flux switching motor based on a neural network according to claim 2 is characterized in that: In step S4, the x radial suspension force F x , y radial suspension force F y The method of obtaining is as follows: Step B1: Calculate the dq axis current i of different power windings using finite element analysis software or analytical model md with i mq , suspension winding current i sx with i sy , rotor eccentricity position x and y, rotor electrical angle θ r The x and y radial suspension forces F acting on the rotor are x 、F y , get i md 、i mq 、i sx 、i sy , x, y, θ r For input, F x 、F y is the output array M[1…N]; Step B2, using the array M[1…N] to train the BP neural network suspension force observer to obtain a stable BP neural network weight coefficient, thereby obtaining a BP neural network suspension force observer BPNN, the BPNN adopts a three-layer BP neural network, wherein there are 7 input nodes, 20 hidden layer nodes, and 2 output nodes; The hidden layer activation function f(x) is the Tanh function, which is expressed as follows Step B3: Motor three-phase stator power winding current i mA 、i mB 、i mC The αβ axis current component i of the stator power winding current is obtained by transforming it into the αβ stationary coordinate system through the 3s / 2s transformation link. mα 、i mβ : Step B4: αβ-axis current component i of the stator power winding current mα 、i mβ The dq axis current component i of the stator power winding current is obtained by transforming it into the dq rotating coordinate system through the 2s / 2r transformation link. md 、i mq : Step B5: The power winding dq axis current i md with i mq , suspension winding current i sx with i sy , rotor eccentricity position x and y, rotor electrical angle θ r Sent to the BPNN suspension force observer, outputting the x and y radial suspension forces F x 、F y .

4. The method for controlling rotor suspension of a bearingless flux switching motor based on a neural network according to claim 3 is characterized in that: The control method is implemented by a control system, which includes an AC power supply, a rectifier circuit, a filter circuit, a DC bus part, a voltage acquisition circuit, a dual three-phase inverter, a dual-winding bearingless motor, a rotary encoder, a rotor radial displacement acquisition circuit, a phase winding current acquisition circuit, a controller, an isolation drive, a human-computer interaction part, and a rotor position angle detection circuit; The DC bus voltage of the dual three-phase inverter is provided by a matching DC power supply; the power tube in the inverter adopts an IGBT or MOSFET with a diode in parallel; the controller adopts a DSP or a single-chip microcomputer; The phase winding current acquisition circuit is composed of a Hall current sensor combined with an operational amplifier, or a winding series power resistor followed by a differential operational amplifier. The voltage acquisition circuit is a DC bus voltage acquisition circuit, which is composed of a Hall voltage sensor combined with an operational amplifier, or a parallel resistor voltage divider followed by a voltage follower composed of an operational amplifier; The rotor position angle detection circuit is composed of a rotary encoder followed by a level conversion circuit, or a rotary transformer followed by a decoding circuit; The rotor radial displacement acquisition circuit is a rotor radial xy offset acquisition circuit, which is composed of an eddy current sensor followed by an operational amplifier, or a linear optical coupler followed by an operational amplifier. The weak current signal output by the current detection circuit, voltage sampling circuit and rotor radial displacement acquisition circuit in the control system is sent to the A / D conversion module of the controller, and the pulse signal output by the position angle detection circuit is sent to the QEP module of the controller; When the control system is running, the radial displacement control errors Δx and Δy in two directions are obtained through the neural network PID to obtain the suspension force given F x * 、F y * ; Suspension force given F x * 、F y * The suspension force F estimated by BP neural network observer x 、F y The difference in suspension force is obtained by subtracting ΔF x , ΔF y , and then calculate the change value of the suspension winding flux Δψ sx , Δψ sy , and then calculate the precise voltage vector u applied to the suspension winding using Formula 12 sx * 、u sy * Finally, the switching tube of the suspended winding inverter is controlled by SVPWM.

5. The method for controlling the rotor suspension of a bearingless flux switching motor based on a neural network according to claim 4 is characterized in that: The suspension force of the suspension winding is provided by the suspension winding flux. The current expression of the suspension winding current that forms the suspension winding flux is: The magnetic flux expression of the suspension winding suspension force is: In the motor, the electrical time constant is much smaller than the mechanical time constant, which is equivalent to the radial displacement x and y of the motor not changing within one control cycle. Therefore, the suspension force F at the next moment is x * 、F y * and its magnetic flux ψ sx * , sy * The relationship is: Subtracting Formula 2 from Formula 3 gives the change in suspension force ΔF within a control cycle. x , ΔF y The change of the magnetic flux of the suspension winding is Δψ sx , Δψ sy The relationship is as follows: Rewrite Formula 4 as the suspension winding flux difference Δψ sx , Δψ sy The relationship between the difference in suspension force and the difference in suspension force is: The voltage vector u applied to the suspension winding is calculated by the difference in the suspension winding flux. sx * 、u sy * : Where R s is the suspension winding resistance; the voltage vector generates a switching signal with a constant switching frequency through the SVPWM module to accurately compensate for the suspension force error and realize direct control of the suspension force.

Images