Bearingless flux switching motor neural network PID suspension method with adaptive learning rate

Abstract

The invention provides a learning rate adaptive bearingless magnetic flux switching motor neural network PID suspension method, which is used for PID magnetic suspension control of a motor PID controller on a rotor, and adjusts a neural network weight coefficient in real time according to a rotor radial displacement control error so as to realize real-time adjustment of parameters of the neural network PID controller. Establishing a neural network PID parameter learning rate value range according to a neural network PID closed-loop control stability requirement; within a learning rate value range, designing a self-adaptive learning rate adjustment algorithm based on fuzzy reasoning for high-steady-state control precision and high-dynamic response of the dynamic eccentric magnetic suspension of the rotor in a wide range; the method can meet the requirements of magnetic suspension high-steady-state control precision and high dynamic response of wide rotor dynamic eccentricity.

Description

Technical Field

[0001] The present invention relates to the technical field of motors, in particular to a neural network PID suspension method for a bearingless flux-switching motor with adaptive learning rate. Background Technique

[0002] In a bearingless flux-switching motor, the permanent magnet is located on the stator, which has the advantages of low risk of demagnetization due to temperature rise of the permanent magnet, no mechanical wear, and suitability for high-speed operation; the stator of a dual-winding bearingless flux-switching motor has three-phase torque windings and three-phase suspension windings, and the torque and magnetic suspension control decoupling performance is relatively good, which has received extensive attention and research in the academic and industrial fields.

[0003] During the magnetic suspension rotation of the rotor of a bearingless flux-switching motor, the geometric center of the rotor and the geometric center of the stator are in a dynamically offset state, and the air gap length changes in real time, resulting in real-time changes in the parameters of the mathematical model of the magnetic suspension channel; if a PID controller with constant parameters is used, it is difficult to meet the requirements of the wide dynamic eccentricity magnetic suspension stability of the rotor, and the constant-parameter PID also brings a contradiction between the high steady-state control accuracy and the high dynamic response of the magnetic suspension channel.

[0004] If a neural network PID controller is used for the magnetic suspension channel, the weight coefficients of the neural network can be adjusted by using the control error information of the radial displacement of the rotor, so as to realize the real-time adaptive adjustment of the PID parameters. However, the learning rate of the parameters must meet the closed-loop control stability conditions of the entire magnetic suspension channel, and the learning rate that meets the stability requirements has not been studied yet. If a learning rate with a constant amplitude is used, it is difficult to meet the requirements of the wide dynamic eccentricity magnetic suspension stability, and it is also difficult to balance the requirements of the high steady-state control accuracy and the high dynamic response of the magnetic suspension channel. How to obtain the neural network PID magnetic suspension control of a bearingless flux-switching motor with high steady-state control accuracy and high dynamic response requirements in a wide dynamic eccentricity range is a scientific problem to be solved.

[0005] Therefore, the present invention patent proposes a neural network PID magnetic suspension control method for a bearingless flux-switching motor with adaptive learning rate. Summary of the Invention

[0006] The present invention proposes a neural network PID suspension method for a bearingless flux-switching motor with adaptive learning rate, which can meet the requirements of high steady-state control accuracy and high dynamic response of the magnetic suspension with wide rotor dynamic eccentricity.

[0007] The present invention adopts the following technical solutions.

[0008] Neural Network PID Levitation Method for Bearingless Flux-Switching Motor with Adaptive Learning Rate. This method is used for the PID magnetic levitation control of the rotor by the motor PID controller, and the neural network weight coefficient is adjusted in real time according to the control error of the rotor radial displacement to realize the real-time adjustment of the parameters of the neural network PID controller. At the same time, the value range of the learning rate of the neural network PID parameters is established according to the stability requirements of the neural network PID closed-loop control. Within the value range of the learning rate, an adaptive learning rate adjustment algorithm based on fuzzy inference is designed for high steady-state control accuracy and high dynamic response of the rotor dynamic eccentricity magnetic levitation in a wide range.

[0009] The motor is a dual-winding bearingless flux-switching motor.

[0010] The method includes the following steps;

[0011] Step 1: The radial offset amounts of the rotor spatial xy orthogonal positions of the dual-winding bearingless flux-switching motor are measured as Δx and Δy respectively by the x-axis radial displacement position sensor and the y-axis radial displacement position sensor.

[0012] Step 2: According to the given values of the xy-axis radial displacements Δx * , Δy * and Δx, Δy, calculate the xy-axis radial displacement control errors e x and e y ;

[0013]

[0014] Step 3: Taking the x-axis as an example, the x-axis radial displacement control error e x is obtained through the x-axis differentiator

[0015] Step 4: Send e x and to the x-axis learning rate fuzzy controller, and output the learning rate increments Δη x1 , Δη x2 , Δη x3 of the proportional, integral and differential coefficients in the x-axis neuron PID;

[0016] Step 5: Send e x and Δη x1 , Δη x2 , Δη x3 to the x-axis neuron PID at the same time, and output the suspension current given value for controlling the x-axis direction

[0017] Step 6: Obtain the suspension current given value for controlling the y-axis direction in the same method as in Step 5

[0018] Step 7: Send Sent to the xy / UVW transformation, outputting the given values of the three-phase floating winding currents U, V, and W Wherein

[0019]

[0020] Step Eight: Send the given values of the three-phase floating winding currents and the measured values i U 、i V 、i W to the floating winding current controller simultaneously, so as to achieve that the three-phase floating winding current tracks its given value, thereby realizing the magnetic suspension of the rotor at the geometric center of the stator.

[0021] In Step Four, the learning rate increments Δη x1 、Δη x2 、Δη x3 are obtained as follows:

[0022] The input variables e x and are mapped from the basic domain range to the quantization domain range:

[0023]

[0024] k1 is the error quantization factor, k2 is the error change rate quantization factor, E x and EC x respectively represent the specific values of e x and ; Q x 、QC x are the quantization variables of E x 、EC x respectively.

[0025] On the quantized domain, Q x 、QC x are respectively divided into seven fuzzy linguistic variables: Negative Big NB, Negative Medium NM, Negative Small NS, Zero ZO, Positive Small PS, Positive Medium PM, Positive Big PB;

[0026] For the membership function settings of the quantization variables Q x 、QC x , the membership function is used to quantify the belonging degree of the fuzzy linguistic variable, and the triangular membership function is adopted;

[0027] According to the coordinate distribution, establish the triangular membership functions μ(Q Figure 5 )、μ(QC x ) as shown in x , where the value of z is determined according to the quantization domain range,

[0028] The fuzzy rule base is shown in Table 1-3 below. The quantization variables Q x and QC x represent the fuzzified e x and

[0029] Table 1 Δη x1 Fuzzy control rules

[0030]

[0031] Table 2 Δη x2 Fuzzy control rules

[0032]

[0033] Table 3 Δη x1 Fuzzy control rules

[0034]

[0035] Establish the learning rate increment Δη x1 , Δη x2 , Δη x3 fuzzy control rule base as shown in Tables 1-3 above.

[0036] In step four, according to Figure 3 , the process of obtaining the learning rate increment Δη x1 , Δη x2 , Δη x3 includes the following steps:

[0037] Step 4.1: First, fuzzify e x and to obtain the corresponding fuzzy variables, namely: one of Negative Big (NB), Negative Medium (NM), Negative Small (NS), Zero (ZO), Positive Small (PS), Positive Medium (PM), Positive Big (PB);

[0038] Step 4.2: Send the fuzzy variables to the Mamdani fuzzy inference link to output the fuzzy variable of the learning rate;

[0039] Step 4.3: Send the fuzzy variable of the learning rate to the defuzzification link to output the final learning rate increment Δη x1 , Δη x2 , Δη x3 .

[0040] The steps to obtain the x-axis direction suspension current setting in step five are as follows:

[0041] Step 5.1: According to the learning rate increment Δη x1 , Δη x2 , Δη x3, obtain learning rates η1, η2, and η3 according to Figure 4 the flowchart;

[0042] Step 5.2, calculate the activation functions S p , S I , S D of the hidden layer neurons according to the control errors E(0) to E(k) of the x-axis radial displacement as follows:

[0043]

[0044] where T s is the digital control period

[0045] Step 5.3, calculate the neural network weight coefficients w1, w2, and w3 according to the learning rates

[0046]

[0047] where is the partial derivative of the actual output with respect to the control quantity, obtained by calculating according to the input and output mathematical model of the controlled object;

[0048] Step 5.4, calculate the neuron output p , S I , S D according to the weight coefficients w1, w2, and w3 and the activation functions S

[0049] This output is the given suspension current in the x-axis direction

[0050] In Step 5.1, the method for obtaining the learning rates η1, η2, and η3 is to set the convergence condition ε. When the absolute value of the error satisfies the convergence condition, the tuning of the learning rate is no longer performed; otherwise, the learning rate change value Δη x1 , Δη x2 , Δη x3 is calculated by the fuzzy controller, and it is judged by the convergence criterion whether the corrected learning rate η satisfies the convergence condition. If it is satisfied, it is sent to the neuron for the tuning of the weights w1, w2, and w3. If it is not satisfied, the over-limit learning rate is given 0.9 times the limit value The above steps are cycled until the error meets the expected value; specifically: first collect the displacement deviation E and its change rate EC. If |E| ≤ ε, the learning rate does not change. Otherwise, calculate Δη1, Δη2, and Δη3 through the fuzzy controller. If then η = η0 + Δη, otherwise Then the neural network tunes w1, w2, and w3 according to the learning rate.

[0051] The PID magnetic levitation control method for the rotor adopts the positional PID control algorithm. The generation of the control output quantity is realized through the computer control system based on the discrete sampling mechanism. The continuous-domain PID controller is converted into a discrete form through time discretization and numerical quantization processing. The discrete expression of the positional PID control algorithm is as follows:

[0052] The neuron PID controller adjusts the weight coefficients w1, w2, and w3 in real time through the BP algorithm. The specific change process is as follows;

[0053] Define the controller performance index function J(k) as:

[0054] J(k) = 0.5E 2 (k) Formula 8;

[0055] The gradient descent algorithm is used to adjust w1, w2, and w3 to minimize the J value. Among them, the learning rate is the weight update coefficient in the gradient descent direction. When the value is relatively high, although it can accelerate the weight update rate, it is easy to induce weight oscillation, resulting in instability of the convergence process. Moreover, the connection weights of the proportional, integral, and differential parts are different. Using a unified learning rate may cause the learning rate of a certain part to be too high and weight oscillation, or the learning rate to be too low and reduce the convergence speed. Therefore, independent learning rate parameters η1, η2, and η3 are set for w1, w2, and w3 respectively:

[0056]

[0057] In the formula, is the partial derivative of the controller performance index with respect to the weight. It can be derived according to the chain rule of differentiation:

[0058]

[0059] Similarly, there is:

[0060]

[0061]

[0062] Among them, is the partial derivative of the actual output with respect to the control quantity;

[0063] Let the Lyapunov function be V(k) = 0.5E 2 (k), then there is:

[0064] ΔV(k) = 0.5(E 2 (k) - E 2 (k - 1)) Formula 13;

[0065]

[0066] Among them, S = [S P , S I , S D T , W = [w1, w2, w3] T , η = diag[η1 η2 η3]

[0067] Since the learning rate is greater than 0, there is To ensure the convergence of the algorithm, that is, ΔV(k) < 0, then there is:

[0068]

[0069] The above formula can be transformed to get:

[0070]

[0071] Let η max = max(η1(k), η2(k), η3(k)), then there is:

[0072]

[0073] If y u (k) is not 0, then when η max satisfies the following formula, ΔV(k) < 0, which can ensure the convergence of the algorithm:

[0074]

[0075] The suspension control device of the motor rotor includes a central controller, a host computer, a rectifier circuit, a filter circuit, a three-phase inverter, a bearingless flux-switching motor, a DC bus voltage acquisition circuit, a three-winding current acquisition circuit, an isolation drive circuit, a rotor radial displacement detection circuit, and a rotor tangential position angle detection circuit.

[0076] The switching tubes of the three-phase inverter adopt insulated gate bipolar transistors or metal oxide semiconductor field effect transistors, and the central controller adopts a digital signal processor or a single-chip microcomputer; the DC bus voltage sampling circuit is composed of a combination of a Hall voltage sensor and an operational amplifier, or is composed of a combination of a parallel resistor voltage divider followed by a voltage follower composed of an operational amplifier, and its output signal is sent to the central controller; the three-phase suspension winding current acquisition circuit is composed of a combination of a Hall current sensor and an operational amplifier, or is composed of a combination of a winding series power resistor followed by a differential operational amplifier, and its output signal is sent to the central controller;

[0077] The rotor position angle detection circuit is composed of a rotary encoder followed by a level conversion circuit, or is composed of a resolver followed by a decoding circuit, and its output pulse signal is sent to the central controller; ​

[0078] The rotor radial displacement detection circuit is composed of an eddy current sensor followed by an operational amplifier, and the output signal is sent to the central controller;

[0079] Based on the obtained signal and the rotor suspension control method, the central controller outputs the switching signals of different bridge arms of the inverter, and controls the switching actions of the power switching tubes in the inverter through isolation drive to achieve stable and accurate suspension of the rotor.

[0080] In order to meet the high stability requirements of the wide rotor dynamic eccentricity magnetic suspension of the bearingless flux-switching motor, the present invention proposes a neural network PID magnetic suspension control method for the bearingless flux-switching motor with adaptive learning rate. According to the rotor radial displacement control error, the neural network weight coefficient is adjusted in real time to achieve real-time adjustment of the parameters of the neural network PID controller; the value range of the learning rate of the neural network PID parameters is established from the perspective of the closed-loop control stability of the neural network PID; within the value range of the learning rate, an adaptive learning rate adjustment algorithm based on fuzzy inference is proposed. The present invention can meet the high steady-state control accuracy and high dynamic response requirements of the wide rotor dynamic eccentricity magnetic suspension.

[0081] The present invention patent proposes a neural network PID magnetic suspension control method for the bearingless flux-switching motor with adaptive learning rate, and its beneficial effects also lie in:

[0082] (1) Compared with the existing PID method with constant parameters, the present invention can be applied to the rotor magnetic suspension stability requirements in a wide range of rotor dynamic eccentricity situations, and the rotor magnetic suspension operation is more stable;

[0083] (2) Within the value range of the learning rate that meets the magnetic suspension stability, the fuzzy inference outputs an adaptive learning rate algorithm, which can ensure the high-precision rotor magnetic suspension steady-state characteristics and the high-dynamic-response rotor magnetic suspension dynamic characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] The following further details the present invention in conjunction with the drawings and specific embodiments:

[0085] Attached Figure 1 is a schematic diagram of the neural network PID magnetic suspension control of the bearingless flux-switching motor with adaptive learning rate of the present invention;

[0086] Attached Figure 2 is a schematic diagram of the structure of the neuron PID controller of the present invention;

[0087] Attached Figure 3 is a schematic diagram of the structure of the learning rate fuzzy controller of the present invention;

[0088] Attached Figure 4 is a schematic diagram of the tuning process of the neuron PID parameters with adaptive learning rate of the present invention;

[0089] Attached Figure 5 is a schematic diagram of a triangular membership function;

[0090] Attached Figure 6 is a schematic diagram of an exemplary hardware structure of the implemented maglev drive system of the present invention;

[0091] Attached Figure 7 is for the present invention's triangular membership function setting schematic diagram for the quantization variables Q x , QC x ;

[0092] Attached Figure 8 is a schematic diagram of the Mamdani inference process of an embodiment of the present invention;

[0093] Attached Figure 9 is a schematic diagram of the relationship between the fuzzy control quantity and the input quantity changing in the embodiment. Specific implementation manner

[0094] As shown in the figure, for the neural network PID suspension method of a bearingless flux-switching motor with adaptive learning rate, the method is used for the PID magnetic suspension control of the rotor by the motor PID controller, and the neural network weight coefficients are adjusted in real time according to the rotor radial displacement control error to realize the real-time adjustment of the parameters of the neural network PID controller; meanwhile, the value range of the learning rate of the neural network PID parameters is established according to the stability requirements of the neural network PID closed-loop control; within the value range of the learning rate, an adaptive learning rate adjustment algorithm based on fuzzy inference is designed for high steady-state control accuracy and high dynamic response of the rotor dynamic eccentricity maglev in a wide range.

[0095] The motor is a dual-winding bearingless flux-switching motor.

[0096] The method includes the following steps;

[0097] Step 1, the radial offset amounts of the rotor spatial xy orthogonal positions of the dual-winding bearingless flux-switching motor are measured as Δx and Δy respectively by an x-axis radial displacement position sensor and a y-axis radial displacement position sensor;

[0098] Step 2, according to the xy-axis radial displacement givens Δx * , Δy * and Δx, Δy, calculate the xy-axis radial displacement control errors e x and e y ;

[0099]

[0100] Step 3, taking the x-axis as an example to explain, the x-axis radial displacement control error e x is obtained through the x-axis differentiator ;

[0101] Step Four: Send e x and to the x-axis learning rate fuzzy controller, and output the learning rate increments Δη x1 , Δη x2 , Δη x3 ;

[0102] Step Five: Send e x and Δη x1 , Δη x2 , Δη x3 to the x-axis neuron PID at the same time, and output the given value of the suspension current in the x-axis direction for control

[0103] Step Six: Obtain the given value of the suspension current in the y-axis direction for control in the same method as Step Five

[0104] Step Seven: Send to the xy / UVW transformation, and output the given values of the suspension winding currents of the U, V, and W phases Wherein

[0105]

[0106] Step Eight: Send the given values of the three-phase suspension winding currents and the measured values i U , i V , i W to the suspension winding current controller at the same time, so as to achieve that the three-phase suspension winding current tracks its given value, thereby realizing the magnetic suspension of the rotor at the geometric center of the stator.

[0107] In Step Four, the learning rate increments Δη x1 , Δη x2 , Δη x3 are obtained as follows:

[0108] Map the input variables e x and from the basic domain range to the quantization domain range:

[0109]

[0110] k1 is the error quantization factor, k2 is the error change rate quantization factor, E x and EC x represent the specific values of e x and respectively, and Q x , QC x are Ex and the quantization variable of EC x .

[0111] On the quantized universe of discourse, divide Q x and QC x into seven fuzzy linguistic variables respectively: Negative Big (NB), Negative Medium (NM), Negative Small (NS), Zero (ZO), Positive Small (PS), Positive Medium (PM), Positive Big (PB);

[0112] For the membership function settings of the quantization variables Q x and QC x , the membership function is used to quantify the belonging degree of fuzzy linguistic variables, and a triangular membership function is adopted;

[0113] According to the coordinate distribution, establish the triangular membership functions μ(Q Figure 5 ) and μ(QC x ) as shown in x . The value of z is determined according to the range of the quantized universe of discourse,

[0114] where the fuzzy rule base is shown in Table 1-3 below. Among them, the quantization variables Q x and QC x represent the fuzzified e x and

[0115] Table 1 Δη x1 Fuzzy control rules

[0116]

[0117] Table 2 Δη x2 Fuzzy control rules

[0118]

[0119] Table 3 Δη x1 Fuzzy control rules

[0120]

[0121] Establish the learning rate increment Δη x1 , Δη x2 , Δη x3 fuzzy control rule base as shown in Tables 1-3 above.

[0122] In step 4, according to Figure 3 , the process of obtaining the learning rate increment Δη x1 , Δη x2 , Δη x3 includes the following steps:

[0123] Step 4.1: First, take e x and Fuzzification is performed to obtain the corresponding fuzzy variables, namely, one of Negative Big (NB), Negative Medium (NM), Negative Small (NS), Zero (ZO), Positive Small (PS), Positive Medium (PM), and Positive Big (PB);

[0124] Step 4.2: Send the fuzzy variable to the Mamdani fuzzy inference section to output the fuzzy variable of the learning rate;

[0125] Step 4.3: Send the fuzzy variable of the learning rate to the defuzzification section to output the final learning rate increment Δη x1 , Δη x2 , Δη x3 .

[0126] The steps to obtain the given suspension current in the x-axis direction in Step 5 include the following: are as follows:

[0127] Step 5.1: According to the learning rate increments Δη x1 , Δη x2 , Δη x3 , obtain the learning rates η1, η2, and η3 according to the Figure 4 flowchart;

[0128] Step 5.2: Calculate the activation functions S p , S I , S D of the hidden layer neurons according to the x-axis radial displacement control errors E(0) to E(k) as follows:

[0129]

[0130] where T s is the digital control period

[0131] Step 5.3: Calculate the neural network weight coefficients w1, w2, and w3 according to the learning rate

[0132]

[0133] where, is the partial derivative of the actual output with respect to the control quantity, obtained by calculating according to the input and output mathematical model of the controlled object;

[0134] Step 5.4: Calculate the neuron output p , S I , S D according to the weight coefficients w1, w2, and w3 and the activation functions S

[0135] u(k) = w1S P + w2S I + w3S D Formula 6;

[0136] This output is the given suspension current in the x-axis direction

[0137] In step 5.1, the method for obtaining the learning rates η1, η2, and η3 is as follows: Set the convergence condition ε. When the absolute value of the error satisfies the convergence condition, the tuning of the learning rate is no longer performed; otherwise, the change value of the learning rate Δη is calculated by the fuzzy controller x1 , Δη x2 , Δη x3 , and it is judged by the convergence criterion whether the corrected learning rate η satisfies the convergence condition. If it is satisfied, it is sent to the neuron for the tuning of the weight values w1, w2, and w3. If it is not satisfied, the excessive learning rate is given 0.9 times the limit value The above steps are looped until the error meets the expected value; specifically: First, collect the displacement deviation E and its change rate EC. If |E| ≤ ε, the learning rate does not change. Otherwise, calculate Δη1, Δη2, and Δη3 through the fuzzy controller. If then η = η0 + Δη, otherwise Then the neural network tunes w1, w2, and w3 according to the learning rate

[0138] The PID magnetic suspension control method for the rotor adopts the positional PID control algorithm. The generation of the control output quantity is realized by the computer control system based on the discrete sampling mechanism. The continuous-domain PID controller is converted into a discrete form through time discretization and numerical quantization processing. The discrete expression of the positional PID control algorithm is as follows:

[0139]

[0140] The neuron PID controller tunes the weight coefficients w1, w2, and w3 in real time through the BP algorithm. The specific change process is as follows;

[0141] Define the controller performance index function J(k) as:

[0142] J(k) = 0.5E 2 (k) Formula 8;

[0143] The gradient descent algorithm is used to adjust w1, w2, and w3 to minimize the J value; among them, the learning rate is used as the weight update coefficient in the gradient descent direction. When the value is relatively high, although it can accelerate the weight update rate, it is easy to induce weight oscillation, resulting in the instability of the convergence process. Moreover, there are differences in the connection weights of the proportional, integral, and differential parts. Using a unified learning rate may cause the learning rate of a certain part to be too high and the weight to oscillate, or the learning rate to be too low and the convergence speed to decrease; therefore, independent learning rate parameters η1, η2, and η3 are set for w1, w2, and w3 respectively:

[0144]

[0145] In the formula, is the partial derivative of the controller performance index with respect to the weight, which can be derived according to the chain rule of differentiation:

[0146]

[0147] Similarly, we have:

[0148]

[0149]

[0150] Among them, is the partial derivative of the actual output with respect to the control quantity;

[0151] Let the Lyapunov function be V(k) = 0.5E 2 (k), then we have:

[0152] ΔV(k) = 0.5(E 2 (k) - E 2 (k - 1)) Formula 13;

[0153]

[0154] Among them, S = [S P , S I , S D T , W = [w1, w2, w3] T , η = diag[η1 η2 η3]

[0155] Since the learning rate is greater than 0, we have To ensure the convergence of the algorithm, that is, ΔV(k) < 0, then we have:

[0156]

[0157] The above formula can be transformed as follows:

[0158]

[0159] Let η max = max(η1(k), η2(k), η3(k)), then we have:

[0160]

[0161] If y u (k) is not 0, then when η max satisfies the following formula, ΔV(k) < 0, which can ensure the convergence of the algorithm:

[0162] ​

[0163] The suspension control device of the motor rotor includes a central controller, a host computer, a rectifier circuit, a filter circuit, a three-phase inverter, a bearingless flux-switching motor, a DC bus voltage acquisition circuit, a three-winding current acquisition circuit, an isolation drive circuit, a rotor radial displacement detection circuit, and a rotor tangential position angle detection circuit.

[0164] The switching tubes of the three-phase inverter adopt insulated gate bipolar transistors or metal oxide semiconductor field effect transistors, and the central controller adopts a digital signal processor or a single-chip microcomputer; the DC bus voltage sampling circuit is composed of a combination of a Hall voltage sensor and an operational amplifier, or adopts a combination of parallel resistor voltage division followed by a voltage follower composed of an operational amplifier, and its output signal is sent to the central controller; the three-phase suspension winding current acquisition circuit is composed of a combination of a Hall current sensor and an operational amplifier, or adopts a combination of winding series power resistors followed by a differential operational amplifier, and its output signal is sent to the central controller;

[0165] The rotor position angle detection circuit is composed of a rotary encoder followed by a level conversion circuit, or adopts a resolver followed by a decoding circuit, and its output pulse signal is sent to the central controller;

[0166] The rotor radial displacement detection circuit is composed of an eddy current sensor followed by an operational amplifier, and the output signal is sent to the central controller;

[0167] The central controller outputs the switching signals of different bridge arms of the inverter according to the obtained signals and the rotor suspension control method, and controls the switching actions of the power switching tubes in the inverter through isolation drive to achieve stable and accurate suspension of the rotor.

[0168] Embodiment 1:

[0169] The schematic diagram of the neural network PID magnetic suspension control of the bearingless flux-switching motor with adaptive learning rate proposed in this example is as Figure 1 shown. The radial offset amounts of the rotor spatial xy orthogonal positions of the double-winding bearingless flux-switching motor are measured as Δx and Δy respectively by the x-axis radial displacement position sensor and the y-axis radial displacement position sensor; according to the xy-axis radial displacement givens Δx * 、Δy * and Δx、Δy, calculate the xy-axis radial displacement control errors e x and e y . The rotor magnetic suspension control algorithms for the x-axis and y-axis are the same, and the x-axis is taken as an example for explanation. The x-axis radial displacement control error e x is obtained by differentiating the x-axis Take e x and Sent to the x-axis learning rate fuzzy controller, outputting the learning rate increments Δη x1 , Δη x2 , Δη x3 ; Send e x and Δη x1 , Δη x2 , Δη x3 to the x-axis neuron PID simultaneously, outputting the controlled x-axis direction suspension current given value i x * ; Similarly, the controlled y-axis direction suspension current given value can be obtained Send to the xy / UVW transformation, outputting the given values of the U, V, and W three-phase suspension winding currents Send the given values of the three-phase suspension winding currents and the measured values i U , i V , i W to the suspension winding current controller simultaneously, realizing the three-phase suspension winding current to track its given value, thereby realizing the rotor to be magnetically suspended at the geometric center of the stator.

[0170] The proposed neuron PID controller structure is as Figure 2 shown. Taking the x-axis magnetic suspension control as an example, the input r(k) represents Figure 1 Δx in * , y(k) represents Figure 1 Δx in, E(k) represents Figure 1 e in x , u(k) represents Figure 1 in , and the controlled object represents Figure 1 the overall of the xy / UVW transformation, the suspension winding current controller, and the BFSPMM in, and ignoring the suspension winding current control error. w1, w2, w3 (w>0) are the weight coefficients between the output neuron and the hidden layer, corresponding to the proportional coefficient K p , the integral coefficient K I , and the differential coefficient K D respectively; S p , S I , S D are the activation functions of the hidden layer neurons.

[0171] The learning rate fuzzy controller structure is as Figure 3 shown. Taking the x-axis magnetic suspension as an example, first send e x and Fuzzification is performed to obtain the corresponding fuzzy variable (i.e., one of NB (Negative Big), NM (Negative Medium), NS (Negative Small), ZO (Zero), PS (Positive Small), PM (Positive Medium), PB (Positive Big)); the fuzzy variable is sent to the Mamdani fuzzy inference section to output the fuzzy variable of the learning rate; the fuzzy variable of the learning rate is sent to the defuzzification section to output the final learning rate increment Δη x1 , Δη x2 , Δη x3 .

[0172] Among them, the fuzzy rule base is shown in Table 1-3. Among them, the quantization variables Q x , QC x respectively represent the fuzzified e x and

[0173] Simultaneously Figures 1 to 3 , establish the neuron PID parameter tuning process for learning rate adaptation as Figure 4 shown.

[0174] There must be errors in the actual situation of the system. Therefore, it is necessary to set the convergence condition ε. When the absolute value of the error satisfies the convergence condition, the tuning of the learning rate is no longer performed; otherwise, the learning rate change value Δη x1 , Δη x2 , Δη x3 is calculated by the fuzzy controller, and it is judged by the convergence criterion whether the corrected learning rate η satisfies the convergence condition. If it is satisfied, it is sent to the neuron for the tuning of the weights w1, w2, w3. If it is not satisfied, the over-limit learning rate is given 0.9 times the limit value The above steps will be looped until the error meets the expected value.

[0175] Example 2:

[0176] In this example, the basic principle of the suspension control method is as follows:

[0177] The positional PID control algorithm is adopted. The computer control system realizes the generation of the control output quantity based on the discrete sampling mechanism. It is necessary to convert the continuous-domain PID controller into a discrete form through time discretization and numerical quantization processing. The discrete expression of the positional PID control algorithm is as follows:

[0178]

[0179] The neuron PID controller tunes the weight coefficients w1, w2, w3 in real time through the BP algorithm. The specific change process is introduced as follows.

[0180] Define the controller performance index function J(k) as:

[0181] J(k) = 0.5E2 (k) (8)

[0182] The gradient descent algorithm is used to adjust w1, w2, and w3 to minimize the J value. Among them, the learning rate is the weight update coefficient in the gradient descent direction. When the value is relatively high, although it can accelerate the weight update rate, it is easy to induce weight oscillation, resulting in instability of the convergence process. Moreover, there are differences in the connection weights of the proportional, integral, and derivative parts. Using a unified learning rate may lead to too high a learning rate for a certain part and weight oscillation, or too low a learning rate and a decrease in the convergence speed. Therefore, in the present invention, independent learning rate parameters η1, η2, and η3 are respectively set for w1, w2, and w3:

[0183]

[0184] In the formula, is the partial derivative of the controller performance index with respect to the weight, which can be derived according to the chain rule of differentiation:

[0185]

[0186] Similarly, there is:

[0187]

[0188]

[0189] Among them, is the partial derivative of the actual output with respect to the control quantity.

[0190] Let the Lyapunov function be V(k) = 0.5E 2 (k), then there is:

[0191] ΔV(k) = 0.5(E 2 (k) - E 2 (k - 1)) (13)

[0192]

[0193] Among them, S = [S P , S I , S D T , W = [w1, w2, w3] T , η = diag[η1 η2 η3]

[0194] Since the learning rate is greater than 0, there is To ensure the convergence of the algorithm, that is, ΔV(k) < 0, then there is:

[0195]

[0196] The above formula can be transformed as follows:​

[0197]

[0198] Let η max = max(η1(k), η2(k), η3(k)), then we have:

[0199]

[0200] If y u (k) is not 0, then when η max satisfies the following formula, ΔV(k) < 0, which can ensure the convergence of the algorithm:

[0201]

[0202] The specific learning rate fuzzy controller adopted in this example is as follows:

[0203] Let Q x , QC x be the quantization variables of E x , EC x respectively. Set the quantization domain range of Q x , QC x to [-6, 6]. The maximum eccentric displacement Δx max of the rotor in the x direction is ±0.3 mm. From E x = 0 - Δx max , it can be seen that the basic domain of E x is [-0.3, 0.3]. When the control period is 50 μs and the maximum displacement change rate value is 0.6 / 0.00005 = 12000, then the basic domain of EC x is [-12000, 12000]. Therefore, let Q x , QC x be related to E x , EC x to satisfy formula (3), where k1 = 20 is the error quantization factor and k2 = 0.0005 is the error change rate quantization factor.

[0204] On the quantized domain, divide Q x , QC x into seven fuzzy language variables respectively: NB (Negative Big), NM (Negative Medium), NS (Negative Small), ZO (Zero), PS (Positive Small), PM (Positive Medium), PB (Positive Big).

[0205] The membership function is used to quantify the belonging degree of fuzzy language variables. The triangular membership function is adopted. In this example of the present invention, the triangular membership functions for the quantization variables Q x , QC x are set as shown in Figure 7 as follows.

[0206] Fuzzy inference is carried out using the Mamdani method, and the centroid defuzzification method is adopted, which can be expressed as:

[0207] u = (∫R(u)udu) / (∫R(u)du) (24)

[0208] where R(u) is the membership function of the fuzzy set result R obtained after fuzzy inference, and u ∈ U.

[0209] The Mamdani inference process is as Figure 8 shown. If E x = 0.16, EC x = 0.0025 / 12, then the corresponding Q x = 3.2, QC x = -2.5. The linguistic variables and their membership degrees corresponding to Q x are PS(0.4), PM(0.6), and the linguistic variables and their membership degrees corresponding to QC x are NM(0.25), NS(0.75). Then, the linguistic variables after inference through Table 1 are ZO, NS, PS, ZO respectively. When inferring the intersection of PS(0.4) of Q x and NM(0.25) of QC x , the output linguistic variable and its membership degree of Δη x1 are obtained as ZO(0.25). The clipped area S1 with a height of 0.25 is taken. Similarly, NS(0.4), PS(0.25), ZO(0.6) and their corresponding areas S2, S3, S4 can be deduced. Then, taking the union of all the inference results, the final fuzzy control quantity can be obtained, that is, the shaded area S in the figure.

[0210] Let the overlapping areas of S4 with S2 and S3 be S 24 and S 34 respectively. Then S satisfies the following relational expression:

[0211] S = S2 + S3 + S4 - S 24 - S 34 (25)

[0212] The centroids corresponding to the areas S2, S3, S4, S 24 , S 34 are -2, 2, 0, -1, 1 respectively. According to the centroid defuzzification method, the precise fuzzy control quantity can be deduced as:

[0213] u x1 = (-2 × S2 + 2 × S3 + 0 × S4 - (-1) × S 24 - 1 × S 34 ) / S ≈ -0.4 (26)

[0214] u x1 、 u x2 、 u x3 are respectively the precise fuzzy control quantities of Δη x1 、 Δη x2 、 Δη x3 . The relationship with the change of the input quantity is as Figure 9 shown. Since the same fuzzy control rules are used, u x1 、 u x2 change consistently with the input quantities Q x 、 QC x . The universes of discourse of u x1 、 u x2 、 u x3 are [-5.33, 5.33].

[0215] To ensure that the learning rate is positive, the following transformation is carried out:

[0216]

[0217] where k x1 、 k x2 、 k x3 are change factors determined according to the actual control system.

[0218] In summary, the learning rate adaptive neuron PID controller in the x direction is constructed as described above, and the y direction controller is the same.

Claims

1. Neural network PID suspension method for bearingless flux-switching motor with adaptive learning rate, characterized in that: The method is used for the PID magnetic levitation control of the rotor by a motor PID controller. The neural network weight coefficients are adjusted in real time according to the control error of the radial displacement of the rotor, so as to realize the real-time adjustment of the parameters of the neural network PID controller. At the same time, the value range of the learning rate of the neural network PID parameters is established according to the stability requirements of the neural network PID closed-loop control. Within the value range of the learning rate, an adaptive learning rate adjustment algorithm based on fuzzy inference is designed for high steady-state control accuracy and high dynamic response of the rotor dynamic eccentricity magnetic levitation in a wide range.

2. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 1, characterized in that: The motor is a dual-winding bearingless flux-switching motor.

3. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 2, wherein: The method includes the following steps; Step 1: The radial offset amounts of the rotor space xy orthogonal positions of the dual-winding bearingless flux-switching motor are measured as Δx and Δy respectively by an x-axis radial displacement position sensor and a y-axis radial displacement position sensor. Step 2: Given the radial displacements Δx * and Δy * along the x and y axes, calculate the control error e of the radial displacements along the x and y axes x and e y ; Step 3: Taking the x-axis as an example for explanation, the control error e of the radial displacement of the x-axis x Obtain e through the x-axis differentiator x ; Step 4: Send e x and to the x-axis learning rate fuzzy controller, and output the learning rate increments Δη x1 , Δη x2 , Δη x3 ; Step 5. Send e x and Δη x1 、Δη x2 、Δη x3 to the x-axis neuron PID simultaneously, and output the given suspension current in the x-axis direction for control Step 6: Obtain the given value of the suspension current in the y-axis direction, i, in the same method as in Step 5 y * ; Step 7. Send to the xy / UVW transformation to output the given values of the U, V, and W three-phase floating winding currents where Step VIII. Feed the given value of the three-phase suspension winding current and the measured values i U , i V , i W to the suspension winding current controller simultaneously, so as to achieve that the three-phase suspension winding current tracks its given value, thereby realizing the magnetic suspension of the rotor at the geometric center of the stator.

4. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 3, characterized in that: In step 4, the learning rate increments Δη x1 , Δη x2 , Δη x3 are obtained as follows: The input variables e x and are mapped from the basic domain range to the quantization domain range: k1 is the error quantization factor, k2 is the error change rate quantization factor, E x and EC x represent the specific values of e x and respectively. Q x , QC x are the quantization variables of E x and EC x respectively. On the quantized universe of discourse, Q x , QC x are respectively divided into seven fuzzy linguistic variables: Negative Big (NB), Negative Medium (NM), Negative Small (NS), Zero (ZO), Positive Small (PS), Positive Medium (PM), Positive Big (PB); For the quantization variable Q x and QC x in the membership function setting, the membership function is used to quantify the belonging degree of fuzzy linguistic variables, and the triangular membership function is adopted; Establish triangular membership functions μ(Q x ), μ(QC x ), where the value of z is determined according to the quantization domain range, and establish the learning rate increments Δη x1 , Δη x2 , Δη x3 fuzzy control rule base.

5. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 4, characterized in that: In Step 4, the learning rate increment Δη x1 , Δη x2 , Δη x3 The obtaining process includes the following steps: Step 4.

1. First, fuzzify e x and to obtain the corresponding fuzzy variables, namely: one of Negative Big (NB), Negative Medium (NM), Negative Small (NS), Zero (ZO), Positive Small (PS), Positive Medium (PM), and Positive Big (PB); Step 4.2: Send the fuzzy variable to the Mamdani fuzzy inference link to output the fuzzy variable of the learning rate. Step 4.3: Send the fuzzy variable of the learning rate to the defuzzification link to output the final learning rate increment Δη x1 , Δη x2 , Δη x3 .

6. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 5, characterized in that: Obtaining the x-axis direction suspension current reference in Step 5 The steps are as follows: Step 5.

1. Obtain learning rates η1, η2, and η3 based on learning rate increments Δη x1 , Δη x2 , Δη x3 . Step 5.2: Calculate the activation functions S p , S I , S D of the hidden layer neurons according to the control errors E(0) to E(k) of the radial displacement on the x-axis as follows: p , S I , S D as follows: Among them, T s is the digital control period Step 5.3: Calculate the neural network weight coefficients w1, w2, w3 according to the learning rate. Among them, is the partial derivative of the actual output with respect to the control quantity, which is obtained by calculating according to the mathematical model of the input and output of the controlled object; Step 5.

4. Calculate the neuron output according to the weight coefficients w1, w2, w3 and the activation functions S p , S I , S D ​ u(k) = w1S P + w2S I + w3S D Formula 6; This output is the given suspension current in the x-axis direction 7. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 6, characterized in that: In step 5.1, the method for obtaining the learning rates η1, η2, and η3 is as follows: Set the convergence condition ε. When the absolute value of the error satisfies the convergence condition, the tuning of the learning rate is no longer performed; otherwise, the change value Δη of the learning rate is calculated by the fuzzy controller. x1 , Δη x2 , Δη x3 , and it is judged by the convergence criterion whether the corrected learning rate η satisfies the convergence condition. If it is satisfied, it is sent to the neuron for the tuning of the weights w1, w2, and w3. If it is not satisfied, the over-limit learning rate is given 0.9 times the limit value. The above steps are cycled until the error meets the expected value. Specifically: First, collect the displacement deviation E and its change rate EC. If |E| ≤ ε, the learning rate does not change. Otherwise, calculate Δη1, Δη2, and Δη3 through the fuzzy controller. If then η = η0 + Δη, otherwise Then the neural network tunes w1, w2, and w3 according to the learning rate.

8. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 6, characterized in that: The PID magnetic levitation control method for the rotor adopts a positional PID control algorithm. The control output quantity is generated based on the discrete sampling mechanism by the computer control system. The continuous-domain PID controller is converted into a discrete form through time discretization and numerical quantization processing. The discrete expression of the positional PID control algorithm is as follows: The neuron PID controller adjusts the weight coefficients w1, w2, w3 in real time through the BP algorithm. The specific change process is as follows; Define the controller performance index function J(k) as: J(k) = 0.5E 2 (k) Formula 8; Use the gradient descent algorithm to adjust w1, w2, and w3 to minimize the J value. Among them, the learning rate is used as the weight update coefficient in the gradient descent direction. When the value is relatively high, although it can accelerate the weight update rate, it is prone to induce weight oscillation, resulting in instability in the convergence process. Moreover, the connection weights of the proportional, integral, and differential parts are different. Using a unified learning rate may lead to a too high learning rate for a certain part and weight oscillation, or a too low learning rate and a decrease in the convergence speed. Therefore, for w1, w2, and w3, independent learning rate parameters η1, η2, and η3 are set respectively: In the formula, is the partial derivative of the controller performance index with respect to the weight, which can be derived according to the chain rule of differentiation: Similarly, there is: Among them, is the partial derivative of the actual output with respect to the control quantity; Let the Lyapunov function be \(V(k) = 0.5E\) 2 (k), then we have: ΔV(k) = 0.5(E 2 (k) - E 2 (k - 1)) Formula 13; where, S = [S P , S I , S D T , W = [w1, w2, w3] T , η = diag[η1η2η3]​ Since the learning rate is greater than 0, there is To ensure the convergence of the algorithm, i.e., ΔV(k) < 0, then there is: The above formula can be changed to: Let η max = max(η1(k), η2(k), η3(k)), then we have: If y u (k) is not 0, then when η max satisfies the following equation, ΔV(k) < 0, which can ensure the convergence of the algorithm:

9. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 2, characterized in that: The suspension control device of the motor rotor includes a central controller, as well as a host computer, a rectifier circuit, a filter circuit, a three-phase inverter, a bearingless flux-switching motor, a DC bus voltage acquisition circuit, a three-winding current acquisition circuit, an isolation drive circuit, a rotor radial displacement detection circuit, and a rotor tangential position angle detection circuit.

10. The bearingless flux-switching motor neural network PID suspension method with adaptive learning rate according to claim 9, characterized in that: The switching tubes of the three-phase inverter adopt insulated gate bipolar transistors or metal oxide semiconductor field effect transistors, and the central controller adopts a digital signal processor or a single-chip microcomputer; the DC bus voltage sampling circuit is composed of a combination of a Hall voltage sensor and an operational amplifier, or is composed of a combination of a parallel resistor voltage divider and a voltage follower composed of an operational amplifier, and its output signal is sent to the central controller; The three-phase suspension winding current acquisition circuit is composed of a combination of a Hall current sensor and an operational amplifier, or is composed of a combination of a winding series power resistor and a differential operational amplifier, and its output signal is sent to the central controller; The rotor position angle detection circuit is composed of a rotary encoder followed by a level conversion circuit, or is composed of a resolver followed by a decoding circuit, and its output pulse signal is sent to the central controller; The rotor radial displacement detection circuit is composed of an eddy current sensor followed by an operational amplifier, and the output signal is sent to the central controller; The central controller outputs the switching signals of different bridge arms of the inverter according to the obtained signals and the rotor suspension control method, and controls the switching actions of the power switching tubes in the inverter through isolation drive to realize the stable and accurate suspension of the rotor.

Images