阅读说明：本技术 一种基于深度学习网络的srm转矩脉动抑制方法 (SRM torque ripple suppression method based on deep learning network ) 是由 李孟秋 蔡辉 沈仕其 于 2019-08-03 设计创作，主要内容包括：一种基于深度学习网络的SRM转矩脉动抑制方法,包括以下步骤：1)建立开关磁阻电机数学模型；2)搭建控制系统框图并建立控制逻辑流程图；3)搭建网络模型结构；4)进行Dropout深度学习网络的离线训练。提出设计一种基于随机Dropout深度学习网络的转矩观测器,通过优化网络结构,使转矩-电流-位置等非线性数据快速收敛拟合,提高了实际转矩的采集准确性及实时性。(a SRM torque ripple suppression method based on a deep learning network comprises the following steps: 1) establishing a mathematical model of the switched reluctance motor; 2) building a control system block diagram and a control logic flow chart; 3) building a network model structure; 4) offline training of the Dropout deep learning network is performed. The torque observer based on the random Dropout deep learning network is provided and designed, nonlinear data such as torque-current-position and the like are subjected to fast convergence fitting through optimizing a network structure, and the acquisition accuracy and the real-time performance of actual torque are improved.)

1. a SRM torque ripple suppression method based on a deep learning network is characterized by comprising the following steps:

1) establishing a mathematical model of the switched reluctance motor;

2) building a control system block diagram and a control logic flow chart;

3) Building a network model structure;

4) offline training of the Dropout deep learning network is performed.

2. The SRM torque ripple suppression method based on the deep learning network according to claim 1, wherein in the step 1), the electromagnetic torque calculation formula is:

where θ is the rotor position angle, ik is the phase current, and Tk is the electromagnetic torque.

3. the SRM torque ripple suppression method based on the deep learning network of claim 1, wherein in the step 1), a torque distribution method (TSF) is proposed to suppress the torque ripple during the phase commutation, and the TSF sets up an overlap region to distribute the fixed output torque to the overlapped two phases to share the output during the phase commutation of the adjacent phases through a certain distribution function.

the reference torque for the k-phase may be defined as:

in a linear TSF, the function frize is defined as:

Whereas in a sinusoidal TSF, it is defined as:

For any TSF, the function ffall is related to the function frise:

f(θ)＝1-f(θ+θ-θ+θ)。

4. the method for suppressing the SRM torque ripple based on the deep learning network as claimed in claim 1, wherein in the step 2), a torque observer based on a dropout deep learning network structure is constructed by training current-angle-torque discrete sample data, current and angle are sampled in real time as two-dimensional input of the observer, actual torque is estimated as output of the observer, hysteresis control is performed on the actual torque and each phase torque, six switching logic signals are output, and conduction of a power device is controlled.

5. The SRM torque ripple suppression method based on the deep learning network is characterized in that in the step 3), the network takes minimum Mean Squared Error loss (MSE) as an optimization target, and the MSE loss is defined as the square sum of residuals between a predicted value and an actual value. . The network adopts a Relu activation function which has high convergence speed and can effectively solve the problem of gradient disappearance, and the function is defined as:

ReLU(x)＝max(0,x)

The network training uses a small batch random gradient descent algorithm as an optimization mode, and the formula is as follows:

Where η is the learning rate.

6. the method for suppressing the SRM torque ripple based on the deep learning network of claim 1, wherein in the step 4), the relevant parameters of the experiment are set as: the learning rate of the optimization algorithm is 1e-3, the batch size (Batchsize) is 360, and the Iteration number (Iteration) is 5000.

dropout has direct influence on the training precision of the depth network model, and the experimental precision of the depth network model under the action of random dropout inactivation parameters with different sizes is analyzed by taking high convergence speed and high fitting precision as targets to select the optimal value of dropout.

Technical Field

the invention belongs to the field of motor control, and relates to a method for restraining SRM torque ripple based on a deep learning network.

background

the switched reluctance motor has the characteristics of simple structure, low cost, strong fault-tolerant capability, high reliability and the like, and is widely applied to the fields of mining, transportation and the like. The special iron core magnetic circuit structure and the nonlinear electromagnetic characteristic thereof cause the switched reluctance torque to have larger pulsation, and particularly, the pulsation is obvious when the sum of the torque output of each phase in a phase change interval is not balanced with the load torque at low speed.

Disclosure of Invention

the invention aims to solve the technical problem of overcoming the defects in the prior art and provides a method for inhibiting the SRM torque ripple based on a deep learning network, which enables nonlinear data such as torque-current-position to be subjected to fast convergence fitting by optimizing a network structure and improves the acquisition accuracy and real-time property of actual torque.

in order to solve the technical problems, the technical scheme adopted by the invention is as follows:

A SRM torque ripple suppression method based on a deep learning network is characterized by comprising the following steps:

1) Establishing a mathematical model of the switched reluctance motor;

2) building a control system block diagram and a control logic flow chart;

3) Building a network model structure;

4) Performing offline training of a Dropout deep learning network;

in the step 1), the electromagnetic torque calculation formula is as follows:

where θ is the rotor position angle, ik is the phase current, and Tk is the electromagnetic torque

in step 1), it is proposed to suppress torque ripple at the time of commutation by a torque distribution method (TSF) which establishes an overlap region in the commutation period of the adjacent phase currents by applying a constant distribution function to a fixed output torque and distributes the overlap region to the two overlapped phases to share the output.

The reference torque for the k-phase may be defined as:

In a linear TSF, the function frize is defined as:

whereas in a sinusoidal TSF, it is defined as:

For any TSF, the function ffall is related to the function frise:

f(θ)＝1-f(θ+θ-θ+θ)

In the step 2), a torque observer based on a dropout deep learning network structure is constructed by training current-angle-torque discrete sample data, current and angle are sampled in real time to serve as two-dimensional input of the observer, actual torque is estimated to serve as output of the observer, hysteresis control is performed on the actual torque and each phase torque, six paths of switch logic signals are output, and conduction of a power device is controlled.

In the step 3), the network takes minimum Mean Squared Error (MSE) as an optimization target, and the MSE loss is defined as the sum of squares of residuals between the predicted value and the actual value. . The network adopts a Relu activation function which has high convergence speed and can effectively solve the problem of gradient disappearance, and the function is defined as:

ReLU(x)＝max(0,x)

The network training uses a small batch random gradient descent algorithm as an optimization mode, and the formula is as follows:

θ＝θ-η·▽J(θ；x；y)

where η is the learning rate.

in the step 4), the relevant parameters of the experiment are set as follows: the learning rate of the optimization algorithm is 1e-3, the batch size (Batchsize) is 360, and the Iteration number (Iteration) is 5000.

dropout has direct influence on the training precision of the depth network model, the experimental precision of the depth network model under the action of random dropout inactivation parameters with different sizes is analyzed by taking high convergence speed and high fitting precision as targets, and the optimal value of dropout is selected.

drawings

FIG. 1 is a block diagram of a control system;

FIG. 2 is a control logic flow diagram;

FIG. 3 is a diagram of a deep network model architecture;

FIG. 4 is an error surface of the training results and the actual torque;

Figure 5 random dropout network training results.

Detailed Description

the present invention will be described in detail below with reference to the accompanying drawings and examples.

The method comprises the steps of establishing a switched reluctance motor mathematical model in the step 1, establishing a control system block diagram and a control logic flow diagram in the step 2, establishing a network model structure in the step 3, and performing offline training of a Dropout deep learning network in the step 4.

in step 1, in the mathematical model of the switched reluctance motor, the torque of each phase of the SRM may be greatly changed with the change of the rotor position and the phase current, and generally, the electromagnetic torque Tk generated by the action of any rotor position angle θ and phase current ik may be calculated by calculating the magnetic common function offset:

where m is the number of motor phases, Wc is defined as:

The two basic equations for the phase flux linkage Ψ k are:

Ψ(θ,i)＝L(θ,i)·i (3)

Where Lk is the phase inductance, uk is the phase voltage and Rk is the phase resistance.

The magnitude of the phase inductance Lk is related to the rotor position angle and the phase current, and when analyzing the phase inductance Lk, the influence of magnetic circuit saturation is generally ignored, and the Lk is not related to the phase current ik but only a function of the rotor position angle theta, so that an electromagnetic torque calculation formula can be obtained:

when the derivative of the phase inductance with respect to θ is positive, the torque produced is also positive, as shown in fig. 1, the phase inductance Lk and its derivative vary periodically. In an ideal case, a positive torque can only be generated between the start and end θ bo and θ eo of the overlap angle of the stator and rotor poles. In practice, however, a positive torque can be generated in the entire region between the asymmetric θ u and the symmetric θ al positions, which is equal to half the rotor pole pitch (τ/2):

Torque ripple during commutation is suppressed by a torque distribution method (TSF) which establishes an overlap region during commutation of the adjacent phase currents by applying a fixed output torque to the overlapping two phases to share the output.

in the non-zero region of the TSF, there are k sub-regions where the phases alone provide the entire motor torque () and there are sub-regions where one or more phases together provide torque (i.e., commutation or overlap regions), satisfied. Only the case where no more than two phases are energized simultaneously in the overlap region is considered herein, under which condition most of any of the three-phase and four-phase SRMs will not generate negative torque. The reference torque for the k-phase may be defined as:

in a linear TSF, the function frize is defined as:

whereas in a sinusoidal TSF, it is defined as:

For any TSF, the function ffall is related to the function frise:

f(θ)＝1-f(θ+θ-θ+θ) (10)

There will be numerous TSF curves depending on the manner in which the three-phase composite command torque is distributed to each phase.

in the step 2, a control system block diagram is established and a control logic flow diagram is established, referring to fig. 1, the control system block diagram is constructed, a torque observer based on a dropout deep learning network structure is constructed by training current-angle-torque discrete sample data, current and angle are sampled in real time to be used as two-dimensional input of the observer, actual torque is estimated to be used as output of the observer, hysteresis control is carried out on the actual torque and each phase torque, six switching logic signals are output, and conduction of a power device is controlled. Referring to fig. 2, a control logic flow diagram is shown. (see FIG. 1, FIG. 2)

In the step 3, in the network model structure, compared with the traditional network model, the dropout network newly defines a hyper-parameter p to represent the activation probability of the neuron node. The network output is:

Wherein, L is the total number of network layers, the Bernoulli function represents that a vector of 0 or 1 is randomly generated by probability p, m (L) is a vector which obeys the Bernoulli distribution, and m (L) is multiplied by an output matrix a (L) of the L layer to obtain sparse output of the L layer. The application of the random dropout algorithm is equivalent to extracting a sub-network from the complete network structure for training. For a deep network comprising n neurons, the dropout algorithm can extract 2n sub-networks with different structures in total, and finally, the generalization capability and stability of the whole network model are greatly improved by performing model fusion on the sub-networks.

a curve of the motor is obtained through finite element analysis and used as a training sample of the neural network, wherein the current value range is 0-250A, a group of samples are taken every 2.5A, the angle value range is 0-0, and 18281 groups of samples are taken every group of samples.

referring to fig. 3, the network takes the minimum mean square error loss as an optimization target, and the MSE loss is defined as the sum of squares of residuals between predicted values and actual values. The network adopts a Relu activation function which has high convergence speed and can effectively solve the problem of gradient disappearance, and the function is defined as:

ReLU(x)＝max(0,x)

the network training uses a small batch random gradient descent algorithm as an optimization mode, and the formula is as follows:

θ＝θ-η·▽J(θ；x；y)

Where η is the learning rate. (see FIG. 3)

In the step 4, in the off-line training of the Dropout deep learning network, proper rotor, current and torque sample data are selected for training, the rotor, the current and the torque sample data are trained through a network structure, and based on the rapid fitting and approximation capability of Dropout, relevant parameters of the experiment are set as: the learning rate of the optimization algorithm is 1e-3, the batch size (Batchsize) is 360, and the Iteration number (Iteration) is 5000.

Discrete data samples among motor current, torque and angles are obtained through finite element simulation, in order to improve training precision, 1 degree is selected as an angle interval, an angle interval is 0-360 degrees (electrical angle), 1A is selected as a current sampling interval, the value range is 0-250A, data are 361 rows and 251 columns, total data are 90611, dropouts have direct influence on the training precision of the depth network model, the target of high convergence speed and high fitting precision is taken, the experimental precision of the depth network model under the action of random dropout inactivation parameters of different sizes is analyzed, and then the optimal value of the dropouts is selected

as can be seen from the comparison of the training results shown in FIG. 4 and FIG. 5, the training result has a smaller error than the actual training result, the training error is kept within 3N m, the maximum torque error is less than the rated torque, and the requirement of the torque calculation accuracy is met. Through the analysis, the torque observer constructed by the random Dropout neural network reduces the extra storage space required by a table look-up method through offline model training, avoids complex operation brought by an analytic method, can well represent the complex nonlinear relation between the SRM torque and the input current and angle, and can be applied to engineering practice. (see FIG. 4, FIG. 5)