CN104852639A

CN104852639A - Parameter self-tuning speed controller of permanent magnet synchronous motor based on neural network

Info

Publication number: CN104852639A
Application number: CN201510275336.3A
Authority: CN
Inventors: 王艳; 刘洋; 吴定会; 纪志成
Original assignee: Jiangnan University
Current assignee: Jiangnan University
Priority date: 2015-05-26
Filing date: 2015-05-26
Publication date: 2015-08-19

Abstract

The invention provides a permanent magnet synchronous motor self-tuning speed controller based on artificial neural network. Based on the small signal model of the speed loop, combined with the advantages of obvious physical meaning and simple structure of the PID controller, and the strong self-adaptive ability of the neural network, the present invention designs a speed controller based on the neural network, which consists of a weight update unit, Composed of retraining units and management units. The weight update unit updates the weight value in real time to adjust the output result, and the retraining unit retrains the weight value of the neural network to limit the output within the allowable range. to determine whether the retraining unit needs to be started. In each control cycle, the management unit supervises the work of the weight updating unit and the retraining unit, so that the two cooperate with each other to achieve good real-time speed control. The invention improves the performance of speed response and has good robustness to parameter transformation.

Description

Parameter self-tuning speed controller for permanent magnet synchronous motor based on neural network

技术领域technical field

本发明属于永磁同步电机控制技术领域，具体涉及一种基于神经网络的永磁同步电机参数自整定速度控制器。The invention belongs to the technical field of permanent magnet synchronous motor control, in particular to a neural network-based permanent magnet synchronous motor parameter self-tuning speed controller.

背景技术Background technique

在永磁同步电机调速系统中，通常要求电机具有良好的动态、稳态特性及抗扰性。传统的PID控制器因其易于实现的特点在永磁同步电机驱动系统中得到了普遍的应用，其控制性能在多数条件下可以满足要求。PID控制器的输出由误差、误差累积以及误差变化率三个部分加权组成，其参数的整定和优化很大程序上依靠工程师的经验；另一方面，环境因素包括电机参数的变化会在相当程度上影响传统定参数PID控制器的控制性能。In the permanent magnet synchronous motor speed control system, the motor is usually required to have good dynamic and steady-state characteristics and anti-interference. The traditional PID controller has been widely used in the permanent magnet synchronous motor drive system because of its easy-to-implement feature, and its control performance can meet the requirements under most conditions. The output of the PID controller is composed of three weighted parts: error, error accumulation, and error change rate. The setting and optimization of its parameters largely rely on the experience of engineers; It affects the control performance of the traditional fixed parameter PID controller.

基于神经网络的控制是一种非线性的控制方案，其输出为输入的加权处理，并根据相应规则实时修正权重以得到更合适的输出结果。使用梯度下降法训练神经网络及更新神经网络权重，其工作原理为：以速度跟踪偏差的平方值为性能指标，按该性能指标相对某权重的负梯度方向为搜索方向，设置学习率来调节更新与训练的强度。相比于PID控制，神经网络控制器对环境变化具有良好的适应性，且参数不需要太多人为干扰，更具有智能性。The control based on neural network is a nonlinear control scheme, its output is the weighted processing of the input, and the weight is corrected in real time according to the corresponding rules to obtain a more suitable output result. Use the gradient descent method to train the neural network and update the weight of the neural network. Its working principle is: the square of the speed tracking deviation is the performance index, and the negative gradient direction of the performance index relative to a certain weight is the search direction, and the learning rate is set to adjust the update. and training intensity. Compared with PID control, the neural network controller has good adaptability to environmental changes, and the parameters do not require too much human interference, and are more intelligent.

发明内容Contents of the invention

本发明的目的是提供一种基于神经网络的永磁同步电机自整定速度控制器，以实现永磁同步电机的高性能速度动态响应控制，提高速度控制器对参数变化的鲁棒性。本发明的实施对象为永磁同步电机，以速度环的小信号模型及PID控制器为基础提出基于神经网络的速度控制器结构，引入重新训练机制，提高控制器的动态响应性能。The purpose of the present invention is to provide a self-tuning speed controller for permanent magnet synchronous motors based on neural network, so as to realize the high-performance speed dynamic response control of permanent magnet synchronous motors and improve the robustness of the speed controller to parameter changes. The implementation object of the present invention is a permanent magnet synchronous motor. Based on the small signal model of the speed loop and the PID controller, a speed controller structure based on a neural network is proposed, and a retraining mechanism is introduced to improve the dynamic response performance of the controller.

1.所述速度控制器的管理部分包括：权重更新单元、重新训练单元和管理单元；权重更新单元实时更新权重值以调节输出结果，重新训练单元对神经网络的权重值重新进行训练，以将输出限定在允许的范围内，管理单元通过对转矩的观测及内部转矩给定发生器，判定是否需要启动重新训练单元；1. the management part of described speed controller comprises: weight updating unit, retraining unit and management unit; The output is limited within the allowable range, and the management unit judges whether it is necessary to start the retraining unit through the observation of the torque and the internal torque given generator;

2.速度控制器的神经元结构：2. The neuron structure of the speed controller:

根据速度环的离散小信号模型及PID控制系统结构得到单神经元控制器结构，该神经元有4个输入，分别为速度给定、速度跟踪误差、速度跟踪误差变化率及一个常值输入(作为偏置)，及各输入对应的权值，输出经过tan-sigmoid函数限幅。速度环的离散小信号模型为According to the discrete small signal model of the speed loop and the PID control system structure, the single neuron controller structure is obtained. The neuron has 4 inputs, which are speed reference, speed tracking error, speed tracking error change rate and a constant input ( As a bias), and the weights corresponding to each input, the output is limited by the tan-sigmoid function. The discrete small-signal model of the speed loop is

${ΔT ΔT}_{e e} ((n no)) = = \frac{{J J}_{m m}}{{T T}_{s the s}} ((Δ Δ ω ω ((n no + + 11)) - - Δ Δ ω ω ((n no)))) + + {B B}_{m m} Δ Δ ω ω ((n no + + 11)),,$

其中ΔT_e(n)当前控制周期上转矩的变化量，Δω(n+1)、Δω(n)分别为期望的速度改变量与当前速度改变量，J_m为电机惯量，B_m为粘滞系数，T_s为控制周期。Among them, ΔT _e (n) is the variation of torque in the current control cycle, Δω(n+1) and Δω(n) are the expected speed change and the current speed change respectively, J _m is the motor inertia, B _m is the viscosity Hysteresis coefficient, T _s is the control cycle.

3.内部转矩发生器：3. Internal torque generator:

根据当前采样周期控制器输出及速度变化量估测负载情况，并根据负载转矩估测值、速度跟踪误差及控制规律计算转矩给定曲线。负载转矩估测值为其中T_L(n)为当前周期估测的负载情况，T_ANN(n)为当前周期神经网络控制器输出，Δω(n)、ω(n)分别为当前周期速度改变量及当前速度。转矩给定曲线选为Estimate the load condition based on the current sampling cycle controller output and speed variation, and calculate the given torque curve based on the estimated load torque value, speed tracking error and control law. The estimated load torque is Among them, T _L (n) is the estimated load condition of the current cycle, _TANN (n) is the output of the neural network controller of the current cycle, Δω(n), ω(n) are the speed change and the current speed of the current cycle, respectively. The torque given curve is selected as

$\begin{matrix} {T T}_{r r e e f f}^{' '} ((n no)) = = {T T}_{L L} ((n no)) + + f f (({ω ω}^{* *} ((n no)),, ω ω ((n no)))) = = {T T}_{L L} ((n no)) + + \\ {T T}_{max max} \cdot \cdot s the s i i g g n no (({ω ω}^{* *} ((n no)) - - ω ω ((n no)))) \cdot \cdot (({e e}^{γ γ \cdot \cdot a a b b s the s (({ω ω}^{* *} ((n no)) - - ω ω ((n no))))} - - 11)) \end{matrix}$

${T T}_{r r e e f f} = = \{\begin{matrix} s the s i i g g n no (({T T}_{r r e e f f}^{' '})) \cdot &Center Dot; {T T}_{max max} & w w h h e e n no & a a b b s the s (({T T}_{r r e e f f}^{' '})) > > {T T}_{max max} \\ {T T}_{r r e e f f}^{' '} & o o t t h h e e r r s the s \end{matrix}$

其中e为自然指数，ω^*为期望速度，γ用于调节控制器的刚性。where e is the natural exponent, ω ^* is the desired velocity, and γ is used to adjust the stiffness of the controller.

4.重新训练机制：4. Retraining mechanism:

在每个控制周期中，所述管理单元监管权重更新单元及重新训练单元的工作，使两者相互配合，保证速度控制器的正常运行及动态响应；结合内部转矩发生器产生的转矩给定曲线及控制器输出，判断输出偏离情况，当控制器的输出偏离转矩给定曲线超过范围[1-ξ,1+ξ]T_ref，则对神经网络重新训练，其中ξ为误差容忍度系数。重新训练过程按照梯度下降法实施，当新产生的权值使输出重新回到[1-ξ,1+ξ]T_ref范围内时即完成重新训练。In each control cycle, the management unit supervises the work of the weight updating unit and the retraining unit, so that the two cooperate with each other to ensure the normal operation and dynamic response of the speed controller; combined with the torque generated by the internal torque generator to give Determine the curve and the output of the controller, and judge the deviation of the output. When the output of the controller deviates from the torque given curve and exceeds the range [1-ξ,1+ξ]T _ref , retrain the neural network, where ξ is the error tolerance coefficient. The retraining process is implemented according to the gradient descent method, and the retraining is completed when the newly generated weight makes the output return to the range of [1-ξ,1+ξ]T _ref .

本发明的具体技术效果体现如下：Concrete technical effect of the present invention is embodied as follows:

1.对于该神经网络控制器，通过权重更新，实现良好的速度指令跟踪性能；1. For the neural network controller, good speed command tracking performance is achieved through weight updating;

2.控制器中作为偏置作用的常值输入使控制器对参数变化的鲁棒性增强，偏置输入及其权值更新能较好地补偿参数变化时引入的误差；2. The constant value input as a bias in the controller enhances the robustness of the controller to parameter changes, and the bias input and its weight update can better compensate the error introduced when the parameter changes;

3.在理想情况下，转矩给定发生器产生的转矩给定曲线使速度响应具有很好的动态和稳态性能；3. Under ideal conditions, the torque given curve generated by the torque given generator makes the speed response have good dynamic and steady-state performance;

4.引入重新训练后，较大地提高整体的动态性能，对于发生负载迅速变化、速度指令出现幅值较大突变时，能有效触发重新训练过程，使控制器输出迅速响应外部变化；4. After the introduction of retraining, the overall dynamic performance is greatly improved. When the load changes rapidly and the speed command has a large amplitude mutation, it can effectively trigger the retraining process, so that the controller output can quickly respond to external changes;

5.重新训练过程与控制器权值更新配合良好，误差容忍度系数的设置避免了由于测量误差、参数变化等因素导致误触发或频繁触发重新训练过程。5. The retraining process cooperates well with the controller weight update, and the setting of the error tolerance coefficient avoids false triggering or frequent triggering of the retraining process due to factors such as measurement errors and parameter changes.

附图说明Description of drawings

图1：永磁同步电机速度控制系统框架。Fig. 1: Framework of permanent magnet synchronous motor speed control system.

图2：神经网络速度控制器神经元结构。Figure 2: Neural network speed controller neuron structure.

图3：神经网络速度控制器整体工作框架。Figure 3: The overall working framework of the neural network speed controller.

图4(a)～(f)：控制器仿真结果。其中，图4(a)是只有权重更新时的速度响应；图4(b)是有权重更新及重新训练时的速度响应；图4(c)是有无重新训练时的速度动态性能的细节比较；图4(d)是控制器在全速度段及负载突变情况下的控制效果；图4(e)是控制器对参数变化的鲁棒性；图4(f)是常规PID控制器的控制效果。Figure 4 (a) ~ (f): Controller simulation results. Among them, Figure 4(a) is the speed response with only weight update; Figure 4(b) is the speed response with weight update and retraining; Figure 4(c) is the details of the speed dynamic performance with or without retraining Comparison; Figure 4(d) is the control effect of the controller in the full speed range and load mutation; Figure 4(e) is the robustness of the controller to parameter changes; Figure 4(f) is the control effect of the conventional PID controller Control effect.

具体实施方式Detailed ways

为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

图1为本发明的速度控制器在永磁同步电机调速系统中的应用，其设计方法如下：Fig. 1 is the application of speed controller of the present invention in permanent magnet synchronous motor speed control system, and its design method is as follows:

1.计算速度环离散小信号模型。1. Calculate the discrete small-signal model of the velocity loop.

电机中，输出转矩与转速之间的关系为In the motor, the relationship between output torque and speed is

T_e＝T_L+B_mω+J_mpω (1)T _e ＝T _L +B _m ω+J _m pω (1)

其中T_e,T_L分别为电磁转矩、负载转矩，J_m为电机及负载惯量，B_m为粘滞系数，p为微分算子，ω为电机转速。在很小的时间段内，认为负载转矩、惯量、粘滞系数等参数不变，可以得到Among them, T _e and T _L are the electromagnetic torque and load torque respectively, J _m is the motor and load inertia, B _m is the viscosity coefficient, p is the differential operator, and ω is the motor speed. In a small period of time, it is considered that parameters such as load torque, inertia, and viscosity coefficient remain unchanged, and it can be obtained

${T T}_{e e} + + {ΔT ΔT}_{e e} = = {T T}_{L L} + + {B B}_{m m} ((ω ω + + Δ Δ ω ω)) + + {J J}_{m m} \frac{d d ((ω ω + + Δ Δ ω ω))}{d d t t} - - - - - - ((22))$

因此可以获得速度环的小信号模型为Therefore, the small signal model of the speed loop can be obtained as

${ΔT ΔT}_{e e} = = {J J}_{m m} \frac{d d ((Δ Δ ω ω))}{d d t t} + + {B B}_{m m} Δ Δ ω ω - - - - - - ((33))$

其离散化形式为Its discretization form is

${ΔT ΔT}_{e e} ((n no)) = = \frac{{J J}_{m m}}{{T T}_{s the s}} ((Δ Δ ω ω ((n no + + 11)) - - Δ Δ ω ω ((n no)))) + + {B B}_{m m} Δ Δ ω ω ((n no + + 11)) - - - - - - ((44))$

其中，in,

ΔT_e(n)＝T_e(n)-T_e(n-1)ΔT _e (n)=T _e (n)-T _e (n-1)

Δω(n+1)＝ω(n+1)-ω(n)＝ω^*(n)-ω(n)＝e(n)Δω(n+1)=ω(n+1)-ω(n)=ω ^* (n)-ω(n)=e(n)

Δω(n)＝ω(n)-ω(n-1)＝e(n-1)，Δω(n)=ω(n)-ω(n-1)=e(n-1),

T_e(n)为当前控制周期输出转矩，ω(n)为当前电机转速。T _e (n) is the current control cycle output torque, ω (n) is the current motor speed.

设计神经网络控制器结构。Design the neural network controller structure.

根据式(4)，得到According to formula (4), we get

T_e(n)＝∑ΔT_e(n)＝f(Δω(n+1),Δω(n))T _e (n)=∑ΔT _e (n)=f(Δω(n+1),Δω(n))

＝f(ω^*(n),ω(n),ω(n-1)) (5)＝f(ω ^* (n),ω(n),ω(n-1)) (5)

结合PID控制器的输出由误差、误差累积以及误差变化率三个部分加权组成的特点，设计的神经网络速度控制器结构如图2所示。该神经元有4个输入，分别为速度给定、速度跟踪误差、速度跟踪误差变化率及一个常值输入(作为偏置)，及各输入对应的权值，输出经过tan-sigmoid函数限幅，如式(7)所示。T_max为电机输出最大转矩。Combined with the characteristics that the output of the PID controller is composed of three parts weighted by error, error accumulation and error change rate, the structure of the designed neural network speed controller is shown in Figure 2. The neuron has 4 inputs, which are speed reference, speed tracking error, rate of change of speed tracking error and a constant value input (as a bias), and the weights corresponding to each input, and the output is limited by the tan-sigmoid function , as shown in formula (7). T _max is the maximum torque output by the motor.

u(n)＝w1(n)·ω^*(n)+w2(n)·e(n)+w3(n)·Δe(n)+bias(n) (6)u(n)=w1(n)·ω ^* (n)+w2(n)e(n)+w3(n)Δe(n)+bias(n) (6)

${T T}_{A A N N N N} ((n no)) = = ((\frac{22}{11 + + {e e}^{- - 22 u u ((n no))}} - - 11)) \cdot \cdot {T T}_{max max} - - - - - - ((77))$

权重更新规则。Weight update rules.

定义速度跟踪情况的性能指标函数为使用梯度下降法更新权重：The performance index function that defines the speed tracking situation is Update the weights using gradient descent:

$w w ((n no)) = = w w ((n no - - 11)) - - \frac{\partial \partial J J ((n no))}{\partial \partial w w ((n no))} \cdot \cdot η η - - - - - - ((88))$

其中η为更新强度。使用链式法则，计算得到where η is the update intensity. Using the chain rule, we get

$\begin{matrix} \frac{\partial \partial J J}{\partial \partial w w} ((n no)) = = \frac{\partial \partial J J ((n no))}{\partial \partial e e r r r r ((n no))} \cdot \cdot \frac{\partial \partial e e r r r r ((n no))}{\partial \partial ω ω ((n no))} \cdot &Center Dot; \frac{\partial \partial ω ω ((n no))}{\partial \partial u u ((n no))} \cdot &Center Dot; \frac{\partial \partial u u ((n no))}{\partial \partial w w ((n no))} \\ = = - - e e r r r r ((n no)) \frac{44 {e e}^{- - 22 u u ((n no))}}{{((11 + + {e e}^{- - 22 u u ((n no))}))}^{22}} i i n no p p u u t t ((n no)) s the s i i g g n no ((\frac{\partial \partial ω ω ((n no))}{\partial \partial u u ((n no))})) \end{matrix} - - - - - - ((99))$

其中，u(n)＝w1(n)·ω^*(n)+w2(n)·e(n)+w3(n)·Δe(n)+bias(n)，w＝{w1,w2,w3,bias}，input＝{ω^*,e,Δe,1}，使用的符号代替其精确值。Among them, u(n)=w1(n)·ω ^* (n)+w2(n)e(n)+w3(n)Δe(n)+bias(n), w={w1,w2, w3,bias}, input={ω ^* ,e,Δe,1}, use The sign of replaces its exact value.

4.计算辅助转矩给定曲线。4. Calculate the auxiliary torque given curve.

根据式(1)，可以估测得到电机负载转矩According to formula (1), the motor load torque can be estimated

${T T}_{L L} ((n no)) = = {T T}_{A A N N N N} ((n no)) - - {J J}_{m m} \frac{Δ Δ ω ω ((n no))}{{T T}_{s the s}} - - {B B}_{m m} ω ω ((n no)) - - - - - - ((1010))$

根据估测到的转矩及当前速度跟踪误差，按照控制规律，定义转矩给定曲线为According to the estimated torque and the current speed tracking error, according to the control law, define the given torque curve as

$\begin{matrix} {T T}_{r r e e f f}^{' '} ((n no)) = = {T T}_{L L} ((n no)) + + f f (({ω ω}^{* *} ((n no)),, ω ω ((n no)))) = = {T T}_{L L} ((n no)) + + \\ {T T}_{max max} \cdot \cdot s the s i i g g n no (({ω ω}^{* *} ((n no)) - - ω ω ((n no)))) \cdot &Center Dot; (({e e}^{γ γ \cdot &Center Dot; a a b b s the s (({ω ω}^{* *} ((n no)) - - ω ω ((n no))))} - - 11)) \end{matrix}$

${T T}_{r r e e f f} = = \{\begin{matrix} s the s i i g g n no (({T T}_{r r e e f f}^{' '})) \cdot &Center Dot; {T T}_{max max} & w w h h e e n no & a a b b s the s (({T T}_{r r e e f f}^{' '})) > > {T T}_{max max} \\ {T T}_{r r e e f f}^{' '} & o o t t h h e e r r s the s \end{matrix} - - - - - - ((1111))$

5.重新训练规则。5. Retrain the rules.

以转矩偏差情况作为重新训练的性能指标，其表达式为Taking the torque deviation as the performance index of retraining, its expression is

$J J ((n no)) = = \frac{11}{22} {(({T T}_{r r e e f f} ((n no)) - - {T T}_{A A N N N N} ((n no))))}^{22} = = \frac{11}{22} e e r r r r {((n no))}^{22},,$

使用梯度下降法训练神经网络，每一步训练的权重修正为Using the gradient descent method to train the neural network, the weight of each step of training is corrected as

其中α为训练强度。根据链式法则， where α is the training intensity. According to the chain rule,

$\begin{matrix} \frac{\partial \partial J J}{\partial \partial w w} ((n no)) = = \frac{\partial \partial J J ((n no))}{\partial \partial e e r r r r ((n no))} \cdot &Center Dot; \frac{\partial \partial e e r r r r ((n no))}{\partial \partial {T T}_{A A N N N N} ((n no))} \cdot &Center Dot; \frac{\partial \partial {T T}_{A A N N N N} ((n no))}{\partial \partial w w ((n no))} \\ = = - - e e r r r r ((n no)) \cdot &Center Dot; \frac{44 {e e}^{- - 22 u u ((n no))}}{{((11 + + {e e}^{- - 22 u u ((n no))}))}^{22}} \cdot &Center Dot; i i n no p p u u t t ((n no)) \end{matrix}$

其中u(n)＝w1(n)·ω^*(n)+w2(n)·e(n)+w3(n)·Δe(n)+bias(n)，w＝{w1,w2,w3,bias}，input＝{ω^*,e,Δe,1}。where u(n)=w1(n) ω ^* (n)+w2(n) e(n)+w3(n) Δe(n)+bias(n), w={w1,w2,w3 ,bias}, input={ω ^* ,e,Δe,1}.

6.管理程序。6. Management procedures.

整个控制器中权重更新单元与重新训练单元相互配合，保证控制器的正常运行及高性能的动态响应。当控制器的输出偏离转矩给定曲线超出一定范围[1-ξ,1+ξ]T_ref时，激活重新训练过程，其中ξ为误差容忍度系数。调度流程如图3所示。The weight updating unit and the retraining unit in the whole controller cooperate with each other to ensure the normal operation of the controller and high-performance dynamic response. When the output of the controller deviates from the given torque curve beyond a certain range [1-ξ,1+ξ]T _ref , the retraining process is activated, where ξ is the error tolerance coefficient. The scheduling process is shown in Figure 3.

图4为本发明中所提出的神经网络控制器在永磁同步电机调速系统上的应用仿真结果。如图4(a)所示，对于该神经网络控制器，通过权重更新，实现了良好的速度指令跟踪性能；图4(b)为引入重新训练机制后的控制器的控制效果，图4(c)为图4(a)和4(b)的比较细节放大，引入重新训练后，超调减小，抗扰性增强，整体控制性能得到提高；图4(d)为该神经网络控制器在全速度段下的控制结果，对于发生负载迅速变化、速度指令出现幅值较大突变时，能有效触发重新训练过程，使控制器输出迅速响应外部变化，重新训练过程与控制器权值更新配合良好，重新训练过程基本没有被误触发或者频繁触发的情况；图4(e)为控制器对参数变化的鲁棒性，可以看到控制性能受参数变化的影响很小；图4(f)为相同条件下使用PID控制器时的控制效果，比较图4(b)、图4(e)和图4(f)可以看出本发明中提出的控制器在动态性能和鲁棒性上优于常规PID控制器。Fig. 4 is the application simulation result of the neural network controller proposed in the present invention on the permanent magnet synchronous motor speed control system. As shown in Figure 4(a), for this neural network controller, good speed command tracking performance is achieved through weight updating; Figure 4(b) shows the control effect of the controller after introducing the retraining mechanism, and Figure 4( c) The comparison details of Figures 4(a) and 4(b) are enlarged. After retraining is introduced, the overshoot is reduced, the immunity is enhanced, and the overall control performance is improved; Figure 4(d) is the neural network controller The control results under the full speed range can effectively trigger the retraining process when the load changes rapidly and the speed command has a large amplitude mutation, so that the controller output can quickly respond to external changes, and the retraining process and the controller weight update The cooperation is good, and the retraining process is basically not triggered by mistake or frequently; Figure 4(e) shows the robustness of the controller to parameter changes, and it can be seen that the control performance is slightly affected by parameter changes; Figure 4(f ) is the control effect when using the PID controller under the same conditions. Comparing Fig. 4(b), Fig. 4(e) and Fig. 4(f) it can be seen that the controller proposed in the present invention is in terms of dynamic performance and robustness Better than conventional PID controllers.

本领域的技术人员容易理解，以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, All should be included within the protection scope of the present invention.

Claims

1. The permanent magnet synchronous motor self-tuning speed controller based on the neural network is characterized by comprising a weight updating unit, a retraining unit and a management unit; the weight updating unit updates the weight value in real time to adjust an output result, the retraining unit retrains the weight value of the neural network to limit the output within an allowable range, and the management unit judges whether the retraining unit needs to be started or not through observation of torque and an internal torque given generator; the neuron structure of the speed controller is a single neuron structure, the neuron has 4 inputs, namely a speed setting input, a speed tracking error change rate, a constant value input and a weight corresponding to each input, and the output is limited by a tan-sigmoid function; discrete small signal model of velocity loop is

Wherein Δ T_e(n) is the amount of change in torque over the current control period, Δ ω (n +1), Δ ω (n) are the desired and current speed changes, respectively, J_mIs the inertia of the motor, B_mIs viscosity coefficient, T_sIs a control period;

the internal torque generator of the speed controller estimates the load condition according to the output of the current sampling period controller and the speed variation, and calculates a torque given curve according to the load torque estimation value, the speed tracking error and the control rule; the load torque estimation value is

Wherein T is_L(n) is the estimated load condition of the current cycle, T_ANN(n) is the output of the neural network controller in the current period, and Δ ω (n) and ω (n) are the change amount of the speed in the current period and the current speed respectively;

the given torque curve is selected as

<math> <mrow> <msub> <mi>T</mi> <mi>ref</mi> </msub> <mo>=</mo> <mfenced open='{' close=''> <mtable> <mtr> <mtd> <mi>sign</mi> <mrow> <mo>(</mo> <msubsup> <mi>T</mi> <mi>ref</mi> <mo>′</mo> </msubsup> <mo>)</mo> </mrow> <mo>·</mo> <msub> <mi>T</mi> <mi>max</mi> </msub> </mtd> <mtd> <mi>when abs</mi> <mrow> <mo>(</mo> <msubsup> <mi>T</mi> <mi>ref</mi> <mo>′</mo> </msubsup> <mo>)</mo> </mrow> <mo>></mo> <msub> <mi>T</mi> <mi>max</mi> </msub> </mtd> </mtr> <mtr> <mtd> <msubsup> <mi>T</mi> <mi>ref</mi> <mo>′</mo> </msubsup> </mtd> <mtd> <mi>others</mi> </mtd> </mtr> </mtable> </mfenced> </mrow> </math>

Wherein e is a natural index, ω^*Gamma is used to adjust the stiffness of the controller for the desired speed.

2. The self-tuning speed controller of permanent magnet synchronous motor based on neural network as claimed in claim 1, wherein in each control cycle, said management unit supervises the operation of the weight updating unit and the retraining unit, so that the weight updating unit and the retraining unit cooperate with each other to ensure the normal operation and dynamic response of the speed controller; when the output of the controller deviates from the torque given curve generated by the internal torque generator and exceeds a certain range [ 1-xi, 1+ xi ]]T_refAnd in time, retraining the neural network, wherein xi is an error tolerance coefficient.

3. The neural network-based PMSM self-tuning speed controller of claim 2, wherein the rotor is a permanent magnet rotorThe new training process is implemented according to a gradient descent method, and when the newly generated weight value makes the output return to [ 1-xi, 1+ xi ] again]T_refWhen the range is within, the retraining is completed.