基于神经网络的感应电机定子电阻参数辨识
|
摘要
直接转矩控制系统自提出以来,已经有了长足的发展。该控制技术以其独特的思想,简单的结构及优良的特性受到国内外学者的广泛关注。它以定子坐标系定向,借助于两点式调节产生PWM信号,直接对逆变器的开关状态进行最佳控制,因而电机的参数对系统的影响小,同时计算量与矢量控制相比要小得多。本课题主要通过实验获得了大量的电压频率、定子电流、定子端温和定子电阻的数据对,并从数据中得到定子电阻与电压频率、定子电流以及定子端部温度之间的变化关系,然后利用神经网络算法对定子电阻与电压频率、定子电流和定子端部温度之间的变化关系进行了建模,并利用MATLAB仿真工具进行了仿真。
定子电阻受很多因素的影响,诸如:电机运行时间,电流,频率和环境温度等等,因此通过传统方法来寻求定子电阻与影响因素之间的关系将非常困难。于是本课题将定子电阻的多种影响因素综合为电压频率、定子电流及定子端温,采用GRNN神经网络构造定子电阻辨识模型。
本文是在全速下,对定子电阻进行辨识。由于在低频范围内,定子电阻参数变化对磁链的特性影响很大。如何提高定子电阻的观测精度是改善无速度传感器直接转矩控制系统低频特性的关键。为了提高神经网络的训练速度,同时减少中高频与低频下的训练样本对网络权值的相互影响,本文将分别对中高频和低频下的定子电阻与影响因素建立各自的神经网络,通过训练分别得到中高频和低频下的定子电阻与影响因素的关系模型。
通过与BP网络对定子电阻辨识结果的比较,本文设计的GRNN神经网络对定子电阻辨识的结果有以下优点:
1、逼近能力强。GRNN网络隐含层节点中的作用函数(基函数)采用高斯函数,高斯函数作为一种局部分布对中心径向对称衰减的非负非线性函数,对输入信号将在局部产生响应,即当输入信号靠近基函数的中央范围时,隐含层节点将产生较大的输出,由此看出这种网络具有较强的局部逼近能力。
2、学习速度快。网络最后收敛于样本量集聚最多的优化回归面,一旦学习样本确定,则相应的网络结构和神经元之间的连接权值也随之确定,网络训练过程实际上只是确定平滑参数的过程。
3、泛化能力强。在样本数据较稀少时,效果也很好,网络可以处理不稳定的数据。
Great developments have taken place since Direct Torque Control system (DTC) was brought forward. It has been a focus among domestic and foreign scholars because it has character with simple structure, novelty thought and wonderful property. DTC system is different from vector control in that it is located in stator coordinate. As a result, few parameters have effect on DTC systems and it has littler calculation. This paper mainly has obtained a great deal of data about the stator resistance which changes with different voltage frequencies, the stator currents, the stator end temperate through the experiment, and obtains stator resistance and voltage frequency, stator current as well as the stator end warm between change relations from the data, then uses the neural network algorithm to construct the model of stator resistance changing with variety voltage frequencies, stator currents and the temperature and has carried on the simulation using the MATLAB simulation tool.
The stator resistance is affected by many factors, such as the driver’s runtime, current magnitude, frequency, and temperature of the environment etc. It is extremely difficult to find the relationship through the tradition methods because of complicated variety of resistance. Therefore the main point of this paper is to apply generalized regression neural network to the detection of stator resistance; to combine many factors into three parameters: frequency, electric current and temperature. Using the generalized regression neural network, this paper structures the model in order to confirm stator resistance.
In full speed area, this paper carries on the identification to the stator resistance. Because of in the low-frequency area, the stator resistance parameter change influences the flux linkage characteristic. The key point which improves low-frequency characteristic of the non-velocity generator direct torque control system is how to enhance the precision of the stator resistance observation. In order to enhance the neural network the training speed, and reduce the mutual influence of the network weight between middle & high- frequency and low- frequency sample in training process, this paper separately establishes respective neural network to middle & high- frequency and low- frequency stator resistance model.
Comparing with the identification result of the stator resistance by the BP neural network, the generalized regression neural network which this paper designs has follow merits:
1. The ability of approach is very strong. The concealment level node function (primary function) of the generalized regression neural network uses the Gauss function. As a non-negative nonlinear function, it takes one kind of partial distribution which weakens the center radial symmetry, the Gauss function will have the response to the input signal in the part, namely the concealed level node will have the big output, when input signal nears the central scope of the primary function, it can be seen from this that this kind of network has the strong ability in approaching part.
2. The study speed is higher than BP network. The network finally converges at the optimization regression surface which gathers the most sample data, once the study samples are fixed, the e the corresponding network structure and the neuron connection weight are also fixed. In fact, the training process of the network is only to confirm the smooth parameter.
3. The ability of extend is very strong. When the sample data are scarce, the effect is also satisfying. The network may process the unstable data, too.
定子电阻受很多因素的影响,诸如:电机运行时间,电流,频率和环境温度等等,因此通过传统方法来寻求定子电阻与影响因素之间的关系将非常困难。于是本课题将定子电阻的多种影响因素综合为电压频率、定子电流及定子端温,采用GRNN神经网络构造定子电阻辨识模型。
本文是在全速下,对定子电阻进行辨识。由于在低频范围内,定子电阻参数变化对磁链的特性影响很大。如何提高定子电阻的观测精度是改善无速度传感器直接转矩控制系统低频特性的关键。为了提高神经网络的训练速度,同时减少中高频与低频下的训练样本对网络权值的相互影响,本文将分别对中高频和低频下的定子电阻与影响因素建立各自的神经网络,通过训练分别得到中高频和低频下的定子电阻与影响因素的关系模型。
通过与BP网络对定子电阻辨识结果的比较,本文设计的GRNN神经网络对定子电阻辨识的结果有以下优点:
1、逼近能力强。GRNN网络隐含层节点中的作用函数(基函数)采用高斯函数,高斯函数作为一种局部分布对中心径向对称衰减的非负非线性函数,对输入信号将在局部产生响应,即当输入信号靠近基函数的中央范围时,隐含层节点将产生较大的输出,由此看出这种网络具有较强的局部逼近能力。
2、学习速度快。网络最后收敛于样本量集聚最多的优化回归面,一旦学习样本确定,则相应的网络结构和神经元之间的连接权值也随之确定,网络训练过程实际上只是确定平滑参数的过程。
3、泛化能力强。在样本数据较稀少时,效果也很好,网络可以处理不稳定的数据。
Great developments have taken place since Direct Torque Control system (DTC) was brought forward. It has been a focus among domestic and foreign scholars because it has character with simple structure, novelty thought and wonderful property. DTC system is different from vector control in that it is located in stator coordinate. As a result, few parameters have effect on DTC systems and it has littler calculation. This paper mainly has obtained a great deal of data about the stator resistance which changes with different voltage frequencies, the stator currents, the stator end temperate through the experiment, and obtains stator resistance and voltage frequency, stator current as well as the stator end warm between change relations from the data, then uses the neural network algorithm to construct the model of stator resistance changing with variety voltage frequencies, stator currents and the temperature and has carried on the simulation using the MATLAB simulation tool.
The stator resistance is affected by many factors, such as the driver’s runtime, current magnitude, frequency, and temperature of the environment etc. It is extremely difficult to find the relationship through the tradition methods because of complicated variety of resistance. Therefore the main point of this paper is to apply generalized regression neural network to the detection of stator resistance; to combine many factors into three parameters: frequency, electric current and temperature. Using the generalized regression neural network, this paper structures the model in order to confirm stator resistance.
In full speed area, this paper carries on the identification to the stator resistance. Because of in the low-frequency area, the stator resistance parameter change influences the flux linkage characteristic. The key point which improves low-frequency characteristic of the non-velocity generator direct torque control system is how to enhance the precision of the stator resistance observation. In order to enhance the neural network the training speed, and reduce the mutual influence of the network weight between middle & high- frequency and low- frequency sample in training process, this paper separately establishes respective neural network to middle & high- frequency and low- frequency stator resistance model.
Comparing with the identification result of the stator resistance by the BP neural network, the generalized regression neural network which this paper designs has follow merits:
1. The ability of approach is very strong. The concealment level node function (primary function) of the generalized regression neural network uses the Gauss function. As a non-negative nonlinear function, it takes one kind of partial distribution which weakens the center radial symmetry, the Gauss function will have the response to the input signal in the part, namely the concealed level node will have the big output, when input signal nears the central scope of the primary function, it can be seen from this that this kind of network has the strong ability in approaching part.
2. The study speed is higher than BP network. The network finally converges at the optimization regression surface which gathers the most sample data, once the study samples are fixed, the e the corresponding network structure and the neuron connection weight are also fixed. In fact, the training process of the network is only to confirm the smooth parameter.
3. The ability of extend is very strong. When the sample data are scarce, the effect is also satisfying. The network may process the unstable data, too.
引文
[1]李梦才,张贞宝.宝钢高线交流传动直接转矩控制技术.宝钢技术,2000,1:46～51
[2]钟述文.直接转矩控制系统ACS600及其特点.宝钢技术,2001,1:48～51
[3]奚国华.基于DSC的电力牵引交流传动控制系统.机车电传动,2001,3(2):74～78
[4]桂武鸣.异步电机直接转矩控制系统的仿真.电力机车技术,2002,3:11～14
[5]孙韶元.全数字化交流电机直接转矩控制系统.北京科技大学学报,1996,5:468～472
[6]何玉彬,李新忠.神经网络控制技术及其应用.北京:科学出版社2000,11:18~21
[7]苑希敏,林宏达,张广贞.动态系统的神经网络辨识方法.《工程力学》增刊,1997:20~23
[8]焦李成.神经网络系统理论.西安:西安电子科技大学出版社,1990:32~35
[9]陈伯时.电力拖动自动控制系统.北京:机械工业出版社(第二版),1997:255～263
[10]李夙.异步电动机直接转矩控制.北京:机械工业出版社,1999:1～140
[11]李永东.交流电机数字控制系统.北京:机械工业出版社,2002:1～20,122～149,210～303
[12]陈伯时,陈敏逊.交流调速系统.北京:机械工业出版社,1998:150～160,216～220
[13]张崇巍,李汉强.运动控制系统.武汉:武汉理工大学出版社,2002:102～105,159～170
[14]陈伯时,谢鸿鸣.交流传动系统的控制策略.电工技术学报,Oct 2000,5:11～15
[15]朱震莲.现代交流调速系统.西安:西北工业大学出版社,Dec 1994:113～117
[16]赵传峰,朱承高.直接转矩控制的发展现状及前景.电气自动化,1999,4:4～7
[17] Kyo-Beum Lee, Joong-Ho Song, Choy I., Ji-Yoon Yoo. Improvement of low-speed operation performance of DTC for three-level inverter-fed induction motors. Industrial Eletronics,IEEE Transaction on Volume:48 Issue: 5,2001:1006-1014
[18] T. G. habetler, F. Profumo, M. Pastorelli, L. M. Tolbert, Direct torque control of induction machines using space vector modulation. Industry Applications, IEEE Transaction on Volume:28 Issue:5,Sep.- Oct. 1992:1045-1053
[19]冯培悌.模糊控制器在异步电机直接转矩控制中的应用.电工技术学报,1997,6:10～14
[20]王建辉.模糊控制技术在异步电机直接转矩控制中的应用.东北大学学报(自然科学版)1997,6:640～643
[21]王萍,张磊.基于DSP的模糊直接转矩控制.微电机,2002,35(6):22～24
[22]王德明.直接转矩控制系统定子电阻的神经网络辨识.江苏理工大学学报,2000,6:73～76
[23]张素兵.直接转矩控制中神经网络状态选择器的仿真研究.煤矿自动化,1998,3:27～29
[24] Shi K. L. and Chan T. F. Direct self control of induction motor based on neural network.IEEE ON IA 37(5) 2001:1290-1299
[25]赵金.基于TMS320F240的直接转矩控制系统研究.电力电子技术,2000,5:52～54
[26]邓启文等.直接转矩控制的发展与展望.维特电机,2002,2:36～38
[27]张乃尧,阎平凡.神经网络与模糊控制.北京:清华大学出版社,1998
[28] Narandra K. S, Parthsatathy K, Identification and control of dynamic systems using neural networks. IEEE Trans. On Neural Networks, 1990, 1 (1)
[29]焦李成.神经网络的应用与实现.西安:西安电子科技大学出版社,1993
[30]徐丽娜.神经网络控制.黑龙江:哈尔滨工业出版社,1999
[31] S. Chen et al. Nonlineat System Identification Using Neural Networks Int. J. Control, 1990,51: 11911214
[32] D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley, 1989
[33] K. A. De Jong. An Analysis of The Behavior of a Class of Genetic Algorithms. IEEE Trans. On Systems, Man, and Cyb., 1986, 16(1): 12212
[34]陈燕庆,鹿浩.神经网络理论及其在控制工程中的应用.西安:西北工业大学出版社,1991
[35] Martin T. Hagan, Howard B. Demuth, Mark H. Beale. Neural Network Design,戴葵等译,李伯民审校.北京:机械工业出版,2002.9(第一版):198-201
[36] K. M. Hornik, M. Stinchcomebe and H. White,”Multilayer feedforward networks are universal approximators”, Neural Networks,vol. 2,no. 5,1989.(Chapter 11):359-366
[37]陶兴华,张俊洪,杨涛.异步电机定子电阻的观测.海军工程大学学报,Vol.14 No.6 Dec.2002:89-92
[38]王兵团.计算方法基础.北京:中国铁道出版社,2000,9(第一版):94-96
[39]胡伍生.神经网络理论及其工程应用.北京:测绘出版社2006,1:63-66
[40] P. D. Wasserman, Advanced Methods in Neural Computing. New York: Van Nistrand Reinhold, 1993:155~161.
[41] GUP, YAN X. Neural network approach to the reconstruction of freeform surface for reverse engineering[J]. Computer-Aided Design, 1995, 27(1):59~64
[42]赵振宇.模糊理论和神经网络的基础与应用[M].北京:清华大学出版社,1996
[2]钟述文.直接转矩控制系统ACS600及其特点.宝钢技术,2001,1:48～51
[3]奚国华.基于DSC的电力牵引交流传动控制系统.机车电传动,2001,3(2):74～78
[4]桂武鸣.异步电机直接转矩控制系统的仿真.电力机车技术,2002,3:11～14
[5]孙韶元.全数字化交流电机直接转矩控制系统.北京科技大学学报,1996,5:468～472
[6]何玉彬,李新忠.神经网络控制技术及其应用.北京:科学出版社2000,11:18~21
[7]苑希敏,林宏达,张广贞.动态系统的神经网络辨识方法.《工程力学》增刊,1997:20~23
[8]焦李成.神经网络系统理论.西安:西安电子科技大学出版社,1990:32~35
[9]陈伯时.电力拖动自动控制系统.北京:机械工业出版社(第二版),1997:255～263
[10]李夙.异步电动机直接转矩控制.北京:机械工业出版社,1999:1～140
[11]李永东.交流电机数字控制系统.北京:机械工业出版社,2002:1～20,122～149,210～303
[12]陈伯时,陈敏逊.交流调速系统.北京:机械工业出版社,1998:150～160,216～220
[13]张崇巍,李汉强.运动控制系统.武汉:武汉理工大学出版社,2002:102～105,159～170
[14]陈伯时,谢鸿鸣.交流传动系统的控制策略.电工技术学报,Oct 2000,5:11～15
[15]朱震莲.现代交流调速系统.西安:西北工业大学出版社,Dec 1994:113～117
[16]赵传峰,朱承高.直接转矩控制的发展现状及前景.电气自动化,1999,4:4～7
[17] Kyo-Beum Lee, Joong-Ho Song, Choy I., Ji-Yoon Yoo. Improvement of low-speed operation performance of DTC for three-level inverter-fed induction motors. Industrial Eletronics,IEEE Transaction on Volume:48 Issue: 5,2001:1006-1014
[18] T. G. habetler, F. Profumo, M. Pastorelli, L. M. Tolbert, Direct torque control of induction machines using space vector modulation. Industry Applications, IEEE Transaction on Volume:28 Issue:5,Sep.- Oct. 1992:1045-1053
[19]冯培悌.模糊控制器在异步电机直接转矩控制中的应用.电工技术学报,1997,6:10～14
[20]王建辉.模糊控制技术在异步电机直接转矩控制中的应用.东北大学学报(自然科学版)1997,6:640～643
[21]王萍,张磊.基于DSP的模糊直接转矩控制.微电机,2002,35(6):22～24
[22]王德明.直接转矩控制系统定子电阻的神经网络辨识.江苏理工大学学报,2000,6:73～76
[23]张素兵.直接转矩控制中神经网络状态选择器的仿真研究.煤矿自动化,1998,3:27～29
[24] Shi K. L. and Chan T. F. Direct self control of induction motor based on neural network.IEEE ON IA 37(5) 2001:1290-1299
[25]赵金.基于TMS320F240的直接转矩控制系统研究.电力电子技术,2000,5:52～54
[26]邓启文等.直接转矩控制的发展与展望.维特电机,2002,2:36～38
[27]张乃尧,阎平凡.神经网络与模糊控制.北京:清华大学出版社,1998
[28] Narandra K. S, Parthsatathy K, Identification and control of dynamic systems using neural networks. IEEE Trans. On Neural Networks, 1990, 1 (1)
[29]焦李成.神经网络的应用与实现.西安:西安电子科技大学出版社,1993
[30]徐丽娜.神经网络控制.黑龙江:哈尔滨工业出版社,1999
[31] S. Chen et al. Nonlineat System Identification Using Neural Networks Int. J. Control, 1990,51: 11911214
[32] D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley, 1989
[33] K. A. De Jong. An Analysis of The Behavior of a Class of Genetic Algorithms. IEEE Trans. On Systems, Man, and Cyb., 1986, 16(1): 12212
[34]陈燕庆,鹿浩.神经网络理论及其在控制工程中的应用.西安:西北工业大学出版社,1991
[35] Martin T. Hagan, Howard B. Demuth, Mark H. Beale. Neural Network Design,戴葵等译,李伯民审校.北京:机械工业出版,2002.9(第一版):198-201
[36] K. M. Hornik, M. Stinchcomebe and H. White,”Multilayer feedforward networks are universal approximators”, Neural Networks,vol. 2,no. 5,1989.(Chapter 11):359-366
[37]陶兴华,张俊洪,杨涛.异步电机定子电阻的观测.海军工程大学学报,Vol.14 No.6 Dec.2002:89-92
[38]王兵团.计算方法基础.北京:中国铁道出版社,2000,9(第一版):94-96
[39]胡伍生.神经网络理论及其工程应用.北京:测绘出版社2006,1:63-66
[40] P. D. Wasserman, Advanced Methods in Neural Computing. New York: Van Nistrand Reinhold, 1993:155~161.
[41] GUP, YAN X. Neural network approach to the reconstruction of freeform surface for reverse engineering[J]. Computer-Aided Design, 1995, 27(1):59~64
[42]赵振宇.模糊理论和神经网络的基础与应用[M].北京:清华大学出版社,1996