深度置信网络拟合异步电动机逆变器控制开关表
|
摘要
利用深度神经网络学习算法拟合异步电动机控制系统的逆变器开关表控制策略关系,并替代原逆变器开关表电路与逆变器一起形成深度学习控制器,是异步电动机智能控制解决方案的关键技术之一。即原逆变器开关表的输入输出数据作为深度置信神经网络的训练样本,从中学习拟合出具有逆变器开关表控制策略的深度网络拓扑与参数,借助AD/DA转换器接入实际系统,完成替代逆变器开关表硬件电路,智能化地实现异步电动机的控制策略。通过仿真比较两类控制器在各种参考信号下的控制效果,验证了深度置信网络拟合逆变器开关表控制的有效性和精确性。
The hardware controllers were designed by means of the LTI control theory,while neural network controllers are created by approximately fitting the strategies of the hardware controllers through pre-training the neural network algorithms. The approach is based on the fact that neural network algorithms can represent nonlinear functions as exactly as desired,whenever the neural networks have sufficient complexity in terms of neurons and hidden layers numbers. To illustrate the theoretical features,the inverter switching table of an induction motor control system being replaced with deep belief network( DBN) algorithm was examined. The input/output data of the inverter switching table were used as the training reference and target data for training the DBN algorithm,which in turn replaced the inverter switching table. This yielded an induction motor system driven by the DBN-algorithm control strategies. Numerical simulations and comparisons in between time responses under the inverter switching table of hardware circuits and those of the suggested DBN ones validate the expected control performances and accuracy.
The hardware controllers were designed by means of the LTI control theory,while neural network controllers are created by approximately fitting the strategies of the hardware controllers through pre-training the neural network algorithms. The approach is based on the fact that neural network algorithms can represent nonlinear functions as exactly as desired,whenever the neural networks have sufficient complexity in terms of neurons and hidden layers numbers. To illustrate the theoretical features,the inverter switching table of an induction motor control system being replaced with deep belief network( DBN) algorithm was examined. The input/output data of the inverter switching table were used as the training reference and target data for training the DBN algorithm,which in turn replaced the inverter switching table. This yielded an induction motor system driven by the DBN-algorithm control strategies. Numerical simulations and comparisons in between time responses under the inverter switching table of hardware circuits and those of the suggested DBN ones validate the expected control performances and accuracy.
引文
[1]李萌,秦品乐,李传朋.基于多新息理论的深度置信网络算法[J].计算机应用,2016,36(9):2521-2525.
[2] KRIZHEVSKY W,SUTSKEVER I,HINTON G. Image net classifica-tion with deep convolutional neural networks[C]//InternationalConference on Neural Information Processing Systems. Nevada:Cur-rent Associates,2012:1097-1105.
[3] QIAN H M,ZHOU J,MAO Y B,et al. Recognizing human actionsfrom silhouettes described with weighted distance metric and kine-matics[J]. Multimedia Tools and Applications,2017,76(21):21889-21910.
[4] QIAN H M,ZHOU J,LU X B,et al. Recognizing human activities u-sing appearance metric feature and kinematics feature[J]. Journalof Electrical Imaging,2017,26(3):15-33.
[5]何汉林,孟爱华,祝甲明,等.基于优化的GRNN和BP神经网络的磁滞曲线拟合对比分析[J].机电工程,2013,30(1):116-120.
[6]陈旭伟,傅刚,陈环.基于串联BP神经网络多函数拟合的研究设计[J].现代电子技术,2013,36(22):14-16.
[7] LI M M,VERMA B. Nonlinear curve fitting to stopping power datausing RBF neural networks[J]. Expert Systems with Applications,2016,45(3):161-171.
[8] AMMAR A,BENAKCHA A,BOUREK A. Closed loop torque SVM-DTC based on robust super twisting speed controller for inductionmotor drive with efficiency optimization[J]. International Journal ofHydrogen Energy,2017,42(28):17940-17952.
[9] DUAN Y J,LV Y S,Zhang J,et al. Deep learning for control:thestate of the art and prospects[J]. Acta Automatica Sinica,2016,42(5):643-654.
[10] JING X J,LI C. An optimal PID control algorithm for training feed-forward neural networks[J]. IEEE Transactions on Industrial E-lectronics,2013,60(6):2273-2283.
[11] MITTELMAN R,LEE H,KUIPERS B. Weakly supervised learningof mid-level features with Beta-Bernoulli process restricted Boltz-mann machines[C]//Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition(CVPR). Portland:Cur-rent Associates,2013:476-483.
[12] TANAKA M,OKUTOMI M. A novel inference of a restricted Boltz-mann machine[C]//Proceedings of the International Conferenceon Pattern Recognition(ICPR2014). Stockholm:Current Associ-ates,2014:1526-1531.
[13]张春霞,姬楠楠,王冠伟.受限波尔兹曼机简介[J].工程数学学报,2013,32(2):159-173.
[2] KRIZHEVSKY W,SUTSKEVER I,HINTON G. Image net classifica-tion with deep convolutional neural networks[C]//InternationalConference on Neural Information Processing Systems. Nevada:Cur-rent Associates,2012:1097-1105.
[3] QIAN H M,ZHOU J,MAO Y B,et al. Recognizing human actionsfrom silhouettes described with weighted distance metric and kine-matics[J]. Multimedia Tools and Applications,2017,76(21):21889-21910.
[4] QIAN H M,ZHOU J,LU X B,et al. Recognizing human activities u-sing appearance metric feature and kinematics feature[J]. Journalof Electrical Imaging,2017,26(3):15-33.
[5]何汉林,孟爱华,祝甲明,等.基于优化的GRNN和BP神经网络的磁滞曲线拟合对比分析[J].机电工程,2013,30(1):116-120.
[6]陈旭伟,傅刚,陈环.基于串联BP神经网络多函数拟合的研究设计[J].现代电子技术,2013,36(22):14-16.
[7] LI M M,VERMA B. Nonlinear curve fitting to stopping power datausing RBF neural networks[J]. Expert Systems with Applications,2016,45(3):161-171.
[8] AMMAR A,BENAKCHA A,BOUREK A. Closed loop torque SVM-DTC based on robust super twisting speed controller for inductionmotor drive with efficiency optimization[J]. International Journal ofHydrogen Energy,2017,42(28):17940-17952.
[9] DUAN Y J,LV Y S,Zhang J,et al. Deep learning for control:thestate of the art and prospects[J]. Acta Automatica Sinica,2016,42(5):643-654.
[10] JING X J,LI C. An optimal PID control algorithm for training feed-forward neural networks[J]. IEEE Transactions on Industrial E-lectronics,2013,60(6):2273-2283.
[11] MITTELMAN R,LEE H,KUIPERS B. Weakly supervised learningof mid-level features with Beta-Bernoulli process restricted Boltz-mann machines[C]//Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition(CVPR). Portland:Cur-rent Associates,2013:476-483.
[12] TANAKA M,OKUTOMI M. A novel inference of a restricted Boltz-mann machine[C]//Proceedings of the International Conferenceon Pattern Recognition(ICPR2014). Stockholm:Current Associ-ates,2014:1526-1531.
[13]张春霞,姬楠楠,王冠伟.受限波尔兹曼机简介[J].工程数学学报,2013,32(2):159-173.