优化DBN在BLDCM控制中应用研究

Application of Optimized DBN in BLDCM Control

深度信念网络(Deep Belief Networks,DBN)是由多层无监督的受限玻尔兹曼机(Restricted Boltzmann Machine, RBM)叠加而成的递归神经网络。针对RBM固定学习率在样本训练过程中很难寻找全局最优,引入动态学习率,用来改进RBM网络以提高特征向量映射的准确度。构造一个含有两层RBM网络,将改进型控制网络应用于无刷直流电机控制系统中,实验结果表明改进的DBN能够有效加快电机响应速度,提高控制准确度。

Deep belief network（DBN） is a kind of recurrent neural network superposed by several layers of unsupervised restricted Boltzmann machines（RBMs）. To solve the defect of RBM during sample training that the fixed learning rate is unfavorable to find the optimal value, a principle of dynamical learning rate was proposed to improve the RBM network so as to increase the accuracy of eigenvector mapping. Moreover, the construction of a DBN with two layers of RBMs can apply improved control network in the control system of the BLDCM.The experimental results indicate that the improved DBN can effectively accelerate response and also improve the accuracy.