Abstract: To forecast the parameters of surface topography under different processing methods,a prediction method for parameters of surface topography was proposed based on hybrid Restricted Boltzmann Machine (RBM).Aiming at the problems that the generalization ability of RBM was poor and the fixed training rate was unfavorable for the network to be free from the minimal point,a sparse autoencoder was utilized to extract the features of prediction values,and a hybrid RBM neural network was designed to predict the parameters values of surface topography.In unsupervised training,a principle of dynamic learning rate was employed to improve the network so as to increase the accuracy of eigenvector mapping.For the purpose of improving the training speed at unsupervised learning stage,the rule of comparison and dispersion was adopted to conduct the rapid training of neural network.Through using the training model of hybrid RBM,the parameters of surface topography on joint surface could be obtained by arbitrarily inputting machining parameters.To guarantee that the parameters of joint surface could be directly applied to engineering,based on the parameters of surface topography,the authors the actual process of contact model application was deduced with fractal theory.In this way,the stiffness and damping value of each node with uneven load under the microstate of joint surface were introduced to a finite element model.Compared with the experimental values of the same specimen,the correctness of implementation method for joint surface was proved,which could provide a basis for optimizing the structure and improving the precision of numerically-controlled machine tools.

Key words: dynamic learning, hybrid restricted boltzmann machine, fractal theory, joint surface

摘要： 为了预测不同加工方式下的表面形貌参数,提出一种基于混合型约束玻尔兹曼机(RBM)的表面形貌参数预测方法,针对RBM泛化能力较低、且固定的训练率不利于网络跳出极小点的问题,应用稀疏自动编码机实现预测数值的特征提取,设计混合型RBM神经网络预测出表面形貌参数值。在无监督训练中,利用一种动态学习率法则改进网络来提高特征向量映射的准确度,为了提高无监督学习阶段的训练速度,使用对比分散准则快速训练神经网络,通过混合型RBM训练模型任意输入加工参数即可获得结合面的表面形貌参数。为了将结合面参数直接应用于工程,基于表面形貌参数、采用分形理论推导了接触模型应用的实现过程,将结合面微观状态不均匀载荷下各节点的刚度、阻尼值植入有限元模型,最终通过与相同试件的实验值对比,验证了结合面实现方法的正确性,为数控机床结构优化与精度提高提供了基础。

关键词: 动态学习, 混合型约束玻尔兹曼机, 分形理论, 结合面