集成参数自适应调整及隐含层降噪的深层RBM算法

摘要

深度置信网络是由若干层无监督的限制玻尔兹曼机(Restricted Boltzmann machines,RBM)和一层有监督的反馈神经网络组成的深层结构,该结构通过对低层输入的逐层抽象转化提取复杂输入及复杂分类数据的有效信息.然而,深度置信网络模型存在隐含层数及特征维数难以确定,后向有监督过程存在"导数消亡"问题,使得低层结构参数得不到有效的训练,而且噪声干扰直接影响识别结果的问题.针对以上问题,提出以下解决方法:每个隐含层位置构建当前层输出与样本标签之间的映射转换矩阵,根据理论标签与实际标签之间的差异,实现隐含层特征维数的自适应调整,缓解"导数消亡"问题,同时在第一隐含层位置进行特征空间降噪,保证计算效率及提高诊断模型的识别效果.复杂工况的齿轮箱故障模拟实验,验证所提方法的有效性.%Deep belief nets consist of several-layered unsupervised restricted Boltzmann machines and one-layered su-pervised feedback neural network. It digs the inner structure and pattern of the complex input data through effective information abstraction layer by layer, which can well reflect the input mode. However, the hidden layer numbers and the feature dimension are difficult to determine. The feedback process exhibits the vanishing gradient problem, which results in ineffective structural parameters training for lower layers. Moreover, noise affects the recognition results directly. To aim at the problem, a transformation matrix between samples and labels is made for each layer to realize adaptive adjustment of the parameter of hidden layer, and the feature of the hidden layer is de-noised for improving recognition accuracy and calculation efficiency. Simulation experiments on fault diagnosis of a gearbox in complex working conditions have proved the effectiveness of the proposed method.