改进的深度信念网络预测模型及其应用

Improved deep belief network prediction model and its application

针对深度信念网络(DBN)模型在非线性系统预测时,由于在模型构建中固定的学习率难以寻找全局最优以及学习速度慢等问题,提出了一种改进的DBN预测模型。将动量学习率引入到DBN无监督预训练阶段,改进了受限波尔兹曼机(RBM)网络以提高特征提取精度及参数在训练过程中的抗振荡能力;同时,将共轭梯度法嵌入DBN微调阶段来提高学习速度;最后,在袋式除尘器数字样机工作性能数据集上进行了验证。实验结果表明,与传统DBN及其变型模型相比,改进的DBN网络模型不仅收敛速度快而且预测精度高。

For the Deep Belief Network （DBN） model in nonlinear system prediction, it is difficult to find the global optimum and the learning speed is slow because of the fixed learning rate in the model construction. An improved DBN prediction model was proposed. The momentum learning rate was introduced into the DBN unsupervised pre-training phase, and the Restricted Bohzmann Machine （RBM） network was improved to enhance the feature extraction precision and the antioscillation ability of the parameters in the training process. At the same time, the conjugate gradient method was embedded in the DBN fine-tuning phase to speed up the learning. Finally, the improved DBN was tested on a digital prototype of bag filter dataset. The experimental results show that, compared with the traditional DBN and its variant model, the improved DBN network model not only has high convergence speed, but also has high prediction accuracy.