基于深度置信网络集成的齿轮剩余寿命预测

Research on Gear Remaining Useful Life Prediction based on Deep Belief Network Ensemble

在齿轮的剩余寿命预测中,针对传统方法需要人工提取特征和单个网络模型在泛化能力不好的问题,提出一种深度置信网络(DBN)集成模型的齿轮剩余寿命预测方法。该方法以齿轮振动信号作为输入,首先,用遗传算法优化多个DBN;其次,应用负相关学习(NCL)选择误差小同时差异度大的几个DBN搭建集成模型,用以描述不同时刻特征值之间的关系;最后,基于齿轮振动加速度数据配置好模型后进行预测,和单个深度置信网络预测比较,结果显示文章提出的集成模型在不同时刻训练集和测试集之间的误差均比单个DBN低。验证了该方法可以完成预测并且能提高泛化性能。

In the remaining useful life prediction of gear,in order to solve the problem that the traditional methods need to extract features manually and the poor generalization ability of a single network,a gear remaining useful life prediction method based on deep belief network(DBN)ensemble is proposed in this paper.In this method,the gear vibration signal is used as the input,firstly,the genetic algorithm is used to optimize the multiple DBN;Secondly,negative correlation learning is used for selecting DBNs that have small error and large difference.The ensemble model is formed of these selected DBNs to describe the relationship between the characteristics values of the gear at different times;Finally,the prediction is carried out based on the gear vibration acceleration data,and compared with the prediction results of a single deep belief network,the results show that the error between the training set and the test set of the ensemble model proposed in this paper is lower than that of a single DBN at different times.So it is verified that this method can be used to predict and improve generalization performance.