Abstract: In the context of industrial big data and the interconnection of everything,the health status of equipment in complex operating conditions needs to be represented from multi-dimensional dimensions.How to effectively integrate multi-dimensional monitoring data is of great significance for the accurate remaining life prediction of randomly degraded equipment.A kind of random degradation device residual life prediction method based on feature fusion depth network was put forward,and the multi-dimensional self-attention time convolutional network was constructed to do depth feature extraction extract for multi-dimensional monitoring data of time window after processing.The feature fusion method of pattern-weighted was designed to obtain the fusion degradation index,and then the stochastic process was used to model the degradation index.The network parameters,mode coefficients and failure threshold were reversibly adjusted by the optimization objective function representing the prediction effect,and a data-model interactive remaining useful life prediction method was formed to realize the automatic matching of degradation index and stochastic model.The accuracy and superiority of the proposed method were verified on the turbofan engine operation data set.

Key words: remaining useful life prediction, self-attentional mechanism, convolutional neural network, characteristics of the fusion, random process, turbofan engine

摘要： 在工业大数据和万物互联背景下,针对复杂运行条件中的设备健康状态需从多维度进行表征,如何有效地融合多维监测数据,对随机退化设备的精准剩余寿命预测具有重要意义,由此提出一种基于深度特征融合网络的数模联动随机退化设备剩余寿命预测方法,构建多维自注意力时间卷积网络对时间窗处理后的多维监测数据做深度特征提取,设计模式加权的特征融合方法获取融合退化指标,然后采用随机过程对退化指标进行建模,通过表征预测效果的优化目标函数对网络参数、模式系数和失效阈值进行反向调整,形成数模联动的剩余寿命预测方法,实现退化指标和随机模型的自动匹配。最后,在涡扇发动机运行数据集上验证了该方法的准确性和优越性。

关键词: 剩余寿命预测, 自注意力机制, 卷积神经网络, 特征融合, 随机过程, 涡扇发动机