基于FA-LN-BiGRU的机械设备剩余寿命区间预测方法  

Remaining Useful Life Interval Prediction of Mechanical Equipment Based on FA-LN-BiGRU

作　　者：梁伟阁 闫啸家 佘博[1] 张钢 田福庆[1] LIANG Weige;YAN Xiaojia;SHE Bo;ZHANG Gang;TIAN Fuqing(College of Weapon Engineering,Naval University of Engineering Wuhan,430033,China;School of Missiles and Naval Guns,Dalian Naval Academy Dalian,116000,China)

机构地区：[1]海军工程大学兵器工程学院,武汉430033 [2]大连舰艇学院导弹与舰炮系,大连116000

出　　处：《振动．测试与诊断》2023年第3期513-519,620,621,共9页Journal of Vibration,Measurement & Diagnosis

基　　金：国家自然科学基金资助项目(61640308);湖北省自然科学基金资助项目(2019CFB362)。

摘　　要：针对数据驱动融合模型存在前后模型不匹配、关键信息丢失等问题,提出了一种端对端的预测方法,即基于特征注意力机制的对数正态分布和双向门控循环单元融合(feature attention-lognorm-bidirectional gated recurrent unit,简称FA-LN-BiGRU)的剩余寿命区间预测方法。首先,利用特征注意力机制从多维度、非线性和大规模的传感器信号中提取出关键特征向量;其次,采用BiGRU网络从前向和后向2个方向对注意力加权特征的时变特性进行建模学习,并通过最大似然估计损失函数来训练网络参数,获得网络隐含状态输出向量的概率分布;最后,计算出基于对数正态分布的概率密度函数,实现设备剩余寿命(remaining useful life,简称RUL)不确定性的衡量。分析结果表明,对于运行条件复杂和故障模式多变的多维监测数据,所提方法能够深入挖掘性能退化信息,有效提高机械设备剩余寿命点预测和区间预测的准确度和可靠性。Prediction models based on deep learning methods is difficult to measure the uncertainty of the re-maining life of mechanical equipment.In particular,statistical data-driven predictive models are difficult to de-scribe the coupling relationship between multi-dimensional sensor data.And the data-driven fusion model has the problem of loss of key information.To solve these problems,an end-to-end remaining useful lifetime interval prediction method is proposed based on feature attention-lognorm-bidirectional gated recurrent unit(FA-LN-Bi-GRU).First,the feature attention mechanism is used to extract key feature vectors from multi-dimensional,nonlinear and large-scale sensor signals.Then,the BiGRU network is used to model the time-varying character-istics of the attention-weighted features from both forward and backward directions.And the network parameters are trained through the maximum likelihood estimation loss function to obtain the probability distribution of the network hidden state output vector.Thus,the probability density function based on log-normal distribution is calculated to realize the measurement of equipment remaining useful life(RUL)uncertainty.The analysis re-sults show that the proposed method can deeply mine performance degradation information for multi-dimension-al monitoring data with complex operating conditions and variable failure modes.The accuracy and reliability of the remaining life point prediction and interval prediction of mechanical equipment are effectively improved.

关 键 词：剩余寿命预测 对数正态分布 融合预测模型 区间预测 特征注意力机制 

分 类 号：TH17[机械工程—机械制造及自动化]