基于少数据样本的滚动轴承寿命分段预测方法  

Segmental prediction method of rolling bearing life based on few data samples

作　　者：张朋 马孝育 王恒迪[1] 李畅 邓四二[1] 邱小彪 ZHANG Peng;MA Xiaoyu;WANG Hengdi;LI Chang;DENG Sier;QIU Xiaobiao(School of Mechanical and Electrical Engineering,Henan University of Science and Technology,Luoyang 471000,China;Shaanxi Huayan Aviation Instrument Co.,Ltd.,Xi an 710199,China;Shandong Chaoyang Bearing Co.,Ltd.,Dezhou 253200,China;Ningbo Bigger Group Co.,Ltd.,Ningbo 315000,China)

机构地区：[1]河南科技大学机电工程学院,河南洛阳471000 [2]陕西华燕航空仪表有限公司,陕西西安710199 [3]山东朝阳轴承有限公司,山东德州253200 [4]宁波更大集团有限公司,浙江宁波315000

出　　处：《机电工程》2024年第8期1415-1422,共8页Journal of Mechanical & Electrical Engineering

基　　金：山东省重点研发计划项目(2020CXGC011003);宁波市重点研发计划暨“揭榜挂帅”项目(2023Z006)。

摘　　要：针对少数据样本下,滚动轴承难以准确预测剩余使用寿命(RUL)的问题,提出了一种结合卷积长短期记忆网络(ConvLSTM)与对抗性判别域自适应网络(ADDA)的轴承寿命分段预测方法。首先,利用稀疏概率自注意力机制对特征集进行了筛选,提取了具有时变性的特征集,以获取最优全局特征,确定分段点以作为ADDA模型的输入;然后,针对不同阶段的退化特点建立了相应的健康评估指标;对处于健康状态的轴承,利用ConvLSTM网络预测了轴承健康阶段的寿命,将健康阶段预测数据作为局部特征输入ADDA网络与最优特征集(全局特征),进行了对抗训练,以实现故障阶段的寿命预测,并使用全连接层输出滚动轴承的预测剩余使用寿命;最后,采用PHM2012数据集与工程试验数据分别对模型进行了验证。研究结果表明:相较于ConvLSTM模型、RNN-HI模型、CNN-LSTM模型,ConvLSTM-ADDA寿命预测方法的平均绝对误差分别降低了78.16%、53.14%、67.13%,平均得分分别提高了66.42%、92.81%、32.37%;相较于LSTM模型、CNN-LSTM模型以及Transformer模型,ConvLSTM-ADDA寿命预测方法的均方误差分别降低了80.11%、54.95%、55.94%。因此,该算法模型能够实现对较少数据样本的轴承寿命进行RUL预测的目的,且具有较高的精度。Aiming at the problem that it is difficult to accurately predict the remaining useful life(RUL)of rolling bearings with few data samples,a life segment prediction method combining convolutional long short-term memory(ConvLSTM)and adversarial discriminative domain adaptation(ADDA)for lifetime segmentation prediction was proposed.Firstly,the feature set was screened by ProbSpare self-attention,and the feature set with time-varying characteristics was extracted to obtain the optimal global features,which were used to determine the segmentation point and used as inputs to the ADDA model.Secondly,the corresponding health assessment indexes were established for the degradation characteristics at different stages.Then,ConvLSTM network was used to predict the life of the bearing in the healthy stage,the predicted data of the healthy stage was input as local features into the ADDA network with the optimal set of features(global features)for adversarial training to realize the life prediction in the failure stage,and output the predicted remaining life of the rolling bearing through the fully connected layer.Finally,the model with the PHM2012 dataset and engineering test data was validated.The research results show that comparing to the ConvLSTM model,the RNN-HI model,and the CNN-LSTM model,the proposed ConvLSTM-ADDA life prediction method reduces the mean absolute error by 78.16%,53.14%,and 67.13%respectively,improves the mean score by 66.42%,66.81%,and 32.37%respectively.Comparing to the LSTM model,the CNN-LSTM model,and the Transformer model,the proposed ConvLSTM-ADDA life prediction method reduces the mean square error by 80.11%,54.95%,and 55.94%respectively.Therefore,the algorithmic model can realize the RUL prediction purpose of bearing life for fewer data samples with high accuracy.

关 键 词：对抗性判别域适应网络 卷积长短期记忆网络 稀疏概率自注意力机制 少数据样本 分阶段寿命预测 剩余使用寿命 

分 类 号：TH133.3[机械工程—机械制造及自动化] TP206.3[自动化与计算机技术—检测技术与自动化装置]