自动扶梯轴承健康指标提取及剩余寿命预测方法研究

Health index extraction and remaining useful life prediction method of escalator bearing

针对小样本情况下滚动轴承的健康指标提取及寿命预测问题,提出了一种结合变分模态分解(VMD)和门控循环网络(GRU)的轴承健康指标提取及寿命预测方法。首先,针对现场设备轴承故障全寿命数据稀缺问题,对样本数据进行扩增,通过VMD分解将一维数据分解为多维数据;然后,对多维数据进行时域特征提取及归一化处理;最后,以归一化特征为输入,以轴承寿命百分比为输出训练GRU网络,用此网络提取出了轴承健康指标,进行了轴承剩余寿命(RUL)预测,并对数据进行了验证和分析。研究结果表明:基于试验台数据,相比于原始数据GRU方法(Raw-GUR)、循环神经网络健康指标方法(RNN-HI)、卷积神经网络-长短时记忆网络方法(CNN-LSTM),采用该方法单调性分别可以提高0.26、0.19和0.08,基于自动扶梯电动机轴承故障数据单调性分别提高0.48、0.3和0.07;基于VMD特征空间扩增及特征提取后,VMD-GRU方法构建的轴承健康指标具有更优的单调性,可更细致获得信号局部特征空间,可以克服传统算法仅对全局样本特征进行提取难以捕捉局部特征的缺点,使GRU网络构建的健康指标性能更优。

In order to solve the problem of life index extraction and prediction of rolling bearings in the case of small samples,a bearing health index extraction and prediction method based on variational mode decomposition(VMD)and gate recurrent unit(GRU)network was proposed.First,in view of the scarcity of bearing fault life data of field equipment,the sample data was expanded,and the one-dimensional data was decomposed into multi-dimensional data by VMD decomposition.Then,the multi-dimensional features were extracted and normalized in time domain.Finally,the normalized features were used as the input and the bearing life percentage was used as the output to train the GRU network,which was used to extract the bearing health index and predict the remaining useful life,and the data were verified and analyzed.The results show that the monotonicity is improved by 0.26,0.19 and 0.08 respectively comparing with GRU without VMD(Raw-GUR),recurrent neural network health index(RNN-HI)and convolutional neural network-LSTM(CNN-LSTM)based on test-bed data;the monotonicity is improved by 0.48,0.3 and 0.07 respectively based on escalator motor bearing fault data.Therefore,after VMD feature space expansion and feature extraction,the bearing health index constructed by VMD-GRU method has better monotonicity,and the local feature space of signals can be obtained in more detail,which overcomes the disadvantage that the traditional algorithm is difficult to capture local features only considering the global feature of the samples,and makes the performance of health index constructed by GRU network better.