基于CBAM-ResNet的轴承剩余寿命预测

Prediction of Bearing Residual Life Based on CBAM-ResNet

针对以往的轴承剩余使用寿命(RUL)预测方法预测出的RUL值偏差过大的问题,提出了一种将卷积注意力机制(CBAM)和残差神经网络(ResNet)相融合的寿命预测模型——CBAM-ResNet模型。首先,使用连续小波变换(CWT)将轴承的一维振动信号转换为二维图像;然后,将其作为CBAM-ResNet模型的输入,使用深度残差网络在空间和通道两个维度上充分提取图像中蕴含的轴承退化信息;最后,模型将提取到的退化信息映射为轴承的剩余使用寿命。使用PRONOSTIA试验台IEEEPHM 2012数据集中的两种工况数据进行了试验,在工况1下对比分析了4种模型的预测效果。结果表明,相比ResNet、CNN、CBAM-CNN模型,本文CBAM-ResNet模型的均方根误差指标平均降低了38.9%,评分函数降低了51.2%,且本文模型具有较强的泛化性,能够提高轴承RUL预测的精度。

Aiming at the problem that the deviation of RUL value predicted by the previous bearing remaining useful life(RUL)prediction method is too large.In this paper,a life prediction model,CBAM-ResNet model,which combines convolutional block attention module(CBAM)and residual neural network(ResNet),was proposed for RUL prediction of bearings.Firstly,the continuous wavelet transform(CWT)was used to convert the one-dimensional vibration signal of the bearing into a two-dimensional image,which was then used as the input of the CBAM-ResNet model.The deep residual network was used to fully extract the bearing degradation information contained in the image in both spatial and channel dimensions.Finally,the model used to map the extracted degradation information to the remaining useful life of the bearing.The two working condition data in the IEEEPHM 2012 data set of the PRONOSTIA test bench were used for testing.The prediction effect of four models was compared and analyzed in working condition 1.The results show that compared with ResNet,CNN and CBAM-CNN models,the root mean square error index of CBAM-ResNet model is reduced by 38.9%on average,and the scoring function is reduced by 51.2%.The model proposed in this paper has strong generalization,and can improve the accuracy of bearing RUL prediction.