基于可变形卷积的轴承剩余寿命预测

Remaining useful life prediction of bearings based on deformable convolution

针对在滚动轴承剩余寿命(RUL)预测任务中神经网络的普通卷积核提取到的特征分布不均问题,建立了基于注意力的深度可变卷积残差网络(ADRN)以提取轴承的退化特征并计算健康因子(HI)。通过连续小波变换(CWT)提取轴承的时频特征,采用ADRN提取轴承时频图中的退化特征,并通过Tanh激活函数得到HI。为提升对异常值的约束能力,在整个网络中采用提出的动态损失函数进行训练。使用Savitzky-golay滤波器平滑HI后,由多项式函数拟合HI得到回归方程,预测出轴承的RUL。在PHM2012数据集上的实验仿真证明,提出的方法得到了更准确的预测结果。

To address the problem of uneven distribution of features extracted from the standard convolutional kernel in the neural network for the prediction task of the Remaining Useful Life(RUL)of rolling bearings,this paper establishes an Attention-based deep Deformable convolutional Residual Network(ADRN)to extract the degradation features of the bearing and calculate the Health Indicator(HI).The time-frequency features of the bearing are extracted by Continuous Wavelet Transform(CWT).The degenerate features of the time-frequency map of the bearing are extracted by ADRN and the HI is computed by Tanh activation function.To improve the constraint ability for abnormal value,the dynamic loss function proposed in this paper is used in the training process for overall network.The HI is smoothed by Savitzky-golay filter,and the regression equation is obtained by fitting the HI with polynomial function to predict the RUL of the bearing.The experimental simulations on the PHM2012 data set prove that the proposed method obtains more accurate prediction results than other methods.