Remaining useful life prediction of bearings based on deformable convolution
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 De-formable 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.
remaining useful life prediction of rolling bearingsdeformable convolutionat-tention mechanismdynamic loss functioncontinuous wavelet transform
NETL
NSTL
万方数据
维普
