Rolling bearing remaining useful life prediction based on dilated causal convolutional DenseNet and an exponential model

Accurate prediction of the remaining useful life (RUL) of rolling bearings is important to ensure the safe operation of mechanical system and to formulate maintenance strategies. Accordingly, this paper proposes a RUL prediction method combining a dilated causal convolutional DenseNet (DCCDenseNet) and an exponential model, and the keys to its implementation include the construction of health indicators that can characterize performance degradation and the determination of first prediction time (FPT). First, continuous wavelet transform is applied to transform time-domain signals into time-frequency images as the input of the network model. Second, a DCCDenseNet model is constructed by introducing the dilated causal convolution into a DenseNet model, which is used to construct health indicators from time-frequency images. It not only inherits the advantages of the DenseNet model, but can also capture the temporal patterns of time series. Finally, an indicator is constructed by using four time-domain statistical indicators to determine the FPT, and it is combined with an exponential model to fit the obtained health indicators to predict RUL. Two rolling bearing datasets are applied to analyze the availability of the suggested method, and its good performance in the RUL prediction is demonstrated by comparing with the results of other methods.