Remaining useful life prediction method based on gated dilation causal convolution

Abstract

Time series modeling is key to predicting the remaining useful life (RUL) of operating equipment. However, the design of traditional convolutional neural networks cannot flexibly adapt to various time window sizes, and when dealing with long sequences, it needs to have a corresponding deep structure, which is prone to the problem of gradient disappearing. These defects make traditional convolutional neural networks have high prediction errors in mechanical RUL prediction, so this paper proposed a new gated dilation causal convolutional (GDCC) network with multi-feature squeeze excitation unit (MSEU) and sinusoidal position coding, and used it for RUL prediction of rolling bearings. Specifically, MSEU adaptively corrects feature responses from both local and global perspectives; Second, sinusoidal position encoding adds a global order relationship to the hidden layer features; Then, build and stack GDCC for further learning of the front layer features; Finally, the RUL of the rolling bearings is predicted by the Relu activation function. The proposed method was evaluated using vibration data from the accelerated degradation test of rolling bearings and the performance of this method was compared with several advanced predictive methods. Experimental results show that the proposed method is effective in improving the prediction accuracy of bearings RUL, and is better than some existing data-driven prediction methods.