Remaining Useful Life Prediction of Wind Turbine Rotating Machinery Based on Dilated Inception Network

In the wind power sector, remaining useful life (RUL) predictions guarantee the reliable and safe operation of rotating machinery. Traditional deep learning models have a complex and fixed structure, which makes it difficult to extract effective information features from bearing degradation patterns. In this paper, a novel RUL prediction method named dilated inception network (DiIncepN et) is proposed. In the proposed DiIncepNet, signals reconstructed by the denoising autoencoder (DAE) are used as input. Dilated convolution and adaptive pooling operations are introduced in Inception v2 to build the DiIncep module. This improvement increases the receptive field based on capturing features at multiple scales and eliminates the need to manually adjust pooling parameters. The stacked DiIncep modules subsequently integrate (bidirectional gated recurrent unit) BiGRU and Dropout layers. This facilitates the network to capture temporal dependencies and destroys weight co-adaptation. Experimental results on the XJTU-SY datasets show that the network is able to accurately predict the RUL with superior generalization ability and robustness.