A denoising semi-supervised deep learning model for remaining useful life prediction of turbofan engine degradation

Remaining useful life (RUL) prediction is significant for reliability analysis and the reduction of maintenance costs for turbofan engine systems. However, most of the existing methods capture temporal or spatial features to predict RUL, which leads to the neglect of deep spatio-temporal correlation in the prediction data. In addition, there are usually noise interference and few labels in feature extraction, which causes difficulty in designing robust RUL prediction models. To address these problems, a denoised semi-supervised model based on fully convolutional denoising autoencoder, convolutional neural network, and long short-term memory network (FCDAE-CNN-LSTM) is proposed to predict RUL, where FCDAE is constructed to denoise the original data by parameter fine-tuning and CNN-LSTM is established based on parallel connection for RUL prognosis. The denoising and generalization capabilities of the RUL prediction model are enhanced by the combination of unsupervised denoising and supervised feature extraction in the case of few labeled training data. The superior features of the method lie in that FCDAE effectively captures globalized and localized features and CNN-LSTM captures multi-layer fused spatial-temporal correlations. The Root Mean Square Error (RMSE) and Score of the proposed method on C-MAPSS dataset are 12.01, 16.62, 11.84, 18.21, and 209.44, 1466.03, 205.07, 2338.93, respectively, which have demonstrated that our method achieves the state-of-the-art performance and outperforms other models.