A method for uncertainty quantification of remaining useful life in tool wear based on an improved deep residual bidirectional long short-term memory network

Abstract Predictive techniques based on bidirectional LSTM (BiLSTM) networks are essential for improving the reliability of industrial systems, enabling accurate degradation modeling, proactive maintenance scheduling, and mitigation of unexpected failures in high-risk operating environments. However, the training of deep BiLSTM networks often encounters problems such as gradient vanishing, model degradation, and uncertainty. To address these challenges, this paper proposes an improved deep residual BiLSTM model (iDRBiLSTM), which introduces a residual structure to enhance the generalization ability and convergence speed of the degraded model. In order to manage the prediction error, the iDRBiLSTM- Gaussian process (GP) regression degradation model is constructed in this study using GP theory to fit the residuals between the predicted and actual values. This approach improves the reliability of the prediction intervals and further quantifies the uncertainty of the prediction results. Experiments conducted on the PHM2010 tool wear dataset and the automatic tool changer (ATC) system cutting force dataset validate that the proposed method outperforms existing popular methods in terms of accuracy and reliability.