Deep Learning-Based Residual Useful Lifetime Prediction for Assets with Uncertain Failure Modes

TL;DR

This work tackles residual life prediction for assets with multiple, uncertain failure modes by integrating mixture $(\mathrm{Log})$-Location-Scale $(\mathrm{LLS})$ distributions into a deep learning prognostic framework. It introduces two architectures, DLBP1 (asset-specific location and scale) and DLBP2 (shared scale across assets), which map degraded signals to the parameters of a mixture $LLS$ and then compute RUL via the mixture mean. Using SWM to handle variable-length degradation signals and an LSTM-based temporal encoder, the models are trained with a negative log-likelihood loss and evaluated on the NASA C-MAPSS FD003 dataset against 11 baselines, consistently achieving superior PS scores and competitive RMSE/RAE. The results demonstrate that incorporating domain knowledge about failure-time distributions into deep learning improves robustness to overlapping signals and unlabeled data; future work could explore self-attention to capture more complex dependencies.

Abstract

Industrial prognostics focuses on utilizing degradation signals to forecast and continually update the residual useful life of complex engineering systems. However, existing prognostic models for systems with multiple failure modes face several challenges in real-world applications, including overlapping degradation signals from multiple components, the presence of unlabeled historical data, and the similarity of signals across different failure modes. To tackle these issues, this research introduces two prognostic models that integrate the mixture (log)-location-scale distribution with deep learning. This integration facilitates the modeling of overlapping degradation signals, eliminates the need for explicit failure mode identification, and utilizes deep learning to capture complex nonlinear relationships between degradation signals and residual useful lifetimes. Numerical studies validate the superior performance of these proposed models compared to existing methods.

Figures (5)

• Figure 1: A demonstration of the SWM. The time-series data is from sensor $\#4$ of engine $\#27$ in the FD003 training dataset, whose real RUL is $320$. The sliding window width is $T_w$. The solid rectangle in red and the dashed rectangle in green represent two truncated signals, and their corresponding RULs are $\tilde{y}_1$ and $\tilde{y}_{50}$, respectively (more details about the dataset will be provided in Section 3).
• Figure 2: The LSTM recurrent cell diagram, where $\sigma$ means applying sigmoid activation function to the output of the input gate, the forget gate, as well as the output gate, and tanh represents the activation function applied to the output of $\mathbf{G}_t$.
• Figure 3: A comparison of RMSE and PS with $x-$axis be the difference between the predicted response and the true value, and $y-$axis be the values of RMSE (in solid red) and PS (in dashed blue).
• Figure 4: The estimated RUL values (orange circles) and actual RUL values (green triangles) for 100 test engines using the proposed methods are shown. The left three subplots represent Model DLBP1, and the right three subplots represent Model DLBP2, under three mixture distributions: mixture log-normal (top row), mixture Weibull (middle row), and mixture log-logistic (bottom row).
• Figure 5: The boxplot of RAE for the proposed methods (left: model DLBP1; right: model DLBP2) under three mixture distributions (from left to right in each subplot: Mixture Log-Normal, Mixture Weibull, Mixture Log-Logistic).