Abstract
Remaining Useful Life (RUL) prediction is a cornerstone of Prognostic and Health Management (PHM). Traditional deep learning models, such as LSTMs, often rely on extensive historical sequences that are frequently unavailable in real-world industrial settings. This paper introduces BACE-RUL, a Bi-directional Adversarial network with Covariate Encoding. By mapping current sensor snapshots into a high-dimensional conditional space and utilizing a bi-directional generative framework, BACE-RUL achieves state-of-the-art performance on aircraft engine and battery datasets without the need for long, consecutive historical records.
Introduction
Predicting when a machine will fail is essential for optimizing maintenance schedules and preventing catastrophic failures. However, current data-driven methods face three major hurdles: a heavy dependency on long historical data sequences, a reliance on domain expertise for physics-based modeling, and rigid statistical assumptions that fail to capture messy, multi-mode degradation.
BACE-RUL addresses these challenges by demonstrating that a deep encoding of the current sensor snapshot contains sufficient information to accurately predict RUL. This approach makes the model highly practical for "brownfield" industrial deployments where historical logs may be incomplete or unavailable.
Methodology
The BACE-RUL architecture is divided into two primary processes: Condition Encoding (CE) and RUL Prediction (RP).
- Condition Encoding (CE): An autoencoder ($E_1, G_1$) and a discriminator ($D_1$) map raw sensor measurements into a high-dimensional "conditional space." This adversarial setup extracts the most informative features of a machine's status without manual feature engineering.
- RUL Prediction (RP): Designed as a conditional generative network, this part takes encoded conditions and noise to generate RUL estimates. A second encoder ($E_2$) maps ground-truth RUL values into a latent space, creating a bi-directional training loop that stabilizes the generation process.
- Asymmetric Loss Function: The model employs a "distortion loss" that penalizes underestimating RUL during early healthy stages while prioritizing high accuracy as the machine nears its end-of-life.
Experimental Setup
The model was evaluated across three major benchmarks representing different mechanical and chemical degradation processes.
- C-MAPSS (NASA): Turbofan engine data with various operating conditions (FD001–FD004).
- NASA Li-Ion Battery Dataset: Experimental degradation data from battery cells.
- Toyota Research Institute Battery Dataset: Large-scale measurements from 124 commercial lithium-ion phosphate (LFP) cells.
Key Results and Insights
BACE-RUL significantly outperformed standard baselines, including LSTMs and Deep CNNs:
- Superior Accuracy: On the C-MAPSS FD001 dataset, BACE-RUL achieved an RMSE of 21.81, compared to 31.41 for LSTM and 32.76 for DCNN.
- Safety-First Performance: In the PHM08 "Score" metric—which heavily penalizes late predictions—BACE-RUL outperformed competitors by a factor of two or more on engine datasets.
- Generalization: The framework proved stable across both mechanical (engines) and chemical (batteries) systems, highlighting its versatility.
Limitations and Future Work
While BACE-RUL offers a significant leap in practicality, it currently relies on an empirically set "early constant value" for RUL during the healthy phase. Future research will focus on integrating automated "change point detection" to identify the transition from healthy to degrading states automatically. Additionally, while the model is highly accurate, its bi-directional architecture is computationally heavier to train than simple regression models, though this is offset by superior inference performance.
Summary updated