点(几何)
计算机科学
汽车工程
工程类
噪音(视频)
钥匙(锁)
变更检测
环境科学
信号处理
工作(物理)
作者
Yanhang Geng,Jianhai Yan,Biao Lu
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-08-01
卷期号:: 407-413
标识
DOI:10.1109/srse70977.2026.11694706
摘要
Remaining Useful Life (RUL) prediction of aeroengines is a critical problem in Prognostics and Health Management (PHM). Traditional change point detection methods rely on a unified baseline across all engines, which often leads to high misjudgment rates due to the heterogeneity of initial health states among individual engines. To address this issue, this paper proposes a two-stage RUL prediction method based on Generative Adversarial Network (GAN)-driven change point detection and stage-specific Long Short-Term Memory (LSTM) networks. The GAN is first trained in an unsupervised manner to learn the highdimensional manifold distribution of healthy operating data. Degradation change points are then identified by monitoring shifts in the generator reconstruction error, which serves as an anomaly indicator when the data distribution deviates from the learned healthy pattern. Using the detected change point as a segmentation boundary, the engine life cycle is partitioned into a healthy stage and a degradation stage, with a dedicated LSTM model trained for each stage to capture stage-specific degradation patterns. Experiments on the C-MAPSS FD002 dataset demonstrate that the proposed method reduces the prediction RMSE from 24.49 to 17.94 and the PHM08 Score from 4450.00 to 2301.68, significantly outperforming the baseline single-LSTM approach. Ablation studies further reveal that the RUL labeling strategy and the change point detection threshold are key factors governing prediction accuracy, with the 95th-percentile threshold yielding the best overall performance.
科研通智能强力驱动
Strongly Powered by AbleSci AI