Scholarlink

Predicting the Remaining Useful Life (RUL) of turbofan engines is critical but often constrained by the scarcity of run-to-failure labels. We study the data efficiency of Active Learning (AL) for RUL prediction on NASA’s CMAPSS FD001, using a GRU regressor and three acquisition strategies—uncertainty sampling, hybrid uncertainty+diversity, and random sampling. Compared to a baseline trained on the full labeled set, uncertainty sampling reaches comparable accuracy using about 20% of the labeled sequences, while hybrid and random sampling achieve similar final accuracy at higher label budgets. We analyze learning-curve behavior, discuss distributional effects of RUL capping, and provide confidence intervals over multiple random seeds. Our results highlight conditions under which AL can reduce the number of labeled sequences needed for accurate RUL modeling, and we release code to facilitate replication.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Active Learning for Cheap RUL Prediction in CMAPSS Dataset

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Esmaail Albarazi

摘要

Predicting the Remaining Useful Life (RUL) of turbofan engines is critical but often constrained by the scarcity of run-to-failure labels. We study the data efficiency of Active Learning (AL) for RUL prediction on NASA’s CMAPSS FD001, using a GRU regressor and three acquisition strategies—uncertainty sampling, hybrid uncertainty+diversity, and random sampling. Compared to a baseline trained on the full labeled set, uncertainty sampling reaches comparable accuracy using about 20% of the labeled sequences, while hybrid and random sampling achieve similar final accuracy at higher label budgets. We analyze learning-curve behavior, discuss distributional effects of RUL capping, and provide confidence intervals over multiple random seeds. Our results highlight conditions under which AL can reduce the number of labeled sequences needed for accurate RUL modeling, and we release code to facilitate replication.