Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Uncertainty-Aware Deep Learning Framework for Remaining Useful Life Prediction in Turbofan Engines with Learned Aleatoric Uncertainty

Krishang Sharma

TL;DR

This work tackles the challenge of predicting Remaining Useful Life (RUL) for turbofan engines while simultaneously quantifying uncertainty, a critical need for risk-aware maintenance. It introduces a hierarchical deep learning framework that combines multi-scale temporal convolutions, bidirectional recurrence, and dual-level attention, coupled with a Bayesian output layer that predicts both the mean RUL and aleatoric uncertainty. A comprehensive preprocessing pipeline—including condition-aware clustering, wavelet denoising, and feature selection—enables robust learning across multi-condition CMAPSS datasets. The model achieves competitive overall RMSE on FD001–FD004, while delivering breakthrough critical-zone accuracy (RUL ≤ 30) in the range of approximately 5–7 cycles and well-calibrated 95% confidence intervals with actual coverage between 93.5% and 95.2%, supporting risk-aware maintenance decisions. The approach balances accuracy, uncertainty, and computational efficiency (≈487K parameters, inference <10 ms on M1) and establishes new benchmarks for safety-critical prognostics in CMAPSS literature.

Abstract

Accurate Remaining Useful Life (RUL) prediction coupled with uncertainty quantification remains a critical challenge in aerospace prognostics. This research introduces a novel uncertainty-aware deep learning framework that learns aleatoric uncertainty directly through probabilistic modeling, an approach unexplored in existing CMAPSS-based literature. Our hierarchical architecture integrates multi-scale Inception blocks for temporal pattern extraction, bidirectional Long Short-Term Memory networks for sequential modeling, and a dual-level attention mechanism operating simultaneously on sensor and temporal dimensions. The innovation lies in the Bayesian output layer that predicts both mean RUL and variance, enabling the model to learn data-inherent uncertainty. Comprehensive preprocessing employs condition-aware clustering, wavelet denoising, and intelligent feature selection. Experimental validation on NASA CMAPSS benchmarks (FD001-FD004) demonstrates competitive overall performance with RMSE values of 16.22, 19.29, 16.84, and 19.98 respectively. Remarkably, our framework achieves breakthrough critical zone performance (RUL <= 30 cycles) with RMSE of 5.14, 6.89, 5.27, and 7.16, representing 25-40 percent improvements over conventional approaches and establishing new benchmarks for safety-critical predictions. The learned uncertainty provides well-calibrated 95 percent confidence intervals with coverage ranging from 93.5 percent to 95.2 percent, enabling risk-aware maintenance scheduling previously unattainable in CMAPSS literature.

Uncertainty-Aware Deep Learning Framework for Remaining Useful Life Prediction in Turbofan Engines with Learned Aleatoric Uncertainty

TL;DR

This work tackles the challenge of predicting Remaining Useful Life (RUL) for turbofan engines while simultaneously quantifying uncertainty, a critical need for risk-aware maintenance. It introduces a hierarchical deep learning framework that combines multi-scale temporal convolutions, bidirectional recurrence, and dual-level attention, coupled with a Bayesian output layer that predicts both the mean RUL and aleatoric uncertainty. A comprehensive preprocessing pipeline—including condition-aware clustering, wavelet denoising, and feature selection—enables robust learning across multi-condition CMAPSS datasets. The model achieves competitive overall RMSE on FD001–FD004, while delivering breakthrough critical-zone accuracy (RUL ≤ 30) in the range of approximately 5–7 cycles and well-calibrated 95% confidence intervals with actual coverage between 93.5% and 95.2%, supporting risk-aware maintenance decisions. The approach balances accuracy, uncertainty, and computational efficiency (≈487K parameters, inference <10 ms on M1) and establishes new benchmarks for safety-critical prognostics in CMAPSS literature.

Abstract

Accurate Remaining Useful Life (RUL) prediction coupled with uncertainty quantification remains a critical challenge in aerospace prognostics. This research introduces a novel uncertainty-aware deep learning framework that learns aleatoric uncertainty directly through probabilistic modeling, an approach unexplored in existing CMAPSS-based literature. Our hierarchical architecture integrates multi-scale Inception blocks for temporal pattern extraction, bidirectional Long Short-Term Memory networks for sequential modeling, and a dual-level attention mechanism operating simultaneously on sensor and temporal dimensions. The innovation lies in the Bayesian output layer that predicts both mean RUL and variance, enabling the model to learn data-inherent uncertainty. Comprehensive preprocessing employs condition-aware clustering, wavelet denoising, and intelligent feature selection. Experimental validation on NASA CMAPSS benchmarks (FD001-FD004) demonstrates competitive overall performance with RMSE values of 16.22, 19.29, 16.84, and 19.98 respectively. Remarkably, our framework achieves breakthrough critical zone performance (RUL <= 30 cycles) with RMSE of 5.14, 6.89, 5.27, and 7.16, representing 25-40 percent improvements over conventional approaches and establishing new benchmarks for safety-critical predictions. The learned uncertainty provides well-calibrated 95 percent confidence intervals with coverage ranging from 93.5 percent to 95.2 percent, enabling risk-aware maintenance scheduling previously unattainable in CMAPSS literature.

Paper Structure

This paper contains 61 sections, 13 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. Problem Statement
  4. Proposed Solution
  5. Key Contributions
  6. Organization
  7. Related Work
  8. Background and Dataset
  9. Existing Deep Learning Approaches
  10. Critical Research Gaps
  11. Our Approach Compared to Literature
  12. Proposed Methodology
  13. System Architecture Overview
  14. Data Preprocessing Pipeline
  15. RUL Calculation and Piecewise Linear Capping
  16. ...and 46 more sections

Figures (3)

  • Figure 1: Complete system architecture showing the three-stage pipeline: (1) Preprocessing with feature selection, wavelet denoising, and condition-aware normalization, (2) Hierarchical feature extraction through multi-scale CNN, bidirectional LSTM, and dual-level attention, (3) Probabilistic prediction with Bayesian output layer learning both mean RUL and uncertainty.
  • Figure 2: Model performance across standard CMAPSS datasets (FD001-FD004) showing RMSE, MAE, MAPE, and R² score metrics. Blue bars represent average performance across multiple runs, while orange bars show best-run results. Complex multi-condition datasets (FD002, FD004) exhibit higher error rates as expected, while single-condition datasets (FD001, FD003) demonstrate superior accuracy.
  • Figure 3: False positive and false negative analysis for maintenance decisions on FD001. Top-left: Confusion matrix at RUL $\leq$ 30 threshold showing excellent classification with only 1 false negative. Top-right: FPR and FNR across varying RUL thresholds, demonstrating optimal operating point around 30 cycles. Bottom-left: Precision-recall trade-off showing robust F1 scores (92%) across thresholds. Bottom-right: Error distribution revealing conservative model behavior with 65% under-predictions (safer) versus 35% over-predictions, ideal for safety-critical applications.