Bifurcated Remaining Useful Life Prediction: A Hybrid Approach for Realistic Uncertainty Characterization

arXiv:2605.31241 · cs.LG · Submitted 2026-05-29 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Bifurcated Remaining Useful Life Prediction".

Jane: The gist The study presents a novel hybrid prognostic framework for uncertainty-aware Remaining Useful Life (RUL) estimation in turbofan engines using the NASA C-MAPSS dataset.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're talking about the paper "Bifurcated Remaining Useful Life Prediction: A Hybrid Approach for Realistic Uncertainty Characterization". It sounds like they're trying to fix a big problem with how we predict engine life.

Jane: Right, it’s about making those predictions more realistic. They aren't just giving you one single number for when an engine will fail; they are giving you a range of possibilities, which is what uncertainty quantification is all about.

Lu: What’s interesting here is that they are specifically tackling the issue of rigid piecewise linear target assumptions that people often use in these kinds of RUL models. That rigid assumption can really limit how flexible the prediction gets when things actually start to degrade.

Meng: From a deployment standpoint, those rigid targets mean you're either getting a very simple model or you're forcing the data into a shape that might not match reality, which is risky when you need reliable maintenance schedules.

Lalam: So, the core idea seems to be creating something hybrid to get better results than just using one single prediction method. They are aiming for uncertainty-aware Remaining Useful Life estimation in turbofan engines using the NASA C-MAPSS dataset.

The paper's summary: Tom: Okay, so they introduce this novel hybrid prognostic framework. It uses a state-aware strategy to split the engine's life into healthy and degraded regimes. It uses an LSTM-based autoencoder to decide which regime the engine is in—healthy or degraded.

Jane: That sounds like a smart way to handle it because not all engines degrade at the same speed, right? By using that autoencoder trained only on nominal data, they monitor the reconstruction error as a proxy for how degraded the engine actually is.

Lu: Exactly. The paper says this autoencoder acts as a robust state classifier by strictly being trained on nominal data where the Remaining Useful Life is greater than one hundred fifty cycles <ref:2605.31241#pg1,as a robust state classifier>. This gives them a solid baseline for defining what 'healthy' looks like in terms of sensor patterns.

Meng: If it can correctly classify the state, then you can switch to two different estimation methods depending on that classification, which sounds much more adaptable than using one universal model for everything.

Lalam: And for the healthy regime, they use a Conditional Weibull Survival Analysis to estimate the Mean Residual Life. But once they hit the degraded regime, that’s where things get probabilistic with a Probabilistic Neural Network and Monte Carlo Dropout to capture both aleatoric and epistemic uncertainty.

The paper's improvements: Tom: One of the key methodological improvements they highlight is introducing a custom Asymmetric Negative Log-Likelihood loss function for that Probabilistic Neural Network. This loss function is designed to train the network to favor early predictions while simultaneously estimating predictive uncertainty, which helps avoid overestimation during critical operation.

Jane: That’s a big deal because in maintenance, you really want confidence when things are healthy and more caution when things start showing signs of trouble. The asymmetric loss function seems like it directly addresses that timing issue.

Lu: They also focus on modeling uncertainty by capturing both aleatoric and epistemic uncertainty, which is necessary for risk-aware decision-making in aviation maintenance. Aleatoric uncertainty is quantified using a Negative Log-Likelihood loss function for the Probabilistic Neural Networks, which penalizes the model for being confidently wrong.

Meng: And they also use Monte Carlo Dropout to capture epistemic uncertainty by sampling sub-networks through M forward passes, and they combine that with the aleatoric uncertainty to get a total predictive uncertainty. That sounds like a rigorous way to measure how much the AI actually trusts its own answer versus just being noisy data.

Lalam: The final prediction is formulated as a weighted average of these two strategies. They use the autoencoder’s degradation probabilities, which they call Pd, to set the weight for their Probabilistic LSTM prediction, and one minus Pd for the Conditional Weibull strategy.

Conclusion: Tom: So to wrap up this paper on "Bifurcated Remaining Useful Life Prediction: A Hybrid Approach for Realistic Uncertainty Characterization," it shows that combining deep learning with statistical survival analysis can yield highly realistic uncertainty bands, which is what they call the result of their weighted ensemble approach.

Jane: The main implication here is that we move away from just getting a point estimate and start getting a safety-first prediction that tells us exactly how much we should trust that number at any given time. It’s about providing physically consistent uncertainty bands, which means they widen when models conflict, especially near the transition point between healthy and degraded states.

Lu: It addresses that 'flat prediction' bias you see in data-driven models during the early stages of a component's life by incorporating Weibull survival analysis on those early steps. That is a specific contribution to improving the baseline model’s predictions while providing realistic uncertainty intervals where degradation signals are weak.

Meng: For us in production, this means if we deploy this, we get a tool that gives us better risk assessment because we can see when the uncertainty bands get wide and tell maintenance teams exactly when to be more cautious. It moves it from a simple prediction to a risk-informed tool for maintenance.

Lalam: The overall point of "Bifurcated Remaining Useful Life Prediction: A Hybrid Approach for Realistic Uncertainty Characterization" is to provide that robust framework for RUL prediction with integrated uncertainty quantification, validating it on the NASA C-MAPSS two thousand eight turbofan engine degradation dataset <ref:2605.31241#pg1,Bifurcated Remaining Useful Life Prediction: A Hybrid Approach for Realistic Uncertainty Characterization>.

Tom: Yeah, so a method that combines state classification from an autoencoder with different statistical survival models depending on the engine's health regime is a way to get better predictions than relying on just one deterministic baseline.

Jane: It’s a lot of moving parts, but it creates much more meaningful outputs for anyone who has to make decisions about engine maintenance. We have seen how this paper tackles uncertainty quantification in complex hybrid methods, and now we see how that applies directly to predicting engine life using real-world data.

Xabier Belaunzaran, Antonio Nappa, Arkaitz Artetxe, Basilio Sierra

Fundación Vicomtech · Basque Research and Technology Alliance (BRTA) · University of the Basque Country (UPV/EHU)

cs.LG

Submitted: 2026-05-29

Updated: 2026-05-29

Comments: Submitted to 9th European Conference of the Prognostics and Health Management Society 2026

Journal ref: Proceedings of the 9th European Conference of the Prognostics and Health Management Society (PHME 2026), Vol. 9, No. 1, pp. 1-12, Oslo, Norway, July 2026

DOI: 10.36001/phme.2026.v9i1.4903

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 77/100

The gist: The gist The study presents a novel hybrid prognostic framework for uncertainty-aware Remaining Useful Life (RUL) estimation in turbofan engines using the NASA C-MAPSS dataset.

Key concepts

State-Aware Strategy: This strategy divides the engine's life into two distinct operational regimes: 'healthy' and 'degraded.' An LSTM autoencoder classifies the current state by monitoring reconstruction error. This classification dictates which specific RUL estimation model (Weibull or Probabilistic NN) is applied, ensuring the prediction method matches the engine's actual condition.
Conditional Weibull Survival Analysis: Used for estimating RUL when an engine is in a 'healthy' state. This statistical method models the time until failure based on survival data, conditioned on the current healthy state determined by the autoencoder. It provides a rigorous way to estimate mean residual life for engines operating well.
Uncertainty Quantification (UQ): The framework captures two types of uncertainty: aleatoric (inherent randomness in the data) and epistemic (uncertainty due to model limitations). Aleatoric uncertainty is handled via a Negative Log-Likelihood loss, while epistemic uncertainty is captured using Monte Carlo Dropout to sample multiple predictions, resulting in robust, physically consistent confidence bands.
Weighted Ensemble Prediction: The final RUL prediction is a weighted combination of the results from the two different estimation strategies (Weibull and Probabilistic NN). The weights are dynamically determined by the autoencoder's output probabilities of being in a healthy or degraded state, leading to an ensemble mean that is more accurate than either model alone.

Terminology

Remaining Useful Life — 2× in this paper
Hybrid Approach — 1× in this paper · explained in 4 episodes
Monte Carlo dropout — 1× in this paper · explained in 2 episodes
Reconstruction Error — 1× in this paper · explained in 2 episodes

Summary

The gist The study presents a novel hybrid prognostic framework for uncertainty-aware Remaining Useful Life (RUL) estimation in turbofan engines using the NASA C-MAPSS dataset.

How it works

The framework employs a state-aware strategy that bifurcates the engine’s operational lifespan into ‘healthy’ and ‘degraded’ regimes, utilizing an LSTM-based autoencoder to act as a robust state classifier The autoencoder is trained strictly on nominal data (RUL > 150 cycles) to monitor reconstruction error for state classification For the healthy regime, a Conditional Weibull Survival Analysis is used for Mean Residual Life estimation For the degraded regime, a Probabilistic Neural Network with Monte Carlo Dropout captures both aleatoric and epistemic uncertainties Rather than using rigid binary labels, a calibrated sigmoid function converts the autoencoder’s output into continuous state probabilities to dynamically weight the final ensemble prediction

Key Components and Methodology

The proposed hybrid approach overcomes limitations of rigid Piecewise Linear (PwL) target assumptions by combining an autoencoder with a Probabilistic LSTM and a Conditional Weibull survival analysis The main contributions include introducing a custom Asymmetric Negative Log-Likelihood loss function to train the Probabilistic LSTM network, which quantifies predictive uncertainty while penalizing dangerous late predictions The methodology involves a multi-step approach: (i) an autoencoder-based classification into healthy and degraded regions, (ii) RUL estimation using Weibull Survival Analysis in the healthy region and Probabilistic LSTM network in the degraded one, and (iii) fusion of predictions and uncertainty sources through probability weights guided by the autoencoder’s labeling

Modeling Uncertainty

Uncertainty Quantification (UQ) is incorporated by capturing both aleatoric and epistemic uncertainty Aleatoric uncertainty is quantified using a Negative Log-Likelihood (NLL) loss function for Probabilistic Neural Networks, which penalizes the model for being confidently wrong Epistemic uncertainty is captured by Monte Carlo Dropout, where passing the input through the model M times samples an ensemble of sub-networks to quantify variance across predicted means The total predictive uncertainty is calculated as the sum of mean aleatoric uncertainty and epistemic uncertainty

Fusion and Prediction

The final RUL prediction is formulated as a weighted average of the statistical (Weibull) and predictive (Probabilistic LSTM) strategies, where weights are provided by the autoencoder’s degradation probabilities The weight for the Probabilistic LSTM prediction is defined as Pd, while the weight for the Conditional Weibull prediction is its complement, wh = 1 − Pd The final ensemble mean RUL, µ, is calculated as a standard linear combination of the individual model means The final ensemble variance, σ squared, is computed using the Law of Total Variance to account for both internal variance and dispersion between models' conflicting predictions

Results and Evaluation

The results demonstrate that the neural network’s inherent lack of predictive resolution in the early healthy zone is effectively compensated by the statistical rigor of the Conditional Weibull Survival Analysis The generated uncertainty bands accurately reflect physical realities, with uncertainty being rightfully large during early stages and narrowing as degradation becomes active The weighted ensemble approach yields a superior Proportional Index (N/P = 0.54) compared to the standalone LSTM model (N/P = 0.29), achieving a vastly superior and realistic coverage probability of 0.85 The overall point-estimate performance remains below that of deterministic baselines, but the methodology successfully provides uncertainty bounds that faithfully capture the underlying variability of the system

Future Work

Future work will focus on increasing the complexity and capability of individual modules, including rigorous hyperparameter optimization for autoencoder probability thresholds and asymmetric loss penalty factor The baseline survival analysis will be extended through the implementation of a dual-Weibull distribution model to independently characterize the time-to-degradation and time-to-failure phases Incorporating mechanisms for degradation-pattern classification could facilitate extension to scenarios involving multiple operational conditions and multiple failure modes The primary objective of this study is not to establish a new state-of-the-art in terms of absolute predictive error (RMSE), but rather to introduce a methodological framework that addresses a critical limitation in PHM: the ’flat prediction’ bias observed in data-driven models during the early stages of a component’s life By incorporating Weibull survival analysis on the early steps, the proposed hybrid approach improves upon the baseline model’s predictions while providing realistic uncertainty intervals that accurately reflect the lack of degradation signals at the start of the trajectory The research was partially funded by the Department of Industry of the Basque Government within the Hazitek program ZE-2025/00021 The paper was written by Xabier Belaunzaran, Antonio Nappa, Arkaitz Artetxe, and Basilio Sierra The research was partially funded by the Department of Industry of the Basque Government within the Hazitek program ZE-2025/00021 The paper is a preprint submitted to arXiv on June 1, 2026 The paper presents a novel hybrid prognostic framework for uncertainty-aware Remaining Useful Life (RUL) estimation in turbofan engines using the NASA C-MAPSS dataset This study addresses the gap by proposing a robust framework for RUL prediction with integrated UQ, addressing the limitation of rigid Piecewise Linear (PwL) target assumptions The approach is validated using the widely recognized NASA C-MAPSS 2008 turbofan engine degradation dataset The study introduces a novel hybrid approach that overcomes the limitations of the rigid Piecewise Linear (PwL) target assumptions commonly used in literature This design leverages the strong predictive performance of LSTM-based models while improving uncertainty estimation in early stages by incorporating survival-based predictions The final ensemble mean RUL, µ, is calculated as the expected value of this mixture, which simplifies to a standard linear combination of the individual model means This framework successfully demonstrates that combining deep learning with statistical survival analysis yields highly realistic and safety-critical uncertainty bands The primary objective of this study is not to establish a new state-of-the-art in terms of absolute predictive error (RMSE), but rather to introduce a methodological framework that addresses a critical limitation in PHM: the ’flat prediction’ bias observed in data-driven models during the early stages of a component’s life The paper is organized as follows: Section 2 reviews the state of the art in data-driven RUL prediction and Uncertainty Quantification Section 3 describes the C-MAPSS benchmark dataset and its specific challenge score function Section 4 presents the theoretical background of the key ideas Section 5 details the proposed hybrid methodology, including the autoencoder segmentation, the Conditional Weibull model, and the Probabilistic LSTM Section 6 presents the experimental results and evaluates the uncertainty metrics Finally, Section 7 concludes the paper and discusses future research directions The dataset was generated on 2008 using the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) tool Frederick et al.

Improvements for AI systems

textbfWeighting Ensemble Prediction via Degradation Probability (Pd): This fusion strategy allows for a mixture distribution where the weight for the Probabilistic LSTM prediction is wd = Pd, while the weight for the Conditional Weibull prediction is wh = 1 − Pd. This results in an ensemble mean RUL calculated as wh · µW eib + wd · µLSTM (10), which dynamically adapts to the engine's health state.
textbfAsymmetric Negative Log-Likelihood Loss Function: The custom loss function, defined as AsNLL = 1/2 log σ2 + C · (y − µ)2 / 2σ2, is optimized to train models that favor early predictions while simultaneously estimating predictive uncertainty, mitigating the risk of overestimation during critical operation.
textbfEpistemic Uncertainty Quantification via Monte Carlo Dropout: The system utilizes MC Dropout to capture model uncertainty, where the variance across the predicted means (µ) of these M forward passes directly quantifies the epistemic uncertainty. This allows engineers to assess model’s confidence in predictions.
textbfAdaptive Uncertainty Band Generation: The final ensemble variance calculation, σ2 = whσ2W eib + wdσ2LSTM + (wh(µW eib − µ)2 + wd(µLSTM − µ)2), ensures that the total uncertainty band widens when models conflict, providing a more conservative, safety-first uncertainty quantification right at the critical degradation transition point.
textbfDynamic State Gating via Autoencoder: The unsupervised LSTM autoencoder functions as a dynamic health-state classifier, where it transitions from a binary ‘healthy’ or ‘degraded’ classification to a continuous probability of degradation using the reconstruction MAE and Eq. 2, effectively separating modeling into distinct operational regimes.
textbfContextual Feature Augmentation: The Probabilistic LSTM incorporates the absolute operational cycle number of the engine at the time of prediction, which corresponds to the last cycle count of the input sliding window, providing additional temporal context that grounds the localized sensor patterns within the broader lifespan of the engine.
textbfPhysically Consistent Uncertainty Characterization: The framework is designed to generate physically consistent uncertainty bands, yielding high-confidence predictions near end-of-life while accurately reflecting the inherent variance of early operation, which provides a "robust tool for risk-informed maintenance.

Abstract

This study presents a novel hybrid prognostic framework for uncertainty-aware Remaining Useful Life (RUL) estimation in turbofan engines using the NASA C-MAPSS dataset. The framework employs a state-aware strategy that bifurcates the engines operational lifespan into "healthy" and "degraded" regimes. An LSTM-based autoencoder, trained strictly on nominal data (RUL > 150 cycles), monitors reconstruction error to act as a robust state classifier. For the healthy regime, a Conditional Weibull Survival Analysis is used for Mean Residual Life estimation. For the degraded regime, a Probabilistic Neural Network with Monte Carlo Dropout captures both aleatoric and epistemic uncertainties. Rather than using rigid binary labels, a calibrated sigmoid function converts the autoencoders output into continuous state probabilities, dynamically weighting the final ensemble prediction. The primary strength of this framework is its generation of physically consistent uncertainty bands, yielding high-confidence predictions near end-of-life while accurately reflecting the inherent variance of early operation, providing a robust tool for risk-informed maintenance.