Remaining Useful Life Estimation for Turbofan Engines: A Comparative Study of Classical, CNN, and LSTM Approaches

Astitva Goel; Samarth Galchar; Sumit Kanu

arxiv: 2604.27234 · v1 · submitted 2026-04-29 · 💻 cs.LG

Remaining Useful Life Estimation for Turbofan Engines: A Comparative Study of Classical, CNN, and LSTM Approaches

Astitva Goel , Samarth Galchar , Sumit Kanu This is my paper

Pith reviewed 2026-05-07 07:59 UTC · model grok-4.3

classification 💻 cs.LG

keywords Remaining Useful LifeTurbofan EnginesLSTM1D CNNXGBoostRidge RegressionC-MAPSS DatasetPrognostics

0 comments

The pith

A single-layer LSTM outperforms prior deep LSTMs on turbofan engine remaining useful life estimation with RMSE of 14.93 on FD001.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper compares classical regression models, XGBoost, a 1D CNN, and an LSTM for estimating remaining useful life on the NASA C-MAPSS turbofan dataset. All approaches run through one shared preprocessing pipeline on the FD001 and FD003 subsets to enable direct comparison. The LSTM reaches RMSE of 14.93 on FD001 and 14.20 on FD003, beating the deeper LSTM from earlier work despite using only a single layer. XGBoost records 13.36 RMSE on FD003. The study shows how fixing the data steps lets researchers attribute performance gaps to the model choice itself.

Core claim

A single-layer LSTM applied to raw sensor sequences achieves RMSE of 14.93 on FD001 and 14.20 on FD003, outperforming the deep LSTM baseline of 16.14 and 16.18 reported by Zheng et al. under identical preprocessing and evaluation conditions; the 1D CNN reaches 16.97 and 15.68 on the same subsets, while XGBoost on engineered features reaches 13.36 on FD003.

What carries the argument

The single-layer LSTM network that ingests raw time-series sensor readings to output remaining useful life predictions.

If this is right

A simpler single-layer LSTM can exceed the accuracy of deeper LSTM variants on these engine datasets when data handling stays fixed.
XGBoost with engineered features delivers competitive or superior results on FD003 compared with the neural models.
The 1D CNN yields competitive accuracy on FD003 while producing more conservative RUL estimates on FD001.
Classical models such as ridge regression remain useful when supplied with engineered features and serve as transparent baselines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Standardizing the preprocessing step across studies could clarify whether model architecture or data preparation drives most gains in prognostics tasks.
The results point toward testing whether the same lightweight LSTM pattern transfers to RUL prediction on other rotating machinery or sensor streams.
If the performance edge persists under real-world noise and missing data, maintenance systems could adopt simpler recurrent models to lower computational overhead during deployment.

Load-bearing premise

That the shared preprocessing pipeline and evaluation protocol produce a truly fair comparison so performance gaps reflect model choice rather than hidden differences in hyperparameter search or data handling.

What would settle it

Reproducing the experiments with separate hyperparameter tuning for each model family and observing that the LSTM no longer leads or that the ordering among models reverses.

read the original abstract

Remaining Useful Life (RUL) estimation is a critical component of Prognostics and Health Management (PHM), enabling proactive maintenance scheduling and reducing unplanned failures in industrial equipment. This paper presents a comparative study of machine learning approaches for RUL estimation on the NASA C-MAPSS turbofan engine dataset: classical baselines (Ridge Regression, Polynomial Ridge, and XGBoost), a 1D Convolutional Neural Network (CNN), and a Long Short-Term Memory (LSTM) network. All models are evaluated on the FD001 and FD003 subsets under an identical preprocessing pipeline to ensure a fair comparison. Among raw-sequence models, the LSTM achieves RMSE of 14.93 and 14.20 on FD001 and FD003 respectively, outperforming the deep LSTM reported by Zheng et al.~\cite{paper} (RMSE 16.14 and 16.18) despite using a simpler single-layer architecture. The 1D CNN achieves RMSE of 16.97 on FD001 and 15.68 on FD003, demonstrating competitive performance on FD003 while producing more conservative RUL predictions on FD001. Ridge Regression is evaluated on raw and engineered features, while other classical models use only engineered inputs. XGBoost achieves an RMSE of 13.36 on FD003, highlighting the competitiveness of nonlinear modeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper supplies fresh RMSE numbers for LSTM and CNN on C-MAPSS FD001/FD003 and claims a win over Zheng et al., but the comparison hinges on an unverified preprocessing match.

read the letter

The headline result is a single-layer LSTM beating a deeper LSTM from Zheng et al. on the C-MAPSS FD001 and FD003 subsets, with RMSEs of 14.93 and 14.20. The paper also shows XGBoost competitive on FD003 and gives a clean internal comparison across classical and deep models. They do the useful thing of running everything through the same preprocessing steps, which lets you trust the relative ordering of their own models. Including Ridge and XGBoost on engineered features provides a practical baseline that many papers skip. The main gap is the cross-paper comparison. Nothing in the work confirms that window size, feature scaling, or RUL clipping matches Zheng et al. exactly. Small differences in those steps routinely change RMSE by 1-3 points on this data, so the outperformance cannot be pinned on the model choice yet. The abstract also leaves out hyperparameter search details and any measure of variability, which makes the numbers harder to interpret. This is the kind of paper that belongs in a reading group focused on applied PHM or predictive maintenance. Someone working on turbofan RUL will find the numbers handy for quick reference. It is not breaking new ground methodologically. I would recommend sending it for peer review. The core empirical claim is worth checking, and a referee can ask for the missing protocol details and statistical support. Once those are in, it becomes a solid benchmark update.

Referee Report

2 major / 1 minor

Summary. The paper conducts a comparative study of RUL estimation methods on the NASA C-MAPSS turbofan dataset (FD001 and FD003 subsets), evaluating classical models (Ridge Regression, Polynomial Ridge, XGBoost), a 1D CNN, and a single-layer LSTM under a shared preprocessing pipeline. It reports that the LSTM achieves RMSE of 14.93 on FD001 and 14.20 on FD003, outperforming Zheng et al.'s deep LSTM (16.14 and 16.18), while XGBoost reaches 13.36 on FD003 and the CNN shows competitive results on FD003.

Significance. If the experimental protocols are verified to match the cited prior work, the results would usefully demonstrate that a simpler single-layer LSTM can outperform deeper variants on a standard PHM benchmark, with implications for efficient model selection in industrial prognostics. The shared-pipeline comparison across classical and deep models is a strength, and the XGBoost result on FD003 adds a practical data point. However, the current significance is limited by the unverifiable nature of the cross-paper comparison.

major comments (2)

[Abstract] Abstract (and Results section): The central outperformance claim (single-layer LSTM RMSE 14.93/14.20 vs. Zheng et al. deep LSTM at 16.14/16.18 on FD001/FD003) is load-bearing, yet no section, table, or appendix verifies that sequence windowing, normalization, RUL target clipping, or train/validation splits replicate the protocol in Zheng et al. (2017). On C-MAPSS, such differences routinely shift RMSE by 1–3 points, preventing attribution of gains to architecture rather than implementation details.
[Experimental Setup] Experimental Setup / Results: The reported RMSE values are presented without details on hyperparameter tuning procedures, cross-validation strategy, number of independent runs, or statistical significance tests. Absence of error bars or variance estimates makes it impossible to determine whether observed differences (e.g., LSTM vs. CNN or vs. classical baselines) are robust.

minor comments (1)

[Methods] The distinction between 'raw-sequence models' and models using engineered features is mentioned in the abstract but would benefit from an explicit table or paragraph in the methods section clarifying feature usage for each model class.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's thorough review and constructive suggestions. We have revised the manuscript to enhance the transparency of our experimental setup and to provide better context for the comparisons with prior work. Below we address each major comment in detail.

read point-by-point responses

Referee: [Abstract] Abstract (and Results section): The central outperformance claim (single-layer LSTM RMSE 14.93/14.20 vs. Zheng et al. deep LSTM at 16.14/16.18 on FD001/FD003) is load-bearing, yet no section, table, or appendix verifies that sequence windowing, normalization, RUL target clipping, or train/validation splits replicate the protocol in Zheng et al. (2017). On C-MAPSS, such differences routinely shift RMSE by 1–3 points, preventing attribution of gains to architecture rather than implementation details.

Authors: We fully agree that the manuscript should provide explicit details to allow verification of the experimental protocol used for the comparison with Zheng et al. (2017). In the revised manuscript, we have added a comprehensive description of our preprocessing pipeline in the Experimental Setup section, including the sequence window length and stride, the method of normalization, the RUL target clipping threshold, and the train/validation split strategy. Additionally, we have included a table that compares our preprocessing choices with those reported in Zheng et al. (2017) and other relevant literature. This addition makes the cross-paper comparison verifiable and allows readers to assess whether the performance gains can be attributed to the model architecture. We have also moderated the language in the abstract and results to reflect that the outperformance is observed under the described standard preprocessing pipeline. revision: yes
Referee: [Experimental Setup] Experimental Setup / Results: The reported RMSE values are presented without details on hyperparameter tuning procedures, cross-validation strategy, number of independent runs, or statistical significance tests. Absence of error bars or variance estimates makes it impossible to determine whether observed differences (e.g., LSTM vs. CNN or vs. classical baselines) are robust.

Authors: We acknowledge this limitation in the original submission. The revised manuscript now includes detailed information on our experimental procedures: hyperparameter tuning was performed using grid search on a validation set (specific ranges and selected values are provided in an appendix), we employed a standard fixed split for training and validation as is common in C-MAPSS benchmarks, and each experiment was repeated over 5 independent runs with different random initializations to compute mean and standard deviation. We have added error bars to the reported RMSE values in the results tables and included a discussion of statistical significance using t-tests between model pairs. These changes allow for a better assessment of the robustness of the differences observed between the LSTM, CNN, and classical models. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical model comparisons on public benchmark

full rationale

The paper reports experimental RMSE results from training classical regressors, a 1D CNN, and an LSTM on the NASA C-MAPSS FD001/FD003 subsets. All performance numbers are direct outputs of model inference against held-out test labels after a shared preprocessing pipeline; no equation or quantity is defined in terms of another quantity that is itself fitted or predicted from the same pipeline. The cited comparison to Zheng et al. is an external reference to previously published numbers on the same public dataset, not a self-citation whose validity is presupposed by the present work. No uniqueness theorem, ansatz, or renaming of a known result is invoked. The derivation chain therefore consists solely of standard supervised learning steps whose outputs are falsifiable against external data and do not reduce to the inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 1 axioms · 0 invented entities

The performance claims rest on empirical evaluation of tuned models on a fixed public benchmark. Hyperparameters for each model (layer sizes, learning rates, tree depths) are fitted to the data. The work assumes the C-MAPSS subsets are representative of real engine degradation and that identical preprocessing removes confounding factors. No new physical entities or ad-hoc theoretical constructs are introduced.

free parameters (3)

LSTM hyperparameters (units, dropout, sequence length)
Specific architecture and training choices determine the reported 14.93 and 14.20 RMSE values.
XGBoost hyperparameters (max depth, learning rate, n_estimators)
Tuned to achieve the 13.36 RMSE on FD003.
Preprocessing parameters (window size, feature scaling)
Applied identically but still chosen to enable the reported results.

axioms (1)

domain assumption The NASA C-MAPSS FD001 and FD003 subsets constitute appropriate and representative benchmarks for turbofan RUL estimation
The comparison relies on this standard dataset without questioning its fidelity to real-world engine behavior.

pith-pipeline@v0.9.0 · 5545 in / 1573 out tokens · 52686 ms · 2026-05-07T07:59:40.368906+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

5 extracted references · 5 canonical work pages

[1]

Long Short-Term Memory Network for Remaining Useful Life Estimation,

S. Zheng, K. Ristovski, A. Farahat, and C. Gupta, “Long Short-Term Memory Network for Remaining Useful Life Estimation,” inProc. IEEE Int. Conf. Prognostics and Health Management (ICPHM), 2017, pp. 88– 95

work page 2017
[2]

Turbofan Engine Degradation Simulation Data Set,

A. Saxena and K. Goebel, “Turbofan Engine Degradation Simulation Data Set,” NASA Ames Research Center, Moffett Field, CA, USA, Tech. Rep., 2008

work page 2008
[3]

Long Short-Term Memory,

S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,”Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

work page 1997
[4]

Learning Long-Term Dependen- cies with Gradient Descent is Difficult,

Y . Bengio, P. Simard, and P. Frasconi, “Learning Long-Term Dependen- cies with Gradient Descent is Difficult,”IEEE Trans. Neural Networks, vol. 5, no. 2, pp. 157–166, 1994

work page 1994
[5]

Deep Convolutional Neural Network Based Regression Approach for Estimation of Remaining Useful Life,

G. S. Babu, P. Zhao, and X.-L. Li, “Deep Convolutional Neural Network Based Regression Approach for Estimation of Remaining Useful Life,” in Proc. Int. Conf. Database Systems for Advanced Applications, Springer, 2016, pp. 214–228

work page 2016

[1] [1]

Long Short-Term Memory Network for Remaining Useful Life Estimation,

S. Zheng, K. Ristovski, A. Farahat, and C. Gupta, “Long Short-Term Memory Network for Remaining Useful Life Estimation,” inProc. IEEE Int. Conf. Prognostics and Health Management (ICPHM), 2017, pp. 88– 95

work page 2017

[2] [2]

Turbofan Engine Degradation Simulation Data Set,

A. Saxena and K. Goebel, “Turbofan Engine Degradation Simulation Data Set,” NASA Ames Research Center, Moffett Field, CA, USA, Tech. Rep., 2008

work page 2008

[3] [3]

Long Short-Term Memory,

S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,”Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

work page 1997

[4] [4]

Learning Long-Term Dependen- cies with Gradient Descent is Difficult,

Y . Bengio, P. Simard, and P. Frasconi, “Learning Long-Term Dependen- cies with Gradient Descent is Difficult,”IEEE Trans. Neural Networks, vol. 5, no. 2, pp. 157–166, 1994

work page 1994

[5] [5]

Deep Convolutional Neural Network Based Regression Approach for Estimation of Remaining Useful Life,

G. S. Babu, P. Zhao, and X.-L. Li, “Deep Convolutional Neural Network Based Regression Approach for Estimation of Remaining Useful Life,” in Proc. Int. Conf. Database Systems for Advanced Applications, Springer, 2016, pp. 214–228

work page 2016