Predicting when a jet engine will fail.

Predictive maintenance (PdM) is essential in aerospace systems where failures incur high safety and financial risks. The transition from reactive or preventive strategies to condition-based, data-driven maintenance enables improved reliability and reduced downtime. Central to this shift is the estimation of RUL, which empowers timely and cost-efficient maintenance planning.

Recent frameworks have explored hybrid and data-driven approaches to enhance prognostic accuracy. Berri et al. proposed a hybrid method combining a dynamical state estimator with a Support Vector Machine (SVM) classifier to monitor system health in real time, achieving ±20% error margins near end-of-life scenarios. This highlights the benefits of combining physical models with machine learning to enhance generalisability across different engines and fault modes [1].

Machine learning models such as Multilayer Perceptrons (MLPs) have been shown to effectively capture nonlinear degradation trends in engine data. Caricato et al. applied an MLP to NASA's C-MAPSS dataset, achieving RMSE and MAE of 17.38 and 12.50 respectively, validating its potential for fleet-level predictive frameworks [2].

More recent efforts have emphasised uncertainty quantification. Lee and Mitici introduced a CNN architecture with Monte Carlo dropout, generating probabilistic RUL estimates that led to a 95.6% reduction in unscheduled maintenance and a 29.3% decrease in overall cost [3]. Chen et al. integrated RUL prediction with maintenance cost modelling, optimising inspection intervals using a Maintenance Cost Rate (MCR) framework [4]. Pater and Mitici applied Sequential Monte Carlo (SMC) methods to track system health under uncertainty [5].

Uncertainty modelling remains a central challenge in AI-driven prognostics. Rastogi [6] and Xuan et al. [7] distinguish between aleatoric (data) and epistemic (model) uncertainty, emphasising the need for confidence intervals in RUL estimates to support safe, risk-aware decisions.

Traditional ML models such as linear regression, logistic regression, Random Forests (RF) and SVMs have played foundational roles in PdM due to their low computational complexity, interpretability and ease of deployment [8], with applications across aircraft systems, semiconductor fabrication and industrial machinery.

Le et al. showed that linear regression paired with t-statistics for feature selection could maintain accuracy while reducing input dimensionality from 27 to 6 in semiconductor maintenance [9]. In aerospace, Maulana et al. proposed an explainable PdM framework using logistic regression and an Unscented Kalman Filter (UKF), achieving RMSE improvements of 34.5–55.6% over previous methods on C-MAPSS [10].

Ensemble tree methods, especially Random Forests, are resilient to noise and high-dimensional data and provide interpretable feature importance, so they're often used for benchmarking or to rank features before deep learning stages [10]. SVMs, including SVR variants, are favoured on medium-sized datasets where kernel methods uncover nonlinear degradation trends [11]. Classical techniques also support hybrid architectures — Asif et al. used moving median filters to refine input sequences before deep learning [13].

However, these models are generally limited in capturing the long-term temporal dependencies inherent in RUL prediction, and require manual feature engineering such as lags or rolling averages, reducing adaptability to complex operating conditions.

High-frequency, multi-sensor time-series data has driven the adoption of deep learning for RUL prediction: DL models learn temporal and spatial patterns directly from raw sensor inputs, removing the need for manual feature engineering.

LSTM networks capture long-term dependencies and sequential degradation trends; Asif et al. and Wu et al. showed LSTM-based models outperform traditional approaches [12][13]. CNNs, treating time-series as structured arrays, extract local features such as gradients and short-term changes, improving early-warning capability [14][16].

Hybrids combine these strengths. Al-Dulaimi et al.'s Hybrid Deep Neural Network (LSTM–CNN) outperformed standalone models on turbofan engines under nonlinear degradation and variable conditions [14], and their CNN–Bidirectional LSTM (NBLSTM) improved stability under real-world sensor noise [15]. Zraibi et al. proposed a CNN–LSTM–DNN hybrid for lithium-ion battery RUL [16]. Wang et al. introduced a Recurrent Convolutional Neural Network with variational inference for probabilistic RUL of bearings and milling cutters [17], and a Deep Separable Convolutional Network for efficient edge deployment [18]. Mo et al. combined a multi-head CNN–LSTM with real-time error analysis [19], and Amin and Kumar tuned an LSTM/RNN/CNN hybrid with genetic algorithms [20].

C-MAPSS was originally created for the PHM'08 Data Challenge [22]. Its simulation model integrates nonlinear damage-progression equations and thermodynamic principles, with a system health index that declines until a failure threshold is reached.

Techniques benchmarked on it range from Random Forests and Gradient Boosting [24] to LSTMs, CNNs and CNN–LSTM hybrids [12][25]. Alomari et al. combined Random Forest, XGBoost and Natural Gradient Boosting with Genetic Algorithm and Recursive Feature Elimination feature selection, reaching RMSE of 11.8, 23.0, 14.6 and 22.3 on FD001–FD004 [24]. Thakkar and Chaoui used deep-layer RNNs with thorough preprocessing on FD001 [25]. Across studies, T50, Ps30, Nf/Nc and flow indicators (W31, W32) are commonly predictive, though optimal subsets vary [24].

Gaps remain: many studies skip simpler baselines or use non-standard preprocessing, making comparison hard, and there's no consensus on the best sensor subset or model for multi-condition, multi-fault data like FD004. This study addresses that by evaluating models under consistent preprocessing and training conditions.

Data quality and preprocessing

Industrial sensor data suffers from noise, missing values and high dimensionality. Handcrafted features may not generalise across machines or fault modes [10][26], while DL models demand large, well-labelled datasets that are often unavailable [27].

Model selection and optimisation

No model is universally optimal; performance depends on dataset characteristics, fault types and operating environments [28]. DL models need careful architecture design and tuning and are prone to overfitting on limited data [29].

Interpretability and trust

In safety-critical domains, "black box" models impede human oversight. Explainable AI techniques such as feature attribution and interpretable surrogates are increasingly used to support decisions [30].

Deployment and maintenance

Concept drift degrades models over time, requiring continual learning or retraining; models must also be robust to anomalies and light enough for embedded deployment [31].

Benchmarking and dataset limitations

PdM lacks standardised benchmarks and diverse open datasets. C-MAPSS can't capture the full spectrum of real-world conditions, complicating fair comparison [32].