Which machine learning algorithm is most effective for predictive maintenance of turbofan engines?

Comprehensive analysis of predictive maintenance algorithms applied to turbofan engines: challenges, technologies, players, costs, and effectiveness.

In summary

Predictive maintenance (or PdM) consists of anticipating component failure before it occurs using sensor data and machine learning algorithms. For dual-flow engines (i.e., with two or more spools), mainly civil or military turbofans, this system is particularly relevant because downtime or removal costs are very high. Several classes of algorithms are used: random forests, gradient boosting (XGBoost), recurrent neural networks (LSTM, Bi-LSTM), convolutional networks (CNN) or hybrid networks (1D-CNN-LSTM). Academic and industrial studies converge on the conclusion that hybrid deep learning approaches combining CNN + LSTM currently offer the best performance for predicting Remaining Useful Life (RUL), but they require a large volume of data and a robust architecture. In practice, the best solution will depend on the dataset, integration with aircraft systems, aviation regulations, and expected return on investment.

1. Context: definition and challenges of turbofan engines

1.1. What is a turbofan engine?

A turbofan engine, often referred to as a turbofan, is a jet engine in which the incoming air is divided: one part passes through the combustion core (via the compressor, combustion chamber, and turbine), while the other part bypasses this core via a bypass duct.

In “turbofan” terminology, we often refer to two spools (concentric rotating shafts): a low-pressure spool (fan + low stage) and a high-pressure spool (high-pressure compressor + high-pressure turbine).

For a dual-flow engine, in general:

the bypass flow represents a significant portion of the total air flow, which improves propulsive efficiency (high bypass ratio).
the two-spool architecture allows for higher compression ratios, a better speed range, and better thermodynamic efficiency, but also greater mechanical complexity.

1.2. Why does this architecture pose specific maintenance challenges?

The mechanical and thermal complexity of a dual-flow engine makes it a critical asset:

Reliability must be maximized; unplanned downtime is very costly.
The engine operates in an extreme environment (high temperatures, high rotational speeds, mechanical stresses), which requires engine health monitoring (EHM).
Maintenance costs (parts, labor, downtime) are significant, making predictive maintenance economically attractive.
Both civil and military engines share availability and performance requirements that justify investment in advanced predictive maintenance algorithms.

2. Predictive maintenance: concept, technology, differences

2.1. Definition of predictive maintenance

Predictive maintenance aims to predict the optimal time to intervene on a system before a failure occurs. It contrasts with purely preventive maintenance (fixed periodic interventions) or reactive maintenance (after failure).
In aviation, this means using flight data, engine status sensors, and maintenance records to predict the Remaining Useful Life (RUL) or detect anomalies before they become critical.

2.2. Technology and architecture

Essential technological components include:

Onboard sensors that measure variables such as pressure, temperature, rotational speed, vibrations, air flow, etc.
A data collection/storage system—often via an engine health monitoring (EHM) system and near real-time data streams.
Connectivity (satellite, datalink) to upload data to the ground or to a cloud server.
Machine learning (ML) or deep learning algorithms to analyze this data and produce predictions.
Dashboards and decision-making systems to alert maintenance teams and plan interventions.

Predictive maintenance differs from condition-based maintenance in that it does not simply trigger an intervention when a threshold is reached, but predicts when the intervention will be most effective and plans the task.

2.3. Expected benefits

According to recent studies:

Reduction in direct maintenance costs of around 12 to 18% in aviation thanks to AI and predictive maintenance.
Reduction in unplanned downtime of 15 to 20% in aircraft fleets.
In a specific case study of a turbofan engine, the optimized planning approach—using a reinforcement learning model—reduced maintenance costs by ~29% and avoided ~94% of unplanned maintenance.

3. Which algorithm is the most effective? Comparative analysis

3.1. Frequently used algorithms

In research on turbofan engines, the following algorithms are commonly used:

Classic machine learning (ML) algorithms: K-Nearest Neighbors (K-NN), Random Forest, Gradient Boosting/XGBoost.
Recurrent neural networks (RNN)/LSTM/Bi-LSTM for time series.
Convolutional neural networks (CNN) or hybrid 1D-CNN + LSTM.
Federated learning for distributed data.
Advanced combinations including attention mechanisms, deep networks, etc.

3.2. Evaluation criteria

The main criteria for comparing algorithms are:

Root mean square error (RMSE) or another indicator of accuracy in predicting RUL. Example: an RMSE of ~16.1 cycles obtained with a 1D-CNN-LSTM on a subset of the C-MAPSS database.
Computational cost, amount of data required, generalizability.
Ability to handle variability in conditions (different missions, regimes, over/under-regimes).
Interpretability and acceptance in the aeronautical regulatory environment.
Implementation cost, robustness in an operational setting (in-service operation).

3.3. Analysis results

Studies show:

In a study on turbofan engines, XGBoost outperformed a “basic” LSTM: RMSE ~21.29 with XGBoost versus equivalent but not better performance for LSTM.
Another study indicates that the combination of 1D-CNN + LSTM (with change point detection and feature engineering) achieves an RMSE of ~16.1 on one of the C-MAPSS sub-sets, which is significantly better than traditional ML methods.
A recent study on turbofan engines shows that the limiting factor for “simpler” machines (K-NN) remains the amount of data and variability.
The data training and segmentation framework plays a major role in the results: for example, the 2025 study (Jung et al.) notes that with 65% training data and segmentation approaches, an RMSE of 4°C to 6°C for turbine temperature (TGT) was obtained on a family of engines.

3.4. What conclusion can be drawn about the “most effective” algorithm?

For dual-flow engines (turbofans), the most effective algorithm today appears to be a hybrid deep learning architecture, typically a combination of CNN + LSTM, or even, in the long term, with an attention mechanism or transformer variants, for the following reasons:

It allows both the *temporal dimension * (sequence of cycles, sensor variables) thanks to LSTM and the *spatial or structural dimension* (automatic extraction of complex features) thanks to CNN.
It outperforms “classical” ML techniques in terms of RUL accuracy in several studies (e.g., RMSE ~16 cycles vs. ~21 for XGBoost).
It is better suited to the complexity of turbofan behavior in service, with large variations in speed, mission, and load.
However, it should be noted that:
This algorithm requires a significant amount of data (flight cycles, sensors, history); otherwise, simpler algorithms may be preferable.
The cost of computation and implementation (preprocessing, feature engineering, segmentation) is higher.
If the dataset is limited or highly variable, a “less sophisticated” algorithm such as Random Forest or XGBoost may be more reliable in the short term.

So, in summary, the best choice depends on the context: for a modern fleet with a large volume of data, the CNN-LSTM hybrid is recommended. For a smaller fleet or limited data, a well-calibrated XGBoost may suffice.

4. Practical implementation in engine maintenance

4.1. How does it work in practice?

Collection of engine status data (internal sensors, temperatures, pressures, speeds, flow rates, etc.), often at each cycle or time interval.
Preprocessing: data cleaning, removal of seasonal trends, drift removal, normalization, data segmentation by engine or engine family. Example: Jung et al. applied detrending to improve learning quality.
Feature engineering: derivative extraction, moving average, speed change, possibly dimensionality reduction (PCA). Example: Melkumian (2024) used PCA + XGBoost and LSTM.
Algorithm selection: XGBoost, Random Forest, CNN, LSTM, hybrid. Model training on historical “healthy” data (with labels indicating time remaining until maintenance or failure).
Validation and testing: performance measures (RMSE, R², AUC, etc.).
Deployment: the model is embedded or connected to the fleet, generating alerts or RUL predictions in near real time.
Maintenance planning: the algorithm provides an optimal maintenance horizon, hangar planning, spare parts, synchronization with flight schedule.
Periodic updates: performance may deteriorate with new operating conditions, new engines, new regimes. The model must be retrained or refined. The “federated learning” approach ” approach allows the model to be updated without centralizing all the data.

4.2. Is it used on fighter jets and what are the challenges?

Although many publications focus on civil turbofan engines (C-MAPSS, etc.), application to fighter jet engines (often with low bypass, more aggressive architecture) is possible, but presents additional challenges:

Life cycles are different (high-stress flights, afterburners, maneuvers).
Flight data is often classified, or there are fewer or shared onboard sensors.
Maintenance must comply with very strict requirements (safety, speed, redundancy).
However, some suppliers of fighter jets or military engines are integrating engine health monitoring (EHM) systems that can benefit from PdM algorithms. The principles remain the same: sensors → RUL prediction → optimized maintenance.
So, yes, use in the defense sector is possible, but with increased security, confidentiality, and certification constraints.

4.3. Costs, return on investment, and savings

Implementation costs include:

Installation or updating of sensors and data collection systems.
Data storage/processing (cloud, edge computing).
Algorithm development or acquisition (license, consultants).
Team training, integration with maintenance systems.
The expected savings are significant:
According to the 2025 study, maintenance costs are reduced by 12-18% and overall fleet availability is increased.
Another study: costs are reduced by ~20-25% and unplanned downtime by up to ~30%.
University study for turbofans: cost reduction of up to ~29% with an optimized RUL + planning approach.
Thus, although the implementation cost is high, the return on investment can be achieved in 12 to 24 months under favorable conditions.

4.4. Regular updates and connectivity

For the algorithm to remain effective, the following is required:

Regular data updates (cycles, new engines, new conditions).
Connectivity to upload data to analysis servers or onboard modules. The absence of fresh data reduces efficiency.
Re-training of the model or at least periodic refinement.
Data security management, which is particularly critical in aviation. Federated learning can reduce confidentiality issues.
Good integration with maintenance/flight planning, otherwise predictions will not be acted upon and value will be lost.

5. Why is the hybrid CNN + LSTM “the most effective”?

5.1. Capturing temporal dynamics and structure

A twin-flow engine generates a huge amount of sensor data: time sequences, subtle variations, mission effects, speed, altitude, load.

LSTM (Long Short-Term Memory) is capable of memorizing long-term dependencies in a time series; it is suitable for data from successive cycles.
CNN (Convolutional Neural Network) allows abstract features to be extracted from the sequence or signal without having to define them manually.
By combining the two, we obtain a model capable of “seeing” wear patterns or degradation signatures in the data and then tracking them over time.

5.2. Superior numerical results

As mentioned, the 2025 study shows RMSE ~16.1 cycles for a 1D-CNN-LSTM model, which is significantly better than traditional ML algorithms.
Furthermore: in the study by Jung et al. (2025) on turbofans, the hybrid model used 65% less data than the conventional train/test method and achieved an error of 4°C to 6°C on the turbine temperature (TGT): this indicates that the architecture is robust even with less data.

5.3. Implementation conditions

This architecture is particularly suitable when:

The volume of data is high (several thousand cycles per engine).
Access to varied and reliable sensor data is available.
Processing power is available (embedded or in the cloud).
The maintenance system is ready to integrate and execute recommendations.
However, if these conditions are not met, a simpler algorithm may be preferable.

6. Limitations, challenges, and precautions

6.1. Data quality and quantity

An advanced algorithm is worthless without good data: faulty sensors, missing data, and mission bias can compromise its validity.
Furthermore, for a single engine or a small fleet, the volume of data may be insufficient for robust deep learning; in this case, a lighter ML algorithm will be more reliable.

6.2. Complexity and implementation cost

CNN+LSTM hybridization requires expertise (data science, engineering), infrastructure (storage, computing), and time for training and validation. The return on investment must be carefully evaluated.
In addition, aviation imposes certification, traceability, and auditability standards: “black box” algorithms pose a problem, which can limit adoption.

6.3. Operational exploitability

The efficiency of the system depends on the entire chain: sensors → algorithms → planned maintenance. The algorithm must provide actionable results: it is not enough to predict if the recommendations are not implemented. As one source points out: “collecting and cataloguing data only works if it is acted upon.”

6.4. Security, connectivity, cybersecurity

Flight data is sensitive; connectivity must be secure. Federated learning offers a partial answer to the question of confidentiality but introduces other complexities.

6.5. Adaptation to military missions/fighter jets

For fighter jet engines, the environment is even more severe: extreme stresses, highly variable mission cycles, specific maintenance protocols. The algorithm must be specifically validated for this context.

7. Recommendations for turbofan engine operators

Assess your data volume: if you have thousands of reliable and consistent sensor cycles, consider a CNN+LSTM model. Otherwise, start with a simpler ML model (XGBoost, Random Forest).
Perform rigorous pre-processing: trend removal, mission segmentation, normalization. As the study shows, segmenting training sets significantly improves results.
Set up a collection and processing infrastructure: include EHM, data transmission, storage, maintenance dashboard.
Ensure the intervention loop: predictions must lead to concrete actions (parts, planning, technicians).
Plan a gradual rollout: test first on a subset of engines or aircraft, then expand.
Anticipate updates: plan for a retraining or continuous learning mechanism, especially if you are using federated learning.
Assess the economic gain: use your own KPIs (MTBF, cost per flight hour, downtime rate) to model the return on investment. Even if studies show a 12-18% reduction in costs, each fleet is different.
In the aerospace or defense environment, ensure certification, algorithm traceability, and transparency (e.g., explainable models): an algorithm that is too much of a “black box” could pose a problem.

8. Summary and outlook

The analysis shows that, to date, for predictive maintenance of turbofan engines, the combination of a 1D convolutional neural network and a recurrent neural network (LSTM) is the most promising solution in terms of accuracy and ability to model the complexity of engine data. However, this recommendation comes with certain conditions: sufficient data volume, suitable infrastructure, controlled operational integration, and clear return on investment.
For fleets with limited data or significant implementation constraints, a simpler algorithm (e.g., XGBoost) remains a pragmatic option.
In the future, the integration of attention mechanisms, transformers, or federated learning could further improve predictions, reduce dependence on large volumes, and expand use to military fleets where data is sensitive.
Ultimately, the choice of algorithm cannot be separated from the overall ecosystem: sensor hardware, data quality, maintenance operations, and economic return. A holistic approach is essential for predictive maintenance to become a truly effective tool for dual-flow engines.

War Wings Daily is an independant magazine.