Digital Twin-Driven Condition Monitoring Approach for Aircraft Carbon Brakes
Author(s)
Jammal, Patsy
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Wheels and brakes are the most significant contributors to aircraft component maintenance costs. Specifically, carbon brakes have a high initial cost, and their current Time-Based Maintenance (TBM) practice, which involves performing maintenance activities at fixed intervals, often leads to unnecessary inspections and increased aircraft downtime. Shifting to Condition-Based Maintenance (CBM), based on the actual health state of the brakes, can help reduce maintenance costs, increase aircraft availability, and enhance carbon brake life and safety as tailored insights can be provided to reduce wear.
CBM is enabled by Prognostics and Health Management (PHM) studies, which focus on diagnosing current component health and predicting future conditions and performance. Carbon brakes and their maintenance are a significant revenue stream for manufacturers, who are also granted access to operational data from their customers. Big data analytics on such data enables innovative solutions to optimize brake maintenance. While traditional condition monitoring techniques lacked real-time predictive capabilities, the advent of Digital Twin (DT) technology and the abundance of multidomain data availability has given rise to a new DT-driven condition monitoring approach using enhanced data-driven methods such as Artificial Intelligence (AI) and Machine Learning (ML). This advancement enables predictive maintenance by allowing for timely detection and prediction of potential risks and failures.
Still, the existing literature on carbon brake wear highlights several gaps. These include the lack of understanding of the effects that varying operational and environmental conditions have on carbon wear, the limited use of advanced ML techniques on high-dimensional datasets from actual operations, and the lack of generalizability assessments and enhancements of predictive models across different domains (e.g., varying aircraft types). Therefore, this research aims to develop an optimized and generalizable data-driven methodology that predicts carbon brake wear based on various parameters, including aircraft-specific parameters, operational conditions, and environmental factors.
This research addresses the identified gaps by developing a robust and repeatable methodology called AIM-Wear (Advanced Implementation of Machine Learning for Wear Monitoring). This framework is not only tailored for monitoring wear in aircraft brakes but can also be adapted to other components subject to wear. First, clustering techniques are used to identify brake wear patterns and varying ranges of aircraft, operational, and environmental parameters corresponding to different levels of wear severity. By integrating multidimensional datasets—including aircraft data, weather conditions, and airport characteristics—this methodology captures a more comprehensive view of the factors influencing brake wear. The findings validate that clustering techniques with optimized hyperparameters reliably categorize brake degradation severity based on operational and environmental variables. The results highlight the ability of clustering algorithms, such as K-Means and DBSCAN, to achieve high accuracy in aligning with true wear labels while uncovering nuanced patterns and trends. This integration of clustering and feature selection through supervised methods like Random Forest not only enhances the understanding of wear dynamics but also enables more informed and proactive predictive maintenance strategies. This approach could also lead to adjustments in duty cycle procedures, i.e., dynamometer brake performance tests, to better reflect actual operations.
Next, supervised ML algorithms are used to develop classification models that better determine the severity of carbon brake wear based on varying operational and environmental conditions. Such capability allows flights experiencing excessive brake wear to be identified along with the influencing factors. A comprehensive benchmarking of classifiers revealed that the LGBM classifier is the best performer, achieving exceptional results across all metrics. A well-tuned Decision Tree classifier also demonstrated comparable performance, offering valuable interpretability. These models not only provide reliable predictions but also quantify prediction uncertainty, ensuring robustness in real-world applications. The findings validate that supervised ML methods are highly effective for predicting brake wear severity, highlighting the relevance of feature importance analysis in uncovering influential factors. Furthermore, the results emphasize the tradeoff between achieving high predictive performance and practicality, particularly minimizing false negatives, where actual high wear is predicted as low—a critical consideration for safety and operational deployment.
The problem is then tackled as a regression task for more accurate and precise brake wear predictions using traditional ML and advanced Deep Learning (DL) techniques, allowing for a benchmark of multiple algorithms and assessing their suitability for DT modeling purposes. The findings reveal that traditional ML models, such as the Decision Tree Regressor, outperform DL models in terms of predictive performance, computational efficiency, and interpretability, which are critical for predictive maintenance in aviation. The Decision Tree Regressor achieves the lowest Mean Squared Error (MSE) and highest Coefficient of Determination (R²), demonstrating consistency and efficiency in training. Advanced DL models, such as Long Short-Term Memory (LSTM), perform adequately (comparable R²) but require significantly longer training times, confirming that the performance gains from DL do not justify their complexity for this task, supporting the preference for traditional ML approaches. The findings also highlight the importance of selecting modeling techniques aligned with specific operational requirements, emphasizing the role of feature engineering and data preprocessing in enhancing model performance. Refined data handling by adopting a more granular approach to the data, such as using per-flight data without condensation or leveraging raw full-flight data, could further improve predictions.
Both regression and classification models rely on the availability of wear pin signals from flight data, which indicate the percentage of carbon brake disk remaining. These signals are reported inconsistently, approximately every ten flights. To address this, Step 3 of the methodology also investigates the optimal frequency for collecting wear pin values, determining how often operators should report carbon brake thicknesses for aircraft lacking electronic wear pin sensors. This enables the development of supervised predictive models for such aircraft. The results confirm that increasing wear pin signal reporting frequency enhances predictive performance up to a threshold, beyond which further increases yield diminishing returns. The findings reveal the optimal reporting frequency for wear pin data to be every six flights, minimizing MSE while balancing data granularity and operational efficiency. This frequency supports precise wear pattern detection without overfitting or excessive noise, enabling maintenance staff to schedule inspections effectively.
Furthermore, the generalizability of the predictive brake wear models is assessed to determine whether specific models need to be developed for different data segments (or domains, such as diverse aircraft types), or if a single model can provide acceptable performance across the entire dataset. The findings reveal that a well-optimized generalized model effectively handles variations within the fleet, achieving robust performance across different aircraft types. This approach confirms that a single, generalized model trained on diverse data is more efficient and better performing than multiple specialized models. It also simplifies the maintenance strategy and reduces computational overhead by eliminating the need to store, deploy, and maintain multiple models.
Lastly, Transfer Learning (TL) techniques are incorporated to explore their potential to enhance model performance across the different data segments, taken as distinct variants of the widebody aircraft. The findings reveal that TL significantly enhances model adaptability and performance on smaller or less representative datasets, such as the large-sized aircraft variant’s dataset. For instance, by leveraging knowledge from the broader datasets of the smaller-sized variants (Variants 1 and 2), significant reductions were achieved on the limited dataset of the larger-sized variant (Variant 3). These results confirm that TL not only boosts model generalizability across varied domains but also improves computational efficiency by reducing training and prediction times. This strategy enables the development of a unified, accurate model for predictive maintenance across diverse contexts. By reducing the need for multiple models and frequent retraining in Step 4 and the use of TL in Step 5, the AIM-Wear approach further optimizes resource use, supports reliable maintenance scheduling, and enhances operational efficiency, directly benefiting the aviation industry's safety and cost-effectiveness.
Sponsor
Date
2025-01-14
Extent
Resource Type
Text
Resource Subtype
Dissertation