Multi-modal LSTM network for anomaly prediction in piston engine aircraft

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Aircraft require regular maintenance due to their complex systems.
• Existing AI health monitoring systems offer efficiency gains over traditional methods.
• Engine anomalies like flameouts are often linked to rapid temperature changes.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting piston engine flameouts.
• To address limitations in previous studies regarding comprehensive anomaly prediction and data variability.
• To investigate the use of engine oil and cylinder head temperatures for failure prediction.

Main Methods:

• Data pre-processing from real-time flight data recorders.
• Application of conventional and deep learning models, including a custom multi-modal regularised Long Short-Term Memory network.
• Utilisation of engine oil and cylinder head temperatures from a Textron Lycoming IO-540 engine.

Main Results:

• The proposed Long Short-Term Memory network achieved improved accuracy with low root mean square errors (0.55 for cylinder head, 3.20 for engine oil temperature).
• Performance significantly outperformed other popular machine learning methods by up to 84%.
• The system demonstrated capability in predicting engine flameout events 2 minutes ahead of occurrence.

Conclusions:

• The developed AI system effectively predicts engine flameouts using temperature data.
• The proposed model enhances generalization and avoids overfitting, making it suitable for variable flight data.
• The system is a viable candidate for integration into aircraft engine control units for enhanced safety and operational efficiency.