Unsupervised Multimodal Anomaly Detection With Missing Sources for Liquid Rocket Engine.

Area of Science:

• Aerospace Engineering
• Machine Learning
• Data Science

Background:

• Anomaly detection (AD) in large equipment like liquid rocket engines (LREs) is critical for reliability.
• Current deep learning methods often struggle with missing data (modality incompleteness) in multimodal sensor streams.

Purpose of the Study:

• To develop an unsupervised multimodal AD method specifically designed to handle missing data sources in LRE systems.
• To improve the robustness and accuracy of anomaly detection in complex engineering environments.

Main Methods:

• A unified deep learning framework using multiple autoencoders (AEs) and a skip-connected AE.
• Intramodality and intermodality fusion techniques to restore missing data and enhance reconstruction.
• Minimizing reconstruction loss while maximizing latent space representation dissimilarity for anomaly scoring.

Main Results:

• The proposed method effectively restores missing sensor data, enabling reliable secondary reconstruction.
• Anomaly scores are derived from reconstruction errors and latent space discrepancies.
• Anomaly score fusion at the decision level further boosts model performance.

Conclusions:

• The developed unsupervised multimodal AD method demonstrates superior performance in LREs with missing data sources.
• The approach offers a promising solution for enhancing the safety and reliability of critical aerospace systems.