Comprehensive Machine Learning-Based Time-Series Anomaly Detection for ADN-Based Thruster

Abstract

Ammonium dinitramide (ADN)-based thrusters are pivotal for future spacecraft propulsion due to their low toxicity, adjustable specific impulse, and environmental benefits. However, complex fault patterns observed during ground tests challenge traditional fault detection methods, which struggle with high-dimensional, nonlinear time-series data. This study proposes a machine learning-based approach for robust fault diagnosis in ADN-based thrusters. Using 189 real engine test time-series datasets, we performed systematic preprocessing and feature engineering to extract statistical and correlation characteristics inside experimental data, creating a standardized dataset of normal and faulty conditions. Ten algorithms—six traditional machine learning and four deep learning—were evaluated for fault identification. The multilayer perceptron achieved 98.2% accuracy and 100% recall, while random forest and XGBoost, attained accuracies of 99.1% and 98.2% respectively, with superior computational efficiency. Deep learning excels in complex scenarios but demands longer training, whereas traditional methods suit real-time applications. Feature analysis highlighted pre-injection pressure and capillary outlet temperature as key fault indicators. A Simcenter AMESim-based simulation model further augmented the dataset, supporting fault mechanism studies. This approach enhances fault diagnosis, health monitoring, and design optimization for ADN-based thrusters, offering significant engineering value.