AF-TAD: Transformer-Based Anomaly Detection for Aircraft Fuel Systems
Xulang Ouyang, Y. H. Zhang, Hua Lü, Yongfeng Yin
Abstract
This paper presents AF-TAD, which stands for Aircraft Fuel Transformer Anomaly Detection, a Transformer-based reconstruction model specifically designed for time-series anomaly detection in aircraft fuel systems. The model improves anomaly detection by encoding and fusing key features such as fuel mass, overload, and RPM. Additionally, it incorporates global information, a multi-scale time window compression technique, and a conditional coding layer that integrates flight-specific information to enhance time-series data analysis, adaptability, and detection accuracy. Experimental results demonstrate that AF-TAD significantly outperforms other state-of-the-art anomaly detection models in detecting interval anomalies, achieving superior performance in key metrics. An ablation study further confirms that critical components, such as multi-window encoding and global trend information, contribute to improving the model’s detection capabilities. Overall, AF-TAD not only demonstrates powerful anomaly detection capabilities in the complex domain of aircraft fuel systems, but also offers new technical methods for fault prediction and maintenance in this field. Additionally, it holds broad potential for application in other complex systems.