A comprehensive framework for real-time sensor fault detection,isolation,and inference in turbojet engines using autoencoder

Abstract

Turbojet engines play a pivotal role in modern aircraft systems by delivering the thrust necessary for flight, with sensors providing critical data for real-time monitoring and fault diagnosis. However, sensor malfunctions can undermine control capabilities and jeopardize flight safety, which requires prompt detection and mitigation strategies. This paper aims to detect anomalies in sensor measurements instantaneously, isolate faulty sensors, and infer accurate data to ensure reliable turbojet engine operation during flight. We propose a modular, real-time fault-detection, isolation, and inference (FDII) framework for turbojet engines that integrates autoencoder-based virtual sensing, adaptive thresholding, and multi-sensor inference. A key innovation is a time-varying adaptive threshold rather than a fixed limit, the threshold adapts online to engine dynamics, greatly enhancing robustness. In our architecture, an autoencoder continuously reconstructs each sensor’s output; when the normalized reconstruction error exceeds the adaptive threshold, a fault is flagged. Fault inference then leverages full-system sliding-window inputs: all sensor data over a short time window are fed into a multi-sensor inference module, which learns normal correlations and reconstructs faulty readings. A detailed evaluation using real-world turbojet engine datasets, subjected to diverse fault scenarios and noise conditions, validates the effectiveness of the architecture in real-time fault diagnosis.