A Multi-layer Deep Learning Framework for Real-Time Fault Detection and RUL Prediction in Avionics Systems

Abstract

To address the challenge of real-time fault detection in avionics equipment, we proposed a comprehensive framework consisting of data , model, and decision layers. Data is collected in real time from multiple IoT sensors and processed based on a standardized framework. The data model layer utilizes three models for electronic equipment anomaly detection, multi-category fault identification, and remaining equipment life prediction . The model layer contains three core modules: (1) an unsupervised anomaly detection model based on stacked denoising autoencoders (SDAEs) to identify early latent faults using reconstruction errors; (2) a long short-term memory network (LSTM-Attention) with an attention mechanism for accurate classification of multi-class faults; and (3) a Wiener degradation process model based on Bayesian updates to achieve probabilistic prediction of RUL. Experiments were conducted based on the NASA C-MAPSS dataset and real flight data provided by partner airlines. The results showed that the framework achieved an F1 score of 0.969 in the anomaly detection task, an average accuracy of 97.2% for multi-class fault classification, and a RUL prediction interval coverage of 91.3%. After quantization and compression, the model inference latency was only 18.3 milliseconds, meeting the stringent requirements of airborne equipment for lightweight and real-time performance. This method not only improves the economy and safety of aviation maintenance but also provides an interpretable, low-latency end-to-end solution for intelligent health management.