Fault Detection And Prognosis Of Aerospace Systems Using Long Short-Term Memory Based Recurrent Neural Networks
thesisposted on 17.09.2021, 19:17 by FAULT DETECTION AND PROGNOSIS OF AEROSPACE SYSTEMS USING LONG SHORT-TERM MEMORY BASED RECURRENT NEURAL NETWORKS
Health monitoring and remaining useful life predictions for the aerospace systems is a challenging and complex task to accomplish. Internal or external complications in these aerospace systems (aircraft and satellites) may lead to extremely hazardous or catastrophic consequences to the entire mission involving human life and budget. Considering the severity and complexity of the problem, this thesis deals in developing a diagnosis and prognosis health management system (DPHM) for the attitude actuator control system that uses reaction wheels in pyramid configuration onboard Kepler spacecraft and for the fleet of air-breathing turbofan engines. The established model is
comparatively effective and computationally light in managing the objective of fault detection and prognostics.
An advanced data-driven DPHM scheme with optimization techniques is developed and evaluated. Initially, a recurrent LSTM (Long Short-Term Memory) neural network model is
established and assessed with the general dataset (Particulate Matter (PM2.5)). Secondly, a statistical-based fault detection method with functional factors of Weibull and mathematical features of frictional parameters showed that reaction wheels 2 and 4 of Kepler spacecraft have an early sign (~2 months) of their respective failures. This statistical method is compared with the proposed LSTM model for validation. Thirdly, the prognostic approach for estimation of remaining useful life (RUL) of the C-MAPPS and PHM08 datasets is successfully achieved. Numerous preprocessing methods such as digital filters (Savitzky-Golay (S-G)), principal component analysis (PCA) are used for standardizing the data. Finally, the optimization tools such as genetic algorithm (GA) and particle swarm optimization (PSO) are merged with LSTM for finetuning the hyper-parameters. Overall, the optimized model performs with better accuracy and can be concluded as a promising algorithm for the health management of complex systems.