Fault Diagnosis and Prognosis of Aerospace Systems Using Growing Recurrent Neural Networks and LSTM

• Published in: 2021 IEEE Aerospace Conference (50100) pp. 1 - 20
• Main Authors: ElDali, Musab, Kumar, Krishna Dev
• Format: Conference Proceeding
• Language: English
• Published: IEEE 06.03.2021
• Subjects: Atmospheric modeling; Indexes; Predictive models; Recurrent neural networks; Space vehicles; Training; Wheels
• DOI: 10.1109/AERO50100.2021.9438432

Due to the increase in complexity in aerospace systems, developing a diagnosis, prognosis, and health monitoring (DPHM) framework is a challenge that must be considered to assure the safety of such systems. This paper discusses this problem by proposing an artificial intelligence technique based on two novel neural networks, the growing neural networks (GNN) and variable sequence LSTM (VarLSTM) model to automate the process of DPHM for aerospace systems. For single-unit datasets, the proposed model estimates a Health Index value using the residuals between the measured telemetry data and the one predicted using the GNN algorithm, and then the HI value is extrapolated for prognostics. For multiple-units datasets, the model makes RUL predictions by directly mapping the RUL of the training units to their corresponding measured features at every measured instant. In this paper, the model optimizes the architecture of a recurrent neural network and was used to make RUL predictions for aircraft engines and detect failure for satellite attitude actuators (Reaction Wheels). It was tested on the CMAPSS and PHM08 aircraft engine datasets (multiple-unit datasets) simulated by NASA, and it was able to make RUL predictions with root mean square errors as low as 14 engine cycles. Another application to test the proposed model was on the Kepler Spacecraft's reaction wheels from which two have failed (single-unit datasets). The model detected the failure of the two failed reaction wheels by estimating a HI value which indicates the probability of failure of the reaction wheels using the residuals between the speed predictions made by the model and measured speed values. Failure was detected using the model almost 105 days and 54 days for reaction wheels two and four respectively. Prognostics were also applied on the Kepler Mission reaction wheels and RUL predictions were made with mean absolute errors ranging between 2-13 days depending on how close the reaction wheel is to fail when the prediction is made. The proposed artificial intelligence algorithm shows promising results in system fault diagnosis and prognosis leading to the development of smart systems for aerospace applications.