A hybrid framework for real-time satellite fault diagnosis using Markov jump-adjusted models and 1D sliding window Residual Networks

Data-driven methods, including Artificial Intelligence (AI) and Machine Learning (ML) techniques, have been becoming more prominent in the field of satellite Fault Diagnosis and Prognosis (FDP) owing to their exceptional pattern recognition capabilities. On the other hand, they have some glaring accompanying issues other than their data dependency that have not been explored in the literature on satellite fault diagnosis. These issues include their inability to accommodate real-time fault diagnosis requirements, failure to account for the fault diagnosis and fault-tolerant modules’ interactions, and being prone to getting overfit due to manually injected faults. Therefore, this work proposes a hybrid framework for real-time fault diagnosis of a single Reaction Wheel (RW) onboard a satellite that capitalizes on both data-driven and model-based methods’ strong suits. The proposed methodology can also be applied to other satellite sub-systems. The presented hybrid framework comprises a Morkov jump-adjusted RW model, a Markov Jump-Adjusted Particle Filter (MJAPF), and a One Dimensional (1D) sliding window Residual Network (ResNet). The Morkov jump-adjusted RW model addresses the under-represented issues of data-driven methods, the MJAPF provides a means of estimating the non-linear RW’s hidden states under non-Gaussian noise conditions while accounting for malfunction dynamics, and the 1D sliding window ResNet model ensures online diagnosis performance. Experiments showed that the hybrid framework can achieve accurate and timely results, even reaching accuracy rates as high as 99% in low-noise conditions. The proposed MJAPF algorithm proved to be a capable estimation technique. However, the proposed MJAPF and ResNet frameworks were incompatible due to the gap in their perceptions of fault dynamics but proved effective on their own merits. Future remarks for making the proposed hybrid framework more robust to noise are also discussed. • Highlighted limitations of AI and ML in satellite fault diagnosis: inability to meet real-time requirements, failure to account for interactions between fault diagnosis and fault-tolerant modules, and susceptibility to overfitting from manually injected faults. • Developed a hybrid approach combining data-driven and model-based methods for real-time fault diagnosis of Reaction Wheels (RWs) on satellites. • Emphasized the adaptability of the framework for other satellite sub-systems. • Created a Markov Jump-Adjusted RW Model to address under-represented issues in data-driven methods. • Developed a Markov Jump-Adjusted Particle Filter (MJAPF) to estimate non-linear RW hidden states under non-Gaussian noise conditions and account for malfunction dynamics. • Designed a One Dimensional (1D) Sliding Window Residual Network (ResNet) to ensure online diagnosis performance. • Demonstrated the hybrid framework’s ability to achieve high accuracy (up to 99%) in low-noise conditions through experimental validation. • Showed that the MJAPF algorithm is a capable estimation technique. • Identified incompatibility between MJAPF and ResNet due to differing perceptions of fault dynamics but validated their individual effectiveness. • Discussed recommendations for enhancing the hybrid framework’s robustness to noise.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.