A deep learning approach to fault detection in satellite power system using Gramian angular field

In this paper, an approach based on time series-to-image mapping is proposed for fault detection in a satellite power system (SPS). This approach exploits the possibilities of encoding the SPS time series data as images using Gramian angular fields (GAF). The resulting images are analysed by a convolutional neural network (CNN) for recognising faulty and normal conditions of SPS. Validation with NASA's advanced diagnostics and prognostics testbed (ADAPT) dataset has demonstrated that the combination of CNN with GAF results in better performance when compared to other image encoding methods such as spectrogram and recurrence plot (RP). The proposed approach yields an accuracy of 85.13% with precision 84% and F1 score 0.91 suggesting that encoding multivariate time series data to images using GAF is worth considering for SPS fault diagnosis when compared to other time series-to-image encoding based approaches.