부산대학교 도서관

Automatic time-series anomaly detection has been utilized for several applications such as general infrastructure systems. While accuracy and quickness are important factors for anomaly detection, explainability is crucial for safety-critical systems such as space systems. In this context, traditional machine learning-based methods of time-series anomaly detection such as Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) are not able to explain why symptoms of anomaly are detected. However, flight controllers of International Space Station (ISS) need to assess trends of telemetries and understand situations before making coordination with other domestic and foreign flight controllers and taking any actions corresponding to anomaly detections. In this research, we conducted the following three steps to detect signs of anomaly with rationales behind the anomaly, i.e., 1) selecting objective and explanatory variables, 2) conducting anomaly detections with time-series machine learning, 3) providing additional information with assessing the anomaly detections. The method was verified with the data of two anomaly events of Low Temperature loop (LTL) pump inverter of Japanese Experimental Module of the ISS in 2012 and 2014. There were some different trends of telemetries in 2012 and 2014. A selected objective variable was successfully predicted based on explanatory variables in normal period and away from the actual measured value in off-nominal period with LSTM for both cases. The analysis results showed the effectiveness of our new methods as the machine learning-based time-series predictive failure detection with identifying possible causes. Our proposed method can contribute to health monitoring of space systems and anomaly detections with showing possible reasons behind anomaly events. It is expected to widely contribute to the safety of space systems.