A Stacked Predictor and Dynamic Thresholding Algorithm for Anomaly Detection in Spacecraft

Anomaly or abnormal behavior detection in downlinked spacecraft telemetry is a key step for determining root cause of subsystem failures. Long Short-Term Memory networks (LSTMs) have been demonstrated to be useful for anomaly detection in time series, due to their predictive power, especially with long sequences that contain contextual information. However, there could be thousands of telemetry channels transmitted from the spacecraft with diverse signal characteristics. A single LSTM predictor may not perform well for all data channels. To enhance adaptability, we propose a stacked predictor framework. This framework stacks three LSTM based predictors with different neuron arrangements and a Support Vector Machine (SVM) based predictor in its first layer. A traditional fully connected layer is used as its second layer. We then estimate the distribution of the error of the predicted output using Kernel Density Estimation (KDE). Finally, a novel dynamic thresholding algorithm is applied to optimally extract anomalous behavior in the data. We present the Stacked Predictor results against the benchmark from the NASA MSL/SMAP anomaly dataset.