California State Polytechnic University – Pomona
Anomaly Detection In Spacecraft Telemetry Using Neural Networks.
Degree: MS, Department of Mathematics & Statistics, 2019, California State Polytechnic University – Pomona
A spacecraft is an encompassing term which is used to define all mission types at NASA???s Jet Propulsion Laboratory (JPL) such as: Mars rovers, outer planetary orbiters, and earth orbiting satellites. Nominal behavior within a spacecraft is defined by a diverse amount of categorical and numerical variables. Identifying when and where an anomaly has occurred within a spacecraft's subsystem is valuable information for the cognizant engineers, as it allows them to mitigate future such issues. Finding anomalies requires a broad predictive model which can learn the difference between shifts in a subsystem's nominal distribution caused by expected changes in state, and the shifts which move the distribution temporarily into a low density region based on its current temporal state. This thesis demonstrates the effectiveness of using a dual stage Long Short Term Memory (LSTM) neural network with attention layers as a model which can differentiate between a spacecraft's nominal and anomalous behavior.
For the purposes of this research, data from Soil Moisture Active Passive (SMAP) is used. SMAP is an earth orbiting satellite which collects information about the moisture level within the first five centimeters of soil to obtain a better understanding of the carbon and water cycle. A channel refers to a specific health measurement of a subsystem, only a small subset of SMAP's channels which have known anomalies are examined. To determine which preprocessing methods and additional inputs best predict anomalies, four different neural networks are trained for each channel. Three of these models use a dual stage LSTM with attention layers, and are tested against a fourth traditional LSTM, referred to as the base model, in predicting anomalies. The three neural networks are differentiated by their model inputs, and compared to the base model which uses the unaltered channel values. An exponentially smoothed model uses a low-pass filtering of the channel values as model input. A time delay embedded model uses a phase space representation of the channel as model input. Lastly, a multiple channel model is constructed using additional channels within the same subsystem as model input. Each model uses the mission's categorical data in the form of a temporal binary encoding. The dimensions of this encoding matrix are reduced by taking a representation of the binary encoding which is maximally informative and removes the most redundant categories. The categorical data are used to inform the model of changes in state which might result in a corresponding shift of a channel's distribution. While no model performs the best uniformly across all channels, the time delay embedding shows promising results for channels with high periodicity, and the multiple channel model shows significant benefit in identifying anomalies that are more pronounced when a channel's interactions with other channels are taken into account.
Advisors/Committee Members: King, Adam (advisor), Cannons, Jillian (committee member).
Subjects/Keywords: neural network
to Zotero / EndNote / Reference
APA (6th Edition):
Francis, C. (2019). Anomaly Detection In Spacecraft Telemetry Using Neural Networks. (Masters Thesis). California State Polytechnic University – Pomona. Retrieved from http://hdl.handle.net/10211.3/212245
Chicago Manual of Style (16th Edition):
Francis, Connor. “Anomaly Detection In Spacecraft Telemetry Using Neural Networks.” 2019. Masters Thesis, California State Polytechnic University – Pomona. Accessed August 24, 2019.
MLA Handbook (7th Edition):
Francis, Connor. “Anomaly Detection In Spacecraft Telemetry Using Neural Networks.” 2019. Web. 24 Aug 2019.
Francis C. Anomaly Detection In Spacecraft Telemetry Using Neural Networks. [Internet] [Masters thesis]. California State Polytechnic University – Pomona; 2019. [cited 2019 Aug 24].
Available from: http://hdl.handle.net/10211.3/212245.
Council of Science Editors:
Francis C. Anomaly Detection In Spacecraft Telemetry Using Neural Networks. [Masters Thesis]. California State Polytechnic University – Pomona; 2019. Available from: http://hdl.handle.net/10211.3/212245