|
|
|
Automated Event Detection and Classification in High Volume Space Physics Time Series Data using Machine Learning
|
|
|
| Modern satellite-based space physics instruments generate vast streams of temporal multi-band data. A trained expert can often examine a suitably displayed portion of the data stream and identify such events by hand, but for high-volume data streams this approach is entirely impractical. Writing robust feature detectors is very time-consuming and requires a high degree of programming skill and
familiarity with pattern recognition techniques -- certainly not skills that every scientist has. We propose to investigate the use of supervised machine learning techniques to make this job much
easier. We intend to develop a general purpose user-friendly software tool for automatically generating event classifiers that operate on multi-band time-series data using advanced machine learning techniques. We intend to look at using the relatively recently developed technique of Support Vector Machines. SVMs have a strong mathematical foundation and have been empirically shown to
be one of the most powerful and robust machine learning technologies available today. We will also consider the use of Boosting and Gaussian Processes, two other recently developed and
mathematically justifiable machine learning paradigms. Using whichever technique works best, we will develop a general software tool that can be applied to a variety of different space physics data
streams. The primary target of our initial research will be magnetospheric data collected by satellites such as the Los Alamos geosynchronous satellites (Magnetospheric Plasma Analyzer, MPA;
energetic particles, SOPA and ESP), The GPS energetic particle detectors, with further application to other data sets existing in-house (POLAR CEPPAD/CAMMICE/HYDRA) as well as upcoming data
from the recently launched Cluster II sets of satellites. |
|
|
