Abstract
The IMS seismic network produces an abundance of time-series data, posing great challenges for on-line processing and unbiased near real-time analysis. To this end, methods borrowed from the field of machine learning and data mining provide elegant solutions. By adhering to the multivariate statistical framework of Dynamic Bayesian Networks we make use of historical data obtained from the LEB bulletin to train a classifier to capture the intrinsic characteristics of signal and noise patterns appearing in seismic data streams. On a per station basis this yields generative statistical models that essentially summarize and generalize the information implicitly contained in the LEB allowing for classifying future and previously unseen seismic data. About 100 waveform snippets of short duration (4-12 secs) are extracted from 1 week of waveform data for training both the signal and noise classes. On a separate test-set we measure (binary) classification accuracy, sensitivity and specificity. Moreover, when testing against unseen data in time we can confirm seasonal dependency of noise characteristics, calling for an adaptive adjustment of the noise class over time which is implemented in a sequential learning fashion. A major obstacle is however the limited comparability between our purely automatic station-level detector and the combined automatic network associator with subsequent manual inspection approach at the IDC. The improvements over SEL3 and LEB bulletins is therefore difficult to quantify without further effort. To allow for a controlled evaluation we generate a semi-synthetic data set from cutting and pasting real waveform data in between station-specific noise samples.