Blind Source Separation in CTBTO Expert Technical Analysis Procedures

Abstract:

Blind Source Separation (BSS) is a widely used technique in many branches of data processing, but not in CTBT related applications so far. BSS methods are attractive for they use minimal a priori information about the signals they are dealing with. Homomorphic deconvolution and cepstral smoothing are probably the only methods not widely used in CTBT that can be attributed to given technique. However Expert Technical Analysis (ETA) in CTBTO may face the problems which cannot be resolved with only certified CTBTO applications and may demand specific techniques not presently used in a practice. The case which has to be considered within the ETA framework is the unambiguous separation of signals with close arrival times. There are two scenarios of interest: (1) separation of two almost co-located explosions conducted within fractions of seconds, and (2) extraction of explosion signals merged with wavetrains from strong earthquake. The importance of resolving the problem related to case 1 is obvious since it’s connected with the correct explosion yield estimation. Case 2 is a well-known scenario of conducting clandestine nuclear tests. These cases can hardly be resolved with the conventional methods implemented at the International Data Centre, especially if the signals have close slowness and azimuth. The approach we use here is applying one of the blind source separation methods - Independent Component Analysis, implying non-Gaussianity of the signal’s mixture underlying processes. We have tested the technique with synthetic data and Monte-Carlo modelling, and with the data from three DPRK tests and mining explosions conducted in Central Russia. The data was recorded by the International Monitoring System of CTBTO and by small-aperture seismic array Mikhnevo (MHVAR) operated by the Institute of Dynamics of Geospheres, Russian Academy of Science. The approach demonstrated good ability of separating sources conducted practically simultaneously and/or having close arrival times and recovering their waveforms from the mixture. Perspectives and limitations of the method are discussed.