GlitchNet: A fast SEIS VBB glitch detection and removal algorithm based on deep learning
Abstract
As the main instrumentation program of the InSight mission, SEIS (Seismic Experiment for Interior Structure) is the first seismic recording system deployed on Mars surface with two main seismometers: a very broadband (VBB) one and a short-period (SP) one. So far, SEIS has been in operation for more than two years, recording hundreds of marsquakes. Besides marsquakes, SEIS records contain various kinds of non-seismic signals or anomalies. Glitches are one of the most frequent types of anomalies in the VBB records, although they also appear in SP records. Glitches can be modeled as the response of the instrument to a step-function in acceleration, i.e. as one-side pulses in the raw data. Glitches have a typical duration of ~ 25 s no matter their amplitudes. They seriously contaminate the records and must be removed before seismological analyses, such as receiver function estimation, ambient noise correlation and marsquake studies, can be conducted. Various techniques, including MPS and UCLA, have been developed for the detection and removal of glitches. But all suffer from problems of low efficiency and reliance on the choice of multiple parameters. Here, we propose an algorithm named GlitchNet, which is based on deep learning with a 5-layer convolutional neural network (CNN) for glitch detection and an autoencoder (AE) for glitch reconstruction and removal. Results from our GlitchNet show great consistency with manual results during nighttime, and our removal operations induce less direct current (DC) component in the deglitched waveforms. Our algorithm is much more efficient than existing methods such as MPS and UCLA (20 x UCLA and 150 x MPS), providing an efficient and reliable alternative to the long-term data analysis of SEIS VBB records.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.P15C2118X