Methodolgy

In order to detect and identify a wide variety of seismic transient and semi-continuous volcano-seismic event types (i.e. volcanic tremor) several methods will be combined for the sake of robustness and redundancy. Classical trigger and waveform matching algorithms build a base system as those techniques are robust, widely spread and well understood in their working principles. Beyond these traditional detection algorithms we will develop a HMM-based system as central detection/classification module. In recent years, first HMM recognition techniques have been successfully employed in seismic event classification for seismic signals of volcanic origin. The modeling of the time sequence of short term signal characteristics from individual or combined data streams by such a stochastic process allow for robust classification in difficult noise environments.

Due to the particular nature of the task force activity, a major difficulty for classification algorithms which are based on supervised learning principles is the potentially lack of prior knowledge about signal characteristics of particular seismic events or the shortcoming of waveform training data. The system under development therefore needs to be able to start from scratch, i.e. contain the capability to record and build its own knowledge base and waveform repository with small temporal delay and during the very first recording phase of the network. Only then, the characteristic features of the relevant seismic activity (recorded at the actual network configuration) can be used for the detection/classification task.

Goals and Workplan

Thus, one main task in this sub-project is to focus on the development of a fast and reliable adaptive training strategy for the classification system allowing for operator interaction to the automatic system and re-scanning for interesting waveform data in the recorded archive data (HMM-Spotting).

The combined evaluation of classical trigger algorithms (STA/LTA, AR), cross-correlation master event matching and stochastic modeling of short term feature maps (HMM-spotting) will result in detection messages sent out to the physical model building software modules. These detection messages may also be used to alert other systems in a multi-sensor network. It seems to be feasible to adapt the described approaches also to the continuous deformation data. However, signal characteristics differ compared to seismic wave data and therefore the detection of deformation transients has to be considered as an experimental add-on to this package.

Back to WP5: Physical model