5:15 PM - 6:45 PM
[SSS04-P05] DiallelX: a fortran code to approximate network cross-correlation coefficients
Keywords:Event detection, Matched-Filter Analysis, Network Cross-correlation Coefficient
Network cross-correlation coefficient (NCC), a correlation between continuous and template waveforms averaged over multiple stations[Gibbons & Ringdal, 2006 GJI], has been widely employed to detect uncataloged seismic events. However, calculating NCCs for long-term analyses with many templates requires high computational costs. In the case of a small seismic event search, the dominant frequency of waveforms would be higher than 10 Hz, and significant downsampling to reduce computation time may not work. For example, Ross et al.[2019 Science] searched small events from 50 Hz, 10-year continuous records using 284,000 template events observed at least four stations per each, which was enabled by a huge and expensive GPU array. On the other hand, analyzing a medium-sized dataset using a reasonable CPU is still necessary for individual researchers. Although SEC-C[Senobari et al., 2019 SRL], based on the MASS algorithm[Mueen et al., 2016 IEEE ICDM], was developed for the requirement, parallelization could have been more efficient.
We developed DiallelX, a modern fortran code to approximate NCC values among continuous and many template waveforms. Instead of fully calculating cross-correlation functions (CFs), DiallelX calculates CFs between templates and fixed-length waveform segments extracted from the continuous record, where each segment overlaps each other so that at least one among neighboring segments includes significant wave packets. After calculating CFs in the Fourier domain, our algorithm can average them along all channels before the inverse fast Fourier transformation (IFFT) thanks to the fixed-length segmentation of the continuous record, which reduces the number of IFFT from the number of channels to only once. Moreover, DiallelX efficiently selects candidates for new events to reduce output data and overlooking.
Using a 12-core AMD CPU, the latest implementation of DiallelX completes calculations between a 100Hz, 24-hour continuous record, and 1,000 template events in 15 channels within 5 seconds; the memory requirement for the processing is at most 2GB. The total processing time for a larger dataset will linearly depend on the number of data (e.g., ∼30 minutes for a 1-year continuous record with the same conditions). Applying our algorithm to ~6TB continuous records in total obtained during multiple rock failure experiments enabled us to find ∼tenfold small events compared to conventional event detection in some experimental runs.
Thus, referring to several thousands of templates, we can now search small events from Terabyte-sized datasets. We will demonstrate the speed, parallelization efficiency, and accuracy of DiallelX compared to conventional schemes and welcome feedback from seismologists experiencing heavy NCC calculations.
We developed DiallelX, a modern fortran code to approximate NCC values among continuous and many template waveforms. Instead of fully calculating cross-correlation functions (CFs), DiallelX calculates CFs between templates and fixed-length waveform segments extracted from the continuous record, where each segment overlaps each other so that at least one among neighboring segments includes significant wave packets. After calculating CFs in the Fourier domain, our algorithm can average them along all channels before the inverse fast Fourier transformation (IFFT) thanks to the fixed-length segmentation of the continuous record, which reduces the number of IFFT from the number of channels to only once. Moreover, DiallelX efficiently selects candidates for new events to reduce output data and overlooking.
Using a 12-core AMD CPU, the latest implementation of DiallelX completes calculations between a 100Hz, 24-hour continuous record, and 1,000 template events in 15 channels within 5 seconds; the memory requirement for the processing is at most 2GB. The total processing time for a larger dataset will linearly depend on the number of data (e.g., ∼30 minutes for a 1-year continuous record with the same conditions). Applying our algorithm to ~6TB continuous records in total obtained during multiple rock failure experiments enabled us to find ∼tenfold small events compared to conventional event detection in some experimental runs.
Thus, referring to several thousands of templates, we can now search small events from Terabyte-sized datasets. We will demonstrate the speed, parallelization efficiency, and accuracy of DiallelX compared to conventional schemes and welcome feedback from seismologists experiencing heavy NCC calculations.