11:30 AM - 11:45 AM
[SCG55-04] Development in CNN-based automatic detection of traveltime in acoustic signal obtained by GNSS-acoustic survey
Keywords:GNSS-Acoustic observation, machine learning, convolutional neural network
In the region where oceanic plates subduct beneath continental plates, as in the Japan Trench, megathrust earthquakes often occur, and the resulting tsunamis can cause major disasters. In order to understand and evaluate the risk of these phenomena, it is important to observe crustal deformation on the seafloor, around the source regions of earthquakes. However, in contrast to land, where a network of GNSS reference sites has been established, a special method is required to observe crustal deformation on the seafloor. GNSS-Acoustic observation is one of the methods, which determines centimeter-level global coordinates of a seafloor reference point by combining GNSS positioning, which determines the global coordinates of a ship on the sea surface, and undersea acoustic ranging, which uses sound waves instead of GNSS radio waves, to determine the relative position between a ship and a seafloor reference point. In acoustic ranging, multiple seafloor stations that configure the seafloor reference point receive and record sound waves transmitted from a ship, and played back to the ship. By receiving them on the ship, the distance between the ship and the seafloor reference point can be determined from the two-way travel time of the sound waves. The position of the seafloor reference station is estimated by repeating the measurements at one-minute intervals while changing the ship's position and integrating the data for about one day. Acoustic ranging, which can reach up to 10 km, requires the use of ultrasonic waves with a relatively low frequency of 10 kHz, which is about 15 cm in wavelength, which has little absorption attenuation in seawater. In order to measure the round-trip travel time with an accuracy of 0.01 ms, which is equivalent to 1/10 of the wavelength, the acoustic signal is designed to have a clear peak with few sidelobes on the cross-correlation function (correlogram) between the transmitted and received waves by modulating phase of a 10 kHz carrier wave with pseudo-random timing according to the M-sequence. However, the actual shape of the correlogram has many sidelobes due to reflected waves so in most cases, the maximum peak is not a true two-way travel time. To solve this problem, now a skilled person selects the correct peak among the sidelobes, defines it as a template, and then automatically detects the two-way travel time by matching process. However, templates need to be created for each incident angle and distance, as a function of the positional relationship between the ship and the seafloor reference station, and the results may differ depending on the creator. In addition, template matching sometimes doesn’t work well when there were large variations in waveforms. Therefore, this study aimed to create an algorithm to automatically determine the two-way travel time of sound waves from correlograms by using a Convolutional Neural Network (CNN), which is one of the machine learning.
In this study, we used data obtained from a total of six campaign surveys conducted from 2020 to 2022 at 18 stations along off the coast of Sanriku. For convenience, the analysis was performed by expressing the one wavelength of correlograms in terms of 8 samples. Correlogram data of 120 samples each before and after the maximum peaks (data length: 241 samples = 241 × 10-5 s), normalized by the maximum correlation value, were used as input values for the CNN. The labels ware the position of the peak determined by the conventional method. 80% of the data obtained from the first four observations was used to train the CNN and 20% was used to verify the accuracy of the CNN. In addition, we examined the results of inputting the data obtained from the two recent observations into the CNN that had been trained.
The results of inputting the validation data into the trained CNN showed that the root mean squares of the difference between the conventional method and the CNN was 2.10 samples (= 2.10 x 10-5 s). Namely, the CNN showed almost the same results as the conventional method. The results of the CNN with the data obtained from the two recent observations were also almost the same for the majority of the data, but some data differed significantly from the conventional method. We present which results of the CNN or the conventional method is correct, and the reasons for the large differences in results.
In this study, we used data obtained from a total of six campaign surveys conducted from 2020 to 2022 at 18 stations along off the coast of Sanriku. For convenience, the analysis was performed by expressing the one wavelength of correlograms in terms of 8 samples. Correlogram data of 120 samples each before and after the maximum peaks (data length: 241 samples = 241 × 10-5 s), normalized by the maximum correlation value, were used as input values for the CNN. The labels ware the position of the peak determined by the conventional method. 80% of the data obtained from the first four observations was used to train the CNN and 20% was used to verify the accuracy of the CNN. In addition, we examined the results of inputting the data obtained from the two recent observations into the CNN that had been trained.
The results of inputting the validation data into the trained CNN showed that the root mean squares of the difference between the conventional method and the CNN was 2.10 samples (= 2.10 x 10-5 s). Namely, the CNN showed almost the same results as the conventional method. The results of the CNN with the data obtained from the two recent observations were also almost the same for the majority of the data, but some data differed significantly from the conventional method. We present which results of the CNN or the conventional method is correct, and the reasons for the large differences in results.