[MTT52-08] Identification of infrasound waveforms from Sakurajima volcanic eruption using a machine learning algorithm
Keywords:infrasound, machine learning, MFCC
1. Introduction: Infrasound defined as the sound wave below the human audible lower frequency limit of 20 Hz are known to arisen from large-scale natural disaster-inducing events such as volcanic eruptions and tsunamis. Infrasound is generated by such events have the property of being able to propagate over long distances, so they are expected to be used for disaster prevention and mitigation against large-scale natural disasters. The Kochi University of Technology has jointly developed a combined infrasound tsunami sensor with SAYA Inc., and a total of 16 sensors are arrayed along the coast of Kochi Prefecture to estimate the direction of arrival and energy of tsunami. Previous researches[1] have developed a system for automatic detection of N shaped waveforms. But it is still impossible to identify the phenomenon that caused the N shaped waveforms. Therefore, we proposed a system to classify volcanic eruption waveforms and noise waveforms using a machine learning algorithm on supervised learning. The purpose of this study is to develop an automatic alert generation system triggered by a large-scale natural disaster using the machine learning algorithm.
2. Feature extraction by MFCC and binary classification by SVM: In this study, feature vectors were constructed from waveform data with a sampling rate of 2 Hz and a data length of 1024 (about 8.5 minutes), and volcanic eruption waveforms and noise waveforms were classified using a machine learning algorithm. The feature vectors were composed of two methods, Mel Frequency Cepstrum Coefficient (MFCC) and delta cepstrum, totaling 24 dimensions. MFCC is a method of compressing the logarythmic power spectrum by dense vocal tract components and sparse fundamental frequency components, and is often used in the field of human speech recognition. Delta cepstrum is a method that uses non-stationary frequency components as feature values. After feature extraction, binary classification was performed using linear Support Vector Machine (SVM), which is a learning method of a linear classifier, using scikit-learn, one of the machine learning libraries.
3. Results of the experiment: Since the purpose of this study is to detect the volcanic eruption waveform in real time with each installed sensor, the data set for supervised learning was configured for each of 16 infrasound sensors. Among them, the volcanic eruption waveform dataset of the best observation point that was judged to have the highest score in the all of dataset was accumulated in Ukibuchi, Kuroshio-cho, Kochi Prefecture, and the accuracy, recall, precision, F-measure were 95±2%, 96±3%, 93±3%, 95±2% respectively. However, the number of the data set for there was N=58. Here, the number of datasets was evaluated using the Speech Commands Dataset (© TensorFlow team, AIY team, 2017), which contains speech of English words with sufficient statistics. As a result, since the number of data sets was required to be at least N=146. N=58 of there did not reach that value, and it becomes an issue to further increase of the dataset number N.
4. Summary: In this study, basic research on real-time detection of volcanic eruption waveforms was performed using proposed machine learning algorithms. At this time, the feature vector was composed of a total of 24 dimensions of MFCC and delta cepstrum, and SVM was selected as the machine learning algorithm for binary classification. As a result, a method to create a high-quality dataset was established, and binary classification of volcanic eruption waveform and noise waveform using remote sensing technology have become possible. In the future, we aim to construct a system for automatically generating alerts for remotely sensed volcanic eruption waveforms using the method of this study.
[1] Sorimachi, R., Optimization method to estimate wave source positions for widely distributed infrasound sensors, Kochi University of Technology 2017 Master Thesis, 2018
2. Feature extraction by MFCC and binary classification by SVM: In this study, feature vectors were constructed from waveform data with a sampling rate of 2 Hz and a data length of 1024 (about 8.5 minutes), and volcanic eruption waveforms and noise waveforms were classified using a machine learning algorithm. The feature vectors were composed of two methods, Mel Frequency Cepstrum Coefficient (MFCC) and delta cepstrum, totaling 24 dimensions. MFCC is a method of compressing the logarythmic power spectrum by dense vocal tract components and sparse fundamental frequency components, and is often used in the field of human speech recognition. Delta cepstrum is a method that uses non-stationary frequency components as feature values. After feature extraction, binary classification was performed using linear Support Vector Machine (SVM), which is a learning method of a linear classifier, using scikit-learn, one of the machine learning libraries.
3. Results of the experiment: Since the purpose of this study is to detect the volcanic eruption waveform in real time with each installed sensor, the data set for supervised learning was configured for each of 16 infrasound sensors. Among them, the volcanic eruption waveform dataset of the best observation point that was judged to have the highest score in the all of dataset was accumulated in Ukibuchi, Kuroshio-cho, Kochi Prefecture, and the accuracy, recall, precision, F-measure were 95±2%, 96±3%, 93±3%, 95±2% respectively. However, the number of the data set for there was N=58. Here, the number of datasets was evaluated using the Speech Commands Dataset (© TensorFlow team, AIY team, 2017), which contains speech of English words with sufficient statistics. As a result, since the number of data sets was required to be at least N=146. N=58 of there did not reach that value, and it becomes an issue to further increase of the dataset number N.
4. Summary: In this study, basic research on real-time detection of volcanic eruption waveforms was performed using proposed machine learning algorithms. At this time, the feature vector was composed of a total of 24 dimensions of MFCC and delta cepstrum, and SVM was selected as the machine learning algorithm for binary classification. As a result, a method to create a high-quality dataset was established, and binary classification of volcanic eruption waveform and noise waveform using remote sensing technology have become possible. In the future, we aim to construct a system for automatically generating alerts for remotely sensed volcanic eruption waveforms using the method of this study.
[1] Sorimachi, R., Optimization method to estimate wave source positions for widely distributed infrasound sensors, Kochi University of Technology 2017 Master Thesis, 2018