An unsupervised learning method for the automatic classification of earthquake and noise signals recorded in continuous waveforms

For the stable operation of an earthquake early warning (EEW) system and the reduction of false alarm issuance due to noise signals, it is important to monitor the condition of seismometers and to check what signals are recorded in their waveforms on a regular basis. On the other hand, in terms of timely warning issuance, a large number of seismometers need to be incorporated into the system. Under such a situation, checking every record by hand would be unrealistic. In this study, we developed a machine learning algorithm to automatically detect and classify earthquake and noise signals recorded in continuous waveforms to assist the seismometer monitoring.

An unsupervised learning technique is employed in our method since noise signals are different depending on individual seismometers. The method is divided into four steps.

- Feature extraction: Features are obtained from an acceleration record by computing running spectra (with a 4-s window and 0.1-s interval) and applying a triangular filter bank (10 bands with a log scale).

- Clustering in the frequency domain: The features are clustered by the k-means algorithm (#cluster = 100).

- Clustering in the time domain: With the assumption of a Markov model with 100 states, a transition matrix is calculated. An undirected weighted graph is constructed based on the matrix, and the spectral clustering algorithm (#cluster = 10) is performed. The 10 classes represent time series models of earthquake or noise signals.

- Classification: Every time window in the record is assigned into the class to which it belongs.

We applied the method to a 24-h waveform (00:00 to 24:00 on March 1, 2017) recorded in Station E.JDJM (a seismometer of the Metropolitan Seismic Observation network (MeSO-net) in Japan), whose records are frequently contaminated with small noise signals by trains. The record was classified as follows: background noise (nighttime) -> Class 2, background noise (daytime) -> Class 8, train noise (nighttime) -> Class 5, train noise (daytime) -> Class 10, earthquake -> Classes 1, 3, 4, 6, and 7 (see the figure). This indicates that our method successfully classified earthquake and various noise signals without prior knowledge and will contribute to the seismometer monitoring.