[S-CG60] Driving Solid Earth Science through Machine Learning

[SCG60-03] Development of anomaly detection using tensor decomposition: An approach based on time series variation data.

*Harigai Shunto¹, Hiroyuki Nagahama¹, Jun Muto¹ (1.Touhoku University)

Keywords:Machine learning(tensor decomposition), atmospheric radon concentration fluctuations, the 2011 off the Pacific coast of Tohoku Earthquake, the 1995 Hyogoken-Nanbu Earthquake

Tensor decomposition, proposed by Francis Llewellyn Hitchcock, is a method to represent a dimensional array called a tensor as a product of tensors or matrix vectors with simpler structures. This method has been used for complex multidimensional data analysis in neuroscience and chemistry due to its ability to handle multidimensional data well and its low rank. Furthermore, it is evolving day by day to handle more complex problems. However, tensor decomposition and data analysis using tensor decomposition have not yet been applied in the field of earth and planetary sciences. The purpose of this study is to apply this tensor decomposition to the field of earth and planetary sciences, and to develop a method for anomaly detection by tensor decomposition using time series variation data.As an application of tensor decomposition, this study developed an anomaly detection method for time series data using tensor decomposition. The anomaly detection method developed in this study was developed by focusing on the similarities between the property of latent variables in the bottleneck layer of the autoencoder to capture the main features of the data and the property of each factor matrix in the tensor decomposition to show the contribution to each dimension of the data and capture the main features of the data, and by referring to the anomaly detection method using autoencoders. The developed anomaly detection method was developed with reference to the auto-encoder-based anomaly detection method. The developed anomaly detection method successfully detected anomalies in periodic time series test data.For the adaptation of the developed anomaly detection method to the field of earth and planetary science, time series data of radon concentration in the atmosphere measured at Fukushima Medical University and Kobe Pharmaceutical University were used as the target data. The developed anomaly detection method was able to detect anomalous fluctuations in atmospheric radon concentrations prior to the 2011 off the Pacific coast of Tohoku Earthquake and the1995 Hyogoken-Nanbu Earthquake in both data sets.As future prospects, anomaly detection and interpretation of data with nonparametric properties in the background are now possible by developing an anomaly detection extraction method with more than two-dimensional parameters by devising the tensor assembly method and by giving interpretability to the factor matrix by tensor decomposition.