A STUDY ON ANOMALOUS SIGNAL DETECTION USING HMM FOR ELF ELECTROMAGNETIC WAVE Yoshinao ITO ∗ , Akitoshi ITAI ∗ , Hiroshi YASUKAWA ∗ , Ichi TAKUMI † and Masayasu HATA ‡ Aichi Prefectural University ∗ , Nagoya Institute of Technology † , Chubu University ‡ Faculty of Information Science and Technology ∗ , Faculty of Engineering † , Collage of Engineering ‡ Ibaragabasama Kumabari Nagakute-cho Aichi ∗ , Gokiso-cho showa-ku Nagoya-shi Aichi † , Matsumoto-cho Kasugai-shi Aichi ‡ 1. INTRODUCTION Anomalous radiations of EM waves due to an earth diastrophism has been recorded in advance of earthquakes and volcanic activities[1]. We have been measuring the EM wave radiation in the ELF band. Our research is directed towards identifying an anomalous radiation of earthquakes from the EM wave data[2]. Observed signals contain undesired components associated with the magnetosphere, the ionized layer and the lightning radiation in the tropics, and so on[3]. Various signal processing techniques have been proposed to detect and understand the anomalous radiation in the ELF band. The normal value method[4] and the principal component analysis[5] is proposed as the simple anomalous signal detection. These methods require the observation signal of several years at the same observation point. It is difficult to detect the anoma- lous signal in a new observation point. The neural network[6] is applied to overcome the weakness of conventional methods. The neural network approach does not require the observed signal recorded over several years at the same observation point. However, in order to achieve the accurate detection, many anomalous signals corresponding to the great earthquake are neces- sary for the training data set. The anomalous signal detection using a linear prediction error can detect seismic signal without anomalous signals. This technique detects abrupt noises as anomalous signals. Requirements for an anomalous signal detection are outlined below: (i) An anomalous signal can be detected from the data observed at various site. (ii) Decrease the number of anomalous signal as training signal. (iii) Decrease the false detection due to an abrupt noise. In this paper, the HMM is applied as the anomalous signal detection satisfying the above requirements. The HMM input signal the amplitude density distribution calculated from the waveform of the EM wave data excluding the anomalous signal. The training data is observed at various seasons and observation points. The observation signal including the abrupt noise is also used as training data to avoid the false detection. Results of the anomalous signal detection will indicate different characteristics when the display scale of the waveform is changed. The number of states of HMM influences the anomalous signal detection accuracy. This paper represents the optimal display scale of the image and the number of state of HMM, and shows possibility anomalous signal can be detected before earthquake occurred. 2. ELECTROMAGNETIC WAVE IN ELF BAND We observed the EM wave radiation in the ELF band (223Hz) as represented by the east-west, north-south, and vertical magnetic field components at about forty observation stations in Japan. Collected data is averaged over 6 seconds interval (14400 points per day) at each station and direction. The EM wave data averaged over 6 and 150 seconds interval is recorded on the data logger established in Nagoya Institute of Technology. The data server provides us the numerical data and its graphical image.