Neural Network Based Electric Field Pattern Recognition for Earthquake Prediction

This study concerns monopolar electric field measurement (MEFM) which is assumed to be related to change in fault stress and evaluation of patterns shaped by acquired data from different locations. Patterns are constructed at a central computer that receives on line data packages from 11 stations through FTP over the internet in Marmara Region (northwestern Turkey). Probable precursory patterns of MEFM data is required to be recognized inside its’ nonperiodic component. A self organizing neural network is used for recognition of predetermined anomaly patterns as priory knowledge in the model based approach. In this work, the mentioned learning and prediction process has been used to forecast earthquakes. Hebbian based principal components analysis has been used in preprocessing, before the classification in a multilayer perceptron. On the other hand a self learning mechanism is used to construct relations where the model is accepted as uncertain and the feature classification of the occurred earthquakes are stored beside the measured data strings. If a correlation exists between the measurement and the events then the outputs begin to converge to physical properties of the forthcoming events. A multilayer perceptron consisting of one input layer, two hidden layers and one output layer that uses back propagation learning algorithm is applied to MEFM data for earthquake prediction.