Compum & Geo.scicn~~~.~ Vol. 2 I. No. 2. pp. 279--28X, 1995 Elsevier Science Ltd. Prinkd in Great Britain 009%3004(94)00072-7 zyxwvutsrqponmlkjihgfedcbaZYXWVUTS NEURAL NETWORKS AND DISCRIMINATION OF SEISMIC SIGNALS GIOVANNI ROMEO, FRANCESCO MELE, and ANDREA MORELLI Istituto Nazionale di Geofisica, via di Vigna Murata 605, 00143 Roma. Italy zyxwvutsrqponmlkjihg e- mail: romeo@in8800.ingrin.it (Received 8 August 1993; accepted 18 March 1994) Abstract-Recent developments in algorithms and computer architecture make neural networks a useful tool in designing pattern recognition systems. We show how a simple multilayer perceptron with 23 neurons can be trained easily and used to classify seismic signals. Applied to broadband seismic signal, the perceptron permitted the recognition of different types of events on the basis of their frequency. Applied to a real-time, automatic, seismic data acquisition system, it saved more than 50% CPU time in a detection procedure. Key Words: Error back-propagation, Multilayer perceptron, Neural network, Perceptron, Seismology, Seismic network. INTRODUCTION A broad variety of problems seems to be easily solved manually or visually by a human operator but difficult to tackle by computer. It may be easy to train an operator to recognize images of objects, shapes, or patterns in a signal, but the same problem cannot be automated by computer: earthquake detection and classification belong to this class of problems (Dowla, Taylor, and Anderson, 1990). A possible approach may lie in trying to reproduce the mechanics of human reasoning. In the brain different areas are specialized to perform different functions, all sharing the same structure: a network of cells densely inter- connected that are termed neurons (see Appendix). Simulation of neural networks by computers or other electronic equipment may be used to solve problems which require an approach similar to human reason- ing and make it easier to solve classical but complex problems (R&h and Tarantola, 1994). DISCRIMINATION OF SEISMIC SIGNALS We describe an experiment consisting of training a small multilayer, fully connected neural network, usually termed perceptron (23 neurons, Fig. 1) to recognize seismic signals recorded by the seismo- graphic stations of the MEDNET network (Boschi, Giardini, and Morelli, 1991). The records consist of high dynamic range digital signals (24 bits per sample) covering a broad frequency brand (8-0.003 Hz). The sampling frequency is 20 Hz. Net- work MEDNET consists of 12 stations distributed around the Mediterranean region which record in a continuous mode. Because of the bulk of incoming data, the analysis of the signal must necessarily rely on automated procedures. The seismographic records used for this study come from an initial selection identified by a detec- tion algorithm as containing nonstationary ‘events’ to be classified. Each of these events may be the result of an actual seismic signal, or of local disturbance inducing ground tremor, or of instrumental effects. An initial detection algorithm monitors variations in time of the spectral ratio between the current signal and the background noise. Spectra are computed for successive, nonoverlapping intervals, 6.4 set long. The algorithm detects an event if the spectral ratio exceeds a predefined threshold in one of several frequency bands in which the spectrum is divided. This detection algorithm is not capable of discrim- inating among the different possible causes for the amplitude variation between seismic and nonseismic sources. For this reason, a definition of efficiency for such an algorithm is never really complete, as it should include false alarms-transient nonseismic signals-that the algorithm cannot distinguish. It is a general feature of such detection algorithms that an attempt to limit the number of missing seismic events results in an unwanted, conspicuous increase in the number of false alarms. The actual discrimination between different classes of events must be performed either by an operator or by an automatic algorithm able to recognize patterns. With the hypothesis that the shape of the spectrum is a diagnostic element for discrimination, it is possible to use this information as an input signal to a neural network trained to class- ify signals in one of a number of predefined classes. For constructing the training set, several signal samples were examined by an operator, assigned to 279