1 Introduction

The search for periodicities in time or spatial dependent signals is a topic of the uttermost relevance in many fields of research, from geology (stratigraphy, seismology, etc.; (Brescia et al. 1996)) to astronomy (Barone et al. 1994) where it finds wide application in the study of light curves of variable stars, AGN's, etc.

The more sensitive instrumentation and observational techniques become, the more frequently we find variable signals in time domain that previously were believed to be constant. Research for possible periodicities in the signal curves is then a natural consequence, when not an important issue. One of the most relevant problems concerning the techniques of periodic signal analysis is the way in which data are collected: many time series are collected at unevenly sampling rate. We have two types of problems related either to unknown fundamental period of the data, or their unknown multiple periodicities. Typical cases, for instance in Astronomy, are the determination of the fundamental period of eclipsing binaries both of light and radial velocity curves, or the multiple periodicities determination of ligth curves of pulsating stars. The difficulty arising from unevenly spaced data is rather obvious and many attempts have been made to solve the problem in a more or less satisfactory way. In this paper we will propose a new way to approach the problem using neural networks, that seems to work satisfactory well and seems to overcome most of the problems encountered when dealing with unevenly sampled data.