The present report shows a possible method to convert data sampled at non-equidistant intervals into equidistant time series. The method is based on the use of neural networks and is equivalent to a non-linear interpolation procedure. From the existing data a statistical model of the time series consisting of two back-propagation networks and one self-organizing map is constructed. The model is used then to reconstruct the time series at arbitrary intervals. The method may be used for reconstruction of both deterministic and pseudo-indeterministic data.
A possible area of application: reconstruction of space data corrupted by unfavourable transmission conditions.
IRF Scientific Report 235
December 1996.