ABSTRACT
Acceleration response spectra (ARS) for mining tremors in the Upper Silesian Coalfield are generated using artificial neural networks trained by means of Kalman filtering. The target ARS were computed on the basis of measured accelerograms. It was proved that the recurrent layered network, trained by the recurrent decoupled extended kalman filter (RDEFK) algorithm is numerically much more efficient than the standard feed-forward NN learnt by DEKF. It is also shown that the considered KF algorithms are better than the traditional Rprop learning method.