Reduced dimensionality representation of strong ground motion records as a superposition of a relatively small number of pulses is studied. Such representation is obtained by the expansion of velocity in orthogonal wavelet series using the Fast Wavelet Transform, and approximation by only the largest energy terms in the series. The Coiflet 5 wavelet family is used, which is orthogonal, smooth and nearly symmetric. The goodness of the approximation is examined for the EQUINFOS for USA Part I database, as representative of a large variety of strong motion records (it consists of 494 three-component records from 106 earthquakes recorded in the western US between 1933 and 1984). The goodness of fit is measured in terms of closeness of predicting several input and output characteristics of a nonlinear oscillator representing a structure: (i) energy of the input ground motion (proportional to integral of velocity squared), (ii) the peak input power, and (iii) the time of collapse of a bi-linear oscillator (considering also collapse due to dynamic instability). The results show very high degree of correlation of these characteristics as estimated from the actual record and from its approximations, even for small number of pulses (relative to the number that would represent the ground velocity exactly). Such reduced representation of strong ground motion is useful for extracting such pulses from strong motion records to study their nature, and for development of new algorithms for the synthesis of artificial earthquake strong motion records.