Special preprocessing techniques for denoising quasi-stationary signals (SN) acquired from the secondary winding of the excitation transformer in a power plant are addressed. Sequences of 30 periods were analyzed. SN were polluted by white Gaussian noise and “average” signals (AS) of one period length were deduced. Because vectors of samples were handled, corrective measures (CM) were applied firstly. They were conceived to diminish the numerical errors generated by fractional number of samples per period. Unlike the quasi-sinusoidal voltages, the currents contain almost horizontal segments changing their magnitude in sharp transitions (ST) due to thyristors. Therefore different criteria and thresholds were used to mark the “not enough stationary” periods as terms not involved in computing AS. Areas of individual periods and, only for voltages, maximum deviations allowed for magnitudes were used. Having AS, 3 sets of periods (1-st, middle and last) were afterward used (now applying CM over AS). Different (currents vs voltages) thresholds based techniques were afterward conceived and used in order to provide immunity against transients and ST effects to the estimated noises (EN=SN-AS). Single and multi-level Wavelet based denoising techniques were finally applied to obtain denoised signals (SD). Root Mean Square Deviations associated to the pair (SD,AS), powers of noises and Signal to Noise Ratios, evaluated in different ways, were used to select the most suitable denoising techniques.