Abstract: The K-L transformation uses the coherence of adjacent seismic traces to remove the random noise. But the de-noise effect is not good for sloping and bending event reflection. Although the advanced time-variation dip sweep stack K-L transformation can remove random noises, the characteristics of effective signals and random noises in frequency domain is not taken into consideration, making the high-frequency effective signals lost. Because the wavelet transformation has a good ability in time-frequency analysis, the K-L transformation in wavelet domain can remove noises in time and frequency domain separately. The principles of suppressing random noises with K-L transformation in wavelet conversion domain are that 1) wavelet decomposition is carried out on seismic signals to form wavelet packet sections in time sharing and frequency division, 2) K-L transformation is utilized to remove noises on sections, and 3) the de-noise wavelet packet sections are reconstituted into seismic sections to remove random noises. The theoretical model computation and actual data processing show that the K-L transformation in wavelet conversion domain can remove random noises effectively and preserve the effective high-frequency signals simultaneously.