This study is devoted to identification of low amplitude components from ECG signals by different time-frequency analysis methods when main power spectrum falls on high-amplitude components. It was also analyzed the problem of choosing correct scale system for determination low-amplitude components on the scalogram by artificial intelligence models. As a result of the study, several modifications of the continuous wavelet transform were proposed. First modification was based on the use of a scaling function and a modified wavelet. Second modification was based on the use of cosine similarity at each iteration of convolution followed by the use of a scaling function. The main idea of the study was to modify the wavelet transform in such a way as to select the components which has the target amplitude and reduce all other components that complicate the neural networks analysis of the interested fragments of the signal. Also, possible procedures for signal restoring were proposed for preserving the effect of using scaling modifications. The testing of the proposed modified algorithms was carried out on the basis of artificially created signals as well as on the basis of real ECG signals with late potentials superimposed on them. Visual analysis of scalograms and signal reconstructions obtained using the modified wavelet transform showed that the modified wavelet transform is capable of extracting low-amplitude components from the signal with much greater spectral power than the transform without modifications. In addition, the ability of common neural network models to distinguish between cardiac cycles with and without late potentials was tested. As a result, it was found that models that were trained on scalograms obtained using a modified wavelet transform train faster and are less susceptible to local minima stucking. The results of classification of signals with and without late potentials based on trained neural network models showed that training using scalograms obtained on the base of a modified wavelet transform allows achieving 99% classification accuracy, which is 1-49% more than that using scalograms obtained on the base of on the classical wavelet transform.