Abstract <p>This paper proposes a novel method for describing earthquake waveforms based on the transformation of seismic waveforms to mesofluctuations (MSF). These cumulative MSF representations ignore high-frequency (fine-scale) oscillations but effectively preserve the waveform envelope and enable accurate detection of the onset, amplitude, and decay duration of P- and S-waves. These key features serve as the foundation for the architecture of the proposed neural network RSNet (Russian Seismic Net) designed for automatic detection of body wave onsets (P- and S-phases) using computed mesofluctuation metrics. In contrast to existing neural network models that rely on raw three-component seismograms, the proposed method operates on reduced and intermediate-scale (mesoscale) metric representations derived from mesofluctuation analysis – specifically, signal range and its first derivative. This approach enhances interpretability and robustness under noise, while reducing the dimensionality of the input data. The model was trained and validated using the STEAD database, which includes over 1 million labeled seismograms. The mean absolute error (MAE) of wave onset detection was 0.04 s for P-waves and 0.13 s for S-waves. When benchmarked against phase picks generated by the PhaseNet neural network, the MAE for S-waves decreased to 0.10 s. These results are comparable to the accuracy of the state-of-the-art neural-network models, while RSNet offers distinct advantages in interpretability and resilience to external noise factors.</p>