D’Hamar Agudelo-Moreno

Temporal super-resolution (SR) aims to reconstruct a high-resolution signal from a low-resolution observation. When hardware limits force low sampling rates, this problem becomes critical for non-stationary signals with abrupt transients and rapid spectral changes. This manuscript reports a deterministic case study using pretrained 1D convolutional models and deterministic evaluation on paired real, synthetic, and mixed EEG-like signals. A compact encoder–linear upsampler–refinement architecture is evaluated at 5× upsampling under four training regimes: synthetic-only, real-only, tuned-real, and mixed. Performance is assessed with Mean Squared Error (MSE), Mean Absolute Error (MAE), Normalized Mean Absolute Error (NMAE), Log Spectral Distance (LSD), and spectral correlation (SCORR). Across 12 model–dataset combinations, mixed-domain training yields the most robust cross-domain behavior, outperforming single-domain checkpoints on real and mixed evaluation subsets. These findings support the practical value of training corpus composition for temporal SR under distribution shift. A focused morphological event analysis further shows that reconstruction error concentrates at abrupt amplitude and frequency boundaries, confirming that these transient regions are the dominant local challenge. An exploratory hybrid wavelet–superlet pilot is also reported; it achieves competitive pointwise error on selected domains but exhibits a substantial spectral-fidelity gap, indicating that frequency-aware inputs alone do not guarantee spectral reconstruction without auxiliary spectral losses.

Temporal super-resolution (SR) aims to reconstruct a high-resolution signal from a low-resolution observation. When hardware limits force low sampling rates, this problem becomes critical for non-stationary signals with abrupt transients and rapid spectral changes. This manuscript reports a reproducible case study using pretrained 1D convolutional models and deterministic evaluation on paired real, synthetic, and mixed ECG-like signals. A compact encoder–linear upsampler–refinement architecture is evaluated at 5× upsampling under four training regimes: synthetic-only, real-only, tuned-real, and mixed. Performance is assessed with Mean Squared Error (MSE), Mean Absolute Error (MAE), Log Spectral Distance (LSD), and spectral correlation (SCORR). Across 12 model–dataset combinations, mixed-domain training yields the most robust cross-domain behavior, outperforming single-domain checkpoints on real and mixed evaluation subsets. These findings support the practical value of training-corpus composition for temporal SR under distribution shift. A focused morphological event analysis further shows that reconstruction error concentrates at abrupt amplitude and frequency boundaries, confirming that these transient regions are the dominant local challenge. An exploratory hybrid wavelet–superlet pilot is also reported; it achieves competitive pointwise error on selected domains but exhibits a substantial spectral-fidelity gap, indicating that frequency-aware inputs alone do not guarantee spectral reconstruction without auxiliary spectral losses.