This paper investigates the application of classical machine learning algorithms for signal demodulation in a high-frequency communication channel described by the Watterson model. The relevance of this task stems from the need to improve noise immunity during broadband data transmission under conditions of multipath propagation and Doppler distortions. The classifiers used include Random Forest, Decision Tree, and k-nearest neighbors (KNN). The models are trained on synthesized data generated in the MATLAB Simulink environment, with varying path delays (0–4 ms) and Doppler shifts (0–10 Hz). The input features comprise the coordinates of signal constellations as well as the history of received symbols, since intersymbol interference significantly affects the position of the current point. The experiments revealed a critical feature: demodulation algorithms exhibit high sensitivity to variations in Doppler shift. A deviation of just 0.1 Hz from the training set parameters alters the structure of the signal constellation, increasing the bit error rate (BER) to 0.5. When the Doppler parameters align with the training grid, the Random Forest algorithm demonstrates the best performance, achieving a BER <0.01 on the full dataset. To address this sensitivity, a local training method with a reduced Doppler shift sampling step (down to 0.1 Hz) is proposed, enabling a BER ≤0.01 for Random Forest. Due to the exponential growth of the training dataset size (up to 1000 GB), data reduction techniques were developed: excluding time parameters and sampling data based on possible combinations of preceding bits. This reduced the data volume by a factor of 31.5 while maintaining a BER <10% for Random Forest with a training step of 0.2 Hz. The study concludes that the Random Forest classifier is suitable for demodulation in HF channels and highlights the necessity of adapting the training step to the required Doppler shift accuracy.
