Analysis of the use of Recurrent Neural Network for Multiple Frequency-Shift Keying signals decoding during Entry, Descent, and Landing phase of interplanetary missions

This paper presents a novel deep learning architecture for robust telemetry decoding during the Entry, Descent, and Landing (EDL) phase of planetary missions, where communication is challenged by rapid Doppler dynamics, trajectory-induced frequency shifts, and significant additive noise. We introduce the Deep Linear Recurrent Unit (DLRU), a neural network model that integrates non-linear multi-layer processing with linear recurrent units for efficient temporal classification of telemetry signals. Focusing on Multi-Frequency Shift Keying (MFSK) signals — common in Mars missions such as NASA’s Curiosity and Perseverance — the DLRU is designed to handle partially Doppler-compensated signals with residual frequency uncertainties of ±45 Hz and Doppler rate uncertainties of ±16.67 Hz/s. The model architecture comprises a linear encoder, skip-connected recurrent units for stable learning, temporal pooling, and a classifier trained using cross-entropy loss and different kinds of optimizers. A custom Python/JAX-based dataset generator was developed to simulate realistic EDL scenarios with adjustable noise, Doppler effects, and signal parameters. Upon evaluation on 8-tone MFSK signals with short-time Fourier transform (STFT) inputs, the DLRU achieved a classification accuracy of 66% under nominal Mars EDL conditions while maintaining real-time computational feasibility. The results highlight the potential of linear recurrent deep learning models to enhance the robustness of telemetry decoding in extreme space communication environments.