Parameter-Efficient Time–Frequency Temporal Convolution with Log-Percentile Normalization for Underwater Acoustic MAC Protocol Recognition

Abstract

Passive recognition of medium access control (MAC) protocols allows an underwater acoustic monitoring node to infer channel-access behavior without decoded control headers or cooperation from the observed network. This study evaluates a parameter-efficient time–frequency temporal convolutional network (RTF-TCN) using exclusively simulated clean waveforms corrupted by independently generated Gaussian or symmetric alpha-stable noise. The processing chain is fully specified from the clean sig arrays through MATLAB’s power-spectral-density output of spectrogram, temporal resampling, cropping, log-percentile normalization, and model evaluation. To reduce leakage from shared simulation geometry, the new experiments use topology group-wise train/validation/test splits rather than the sample-wise split used by the inherited benchmark. Across five seeds, in-domain performance remained stable over the Gaussian −5 to +5 dB range. At 0 dB for the study-defined scale-based signal-to-noise measure (scale-GSNR), performance remained high at alpha = 1.8 and 1.7, became unstable at alpha = 1.6 (70.75% ± 18.66%), and approached the balanced five-class performance floor at alpha = 1.5 and 1.2. Checkpoints trained at alpha = 1.8 also degraded when directly transferred to heavier-tailed conditions. An inherited sample-wise ablation found higher Macro-F1 for a standard 3 × 3 frontend, but that variant used 3.5× as many parameters as the asymmetric frontend. RTF-TCN contains 158,149 parameters and has profiled costs of 289.08 million multiply-accumulate operations and 578.16 million floating-point operations for one 1 × 1 × 100 × 580 input. The results support parameter efficiency within the tested conditions, but they do not establish generalization to measured sea data.