A Lightweight Grouped Fully Connected Network for Single-Interference Recognition from Time-Frequency Data

Interference recognition from time–frequency images requires lightweight models that can maintain recognition performance when signal parameters deviate from their nominal settings. We propose ASD-Net, a decreasing-group fully connected architecture that progressively merges feature blocks while controlling parameter growth. After development-set selection, the fixed architecture was evaluated on held-out single-channel time–frequency data generated under nominal and parameter-shifted conditions. Across ten training seeds, ASD-Net achieved 94.29±0.82% accuracy under nominal conditions on the fixed jamming-to-signal ratio (JSR) grid and 77.88±2.24% under independently sampled parameter perturbations, with further tests showing sustained performance as the parameter deviations increased beyond the predefined range. The model contains only 25,014 parameters and requires 17,888 multiply–accumulate operations (MACs) per sample. Compared with a same-width unrestricted Dense architecture, grouped connectivity reduced the parameter count by a factor of approximately 179 while improving parameter-perturbed accuracy by 14.51 percentage points. These results demonstrate that the proposed decreasing-group hierarchy combines compact computation with competitive recognition performance and improved robustness to signal-parameter variations.