Spectral Encoders for Signal Classification

Learning from Spectra

Traditional signal classification pipelines extract hand-crafted features from spectral data before feeding them to a classifier. Spectral encoders bypass this step by learning features directly from frequency-domain representations.

Architecture Design

Our spectral encoder uses a 1D convolutional backbone operating on the power spectral density, followed by attention layers that learn to focus on the most informative frequency bands. The encoder maps input spectra to a compact latent space where structural similarity corresponds to geometric proximity.

class SpectralEncoder(nn.Module):
    def __init__(self, n_freq=512, latent_dim=64):
        super().__init__()
        self.conv = nn.Sequential(
            nn.Conv1d(1, 32, 7, padding=3),
            nn.GELU(),
            nn.Conv1d(32, 64, 5, padding=2),
            nn.GELU(),
            nn.AdaptiveAvgPool1d(64),
        )
        self.head = nn.Linear(64 * 64, latent_dim)

    def forward(self, x):
        h = self.conv(x.unsqueeze(1))
        return self.head(h.flatten(1))

Training Strategy

We train with a contrastive loss that pulls spectra from the same structural class together and pushes spectra from different classes apart. This produces a latent space where k-NN classification achieves high accuracy.