LWM-Spectro: A Foundation Model for Wireless Baseband Signal Spectrograms

TL;DR

LWM-Spectro addresses the challenge of learning transferable representations from raw wireless I/Q signals by pretraining a transformer-based foundation model on 9.2 million spectrograms across WiFi, LTE, and 5G. The model combines masked spectrogram modeling, contrastive learning, and a mixture-of-experts with protocol-specific encoders and a lightweight router, enabling efficient, protocol-aware feature extraction. It demonstrates strong few-shot and cross-protocol performance on modulation classification and joint SNR/Doppler tasks, achieving high accuracy with limited labeled data and robust generalization to unseen environments. This work provides a scalable, unified foundation for wireless learning, aiding practical deployment in diverse propagation conditions and standards.

Abstract

The received in-phase and quadrature (I/Q) baseband signals inherently encode physical-layer and channel characteristics of wireless links. Learning robust and transferable representations directly from such raw signals, however, remains challenging due to heterogeneous communication systems, diverse propagation environments, and limited labeled data. To address this, we present LWM-Spectro, a transformer-based foundation model pretrained on large-scale I/Q data represented as time-frequency spectrograms. The model leverages self-supervised masked modeling, contrastive learning, and a mixture-of-experts (MoE) architecture to learn general-purpose wireless representations. These representations transfer effectively to downstream tasks such as modulation classification and joint SNR/mobility recognition, even with minimal supervision. Across tasks, LWM-Spectro consistently outperforms state-of-the-art deep learning baselines in both few-shot and data-rich regimes, providing a unified foundation for wireless learning.

Figures (5)

• Figure 1: Block diagram of the baseband-equivalent spectrogram generation pipeline. Information bits are encoded, interleaved, mapped, and pulse-shaped to form the transmitted signal $x[n]$, which propagates through a ray-tracing–based fading channel with AWGN. At the receiver, carrier- and sampling-frequency offsets (CFO/SFO) are compensated, and the resulting baseband signal $y[n]$ is converted to a spectrogram $\mathbf{S}$ via STFT. Only baseband processing is considered, assuming ideal DAC/ADC.
• Figure 2: WiFi BPSK spectrograms (128 × 128 pixels, coding rate = 1/2, SNR = 20 dB) under static and vehicular mobility.
• Figure 3: t-SNE visualization of 5G samples color-coded by SNR. Raw spectrograms show heavy overlap, while LWM-Spectro embeddings form well-separated, discriminative clusters by SNR order.
• Figure 4: t-SNE plot of the same samples color-coded by modulation type.
• Figure 5: Validation macro-F1 score for modulation classification across training set sizes. The LWM embedding sustains higher accuracy than baseline models, particularly in few-shot regimes.