Signal classification based on multi-Scale time-Frequency transformer

Abstract

Recognizing communication signals under non-cooperative remains challenging due to long waveform durations, noise corruption, and complex time-frequency structures. Existing convolutional neural network-, recurrent neural network-, and Transformer-based methods often rely on a single modeling paradigm and struggle to jointly capture local waveform patterns, long-range temporal dependencies, and complementary spectral information. To address these limitations, we propose MTF-Former, a Multi-scale time-frequency Transformer tailored for one-dimensional radio-frequency (RF) signals. MTF-Former combines signal-oriented data augmentation, a lightweight Time-Frequency Enhancement (TFE) block that injects frequency-aware modulation, and a hierarchical windowed Transformer encoder for efficient multi-scale temporal modeling. This unified design effectively integrates temporal and spectral cues while reducing computational cost compared with standard global self-attention Transformers for long sequences. Experiments on specific emitter identification (SEI) and automatic modulation recognition (AMR) benchmarks demonstrate that MTF-Former consistently outperforms many methods, achieving more notable performance gains at lower signal to noise ratio, with accuracy improvements of 1.30% on SEI (5 dB) and 2.42% on AMR (0 dB), and ablation studies further validate the contribution of each component.