Short-wave specific signal recognition based on feature sequence learning and pre-training

This study aims to develop a deep learning framework for recognizing specific short-wave communication signals by identifying their feature sequences (e.g., preambles and synchronization codes) in narrowband data. Conventional methods rely on manual feature extraction or template-based correlation, which suffer from high computational cost, linear scaling with the number of signal types, and poor robustness to symbol-rate variations and channel impairments. We seek to address these limitations by designing neural architectures that are receptive-field-aware and context-aware, while also tackling the practical problem of limited labeled real-world data through a simulation-based pre-training strategy. The ultimate goal is to achieve high recognition accuracy, robustness to fading and frequency offsets, and significantly faster processing than traditional approaches, in an open-set scenario where non-target signals and background noise are prevalent.

We process fixed-length sliding windows (6 s, 9600 Hz) of narrowband data and classify each window into one of several feature-sequence classes or a background class. Two lightweight 1D CNN backbones—VGGNet-16 and ResNet-18—are adapted from their image versions. Two key architectural improvements are introduced. First, the receptive field is enlarged by adjusting down-sampling strides of intermediate layers to ensure that the field covers at least two symbol durations for low-rate signals (e.g., 50 Baud FSK) while avoiding over-coarse granularity for high-rate signals (e.g., 2400 Baud 3G-ALE). Second, a multi-head self-attention (MHSA) layer is appended after the convolutional stages to capture long-range dependencies across local symbol features, because feature sequences often contain dozens to hundreds of symbols. The MHSA also enhances attention to the sequence portions, suppressing interference from non-target segments. A sigmoid layer with a fixed threshold (0.5) enables open-set background rejection, and the network is trained with binary cross-entropy loss.

To overcome data scarcity, we propose a pre-training scheme using extensively simulated signals. The simulation covers PSK/QAM (orders 2–16, symbol rates 1200–2400 Baud) and FSK (orders 2–8, rates 50–200 Baud) with varied parameters. For each modulation setting, multiple random feature sequences are generated, and for each sequence we create three sample types: (a) feature and non-feature parts share the same modulation, (b) they differ, and (c) the signal contains no feature sequence (background). The total pre-training dataset has 913 classes and over 270000 samples. Online augmentation includes random zero-padding, center-frequency offset, additive noise (SNR 5–20 dB), and three-tap Rayleigh fading. The pre-trained model is then fine-tuned on real-world data with mild frequency-offset augmentation. Real-world data consist of seven signal types (3G-ALE, 110A, 110C, 2FSK_1, 2FSK_2, ANDVT, PSK_1) collected from multiple frequencies over one month, plus a large background class. Class imbalance is mitigated by weighted random sampling per epoch.

Experiments on real-world data with three random train-test splits show that the proposed modifications consistently improve performance. Enlarging the receptive field by about 4× increases the F1 score for low-rate 2FSK signals by 6–7%, while over-enlargement (8×) degrades high-rate signals by 2–4%, confirming the importance of symbol-rate-aware design. Adding the MHSA layer further improves the F1 by 2.25–2.96% for the two backbones with negligible extra computation. Without pre-training, the improved ResNet-18 achieves 90.89% average F1, comparable to the traditional correlation-based FFT method (91.98%). With 913-class pre-training, it reaches 94.50%, outperforming correlation and also the STFT+Swin-Transformer approach (93.43%). Pre-training with a smaller 97-class set yields 93.24%, indicating that richer simulation data enhance transferability. Confusion matrices reveal that most errors occur between targets and background (false alarms/misses), while cross-target confusions are rare, thanks to distinct modulation parameters and feature sequences. Ablation studies confirm that all three types of simulated samples contribute positively to fine-tuning performance. In processing speed, the improved ResNet-18 handles a 6-second sample in 4.30 ms on a Tesla V100, while the correlation method with 16-thread parallelism requires 205.90 ms for ten correlations per sample. Moreover, neural computation time remains nearly constant as the number of target types grows, whereas correlation scales linearly. ResNet-18 also has much lower parameters (6.86M) and FLOPs (6.56G) than VGGNet-16 (24.88M, 19.74G), making it more suitable for deployment.

This work demonstrates that deep learning, with receptive-field adjustment and self-attention, can effectively recognize short-wave feature sequences with high accuracy and speed. The proposed simulation-based pre-training successfully alleviates real-world data limitations and enables the model to surpass traditional correlation methods even with limited labeled samples. The neural approach offers practical advantages: it requires no prior knowledge of sequence contents, its processing time is independent of the number of target types, and it can exploit modulation-parameter features to distinguish signals lacking feature sequences. However, its reliability depends on training data richness, and the open-set environment necessitates continuous data collection. Future work will explore hybrid schemes combining neural networks with template correlation to further enhance robustness while maintaining throughput. Overall, the proposed framework provides a scalable and efficient solution for intelligent short-wave monitoring in complex electromagnetic scenarios.