ABSTRACT Separating single‐channel multi‐component linear frequency modulation (LFM) signals presents significant challenges, including the absence of prior information, difficult weak signal detection under low signal‐to‐noise ratios (SNR) and poor resolution in overlapping regions. To overcome these limitations, a blind separation method based on a multi‐feature enhanced Swin‐Transformer is proposed. The separation task is reformulated as a semantic segmentation problem on the spectrogram, employing an end‐to‐end processing pipeline of ‘sample set construction—model training—signal reconstruction’. First, the time‐domain signal is mapped to a spectrogram via the short‐time Fourier transform (STFT) to preserve frequency modulation trajectories. Concurrently, data augmentation is employed to expand the sample size. Second, an unsupervised blind annotation strategy based on energy peak detection and region growing is used to generate pixel‐level labels. Subsequently, with the Swin‐Transformer as the foundational framework, a tailored block embedding layer is optimised by incorporating the convolutional block attention module (CBAM) and multi‐scale feature fusion modules. This enhanced backbone is paired with a lightweight decoder and a hybrid loss function to enhance segmentation accuracy. Finally, pixel‐level filtering is performed on the original time‐frequency map using component‐specific masks to extract corresponding component time‐frequency maps. These are then reconstructed into independent LFM time‐domain signals via inverse time‐frequency transformation. Experimental results on both simulated and field data demonstrate that the proposed approach outperforms baseline models and mainstream methods—including LSTM‐TasNet, CNN‐BiLSTM and MLFMSS‐Ne—in terms of correlation coefficients, intersection over union (IoU) and SI‐SDR metrics under both standard and low signal‐to‐noise ratio (SNR) conditions. An effective solution for signal separation in scenarios such as radar detection and wireless communications is thereby provided.
Sun et al. (Thu,) studied this question.