An End-to-End Approach for Enhanced Weak Signal Detection in Time-frequency Analysis

Time-frequency analysis (TFA) is valuable in crucial application domains such as biomedical sciences, communication systems, and environmental monitoring. Existing TFA approaches, including those based on neural network algorithms and Synchrosqueezing Transform (SST), exhibit limitations when dealing with intersecting strong and weak Instantaneous Frequency (IF). To address this issue, we introduce an end-to-end neural network specifically designed for weak signal detection, WeakSig-TFANet (WTFA-Net), guided by traditional weak signal processing schemes. The network employs a combination of time-domain and convolutional windows to attenuate the sidelobes of strong signals in the convolutional input, demonstrating exceptional recognition capabilities in scenarios involving weak signals such as faint respiratory and cardiac signals. To enhance resolution in cases of IF intersections, the proposed model structure further integrates complex attention modules and multi-coarse estimation basis pixel-level attention mechanisms. These mechanisms counteract the decline in the ability to detect adjacent signals, a decline caused by the widening of the main lobe due to time-domain windowing of strong signals. Through experiments across multiple application scenarios, including both simulated and real-world signals, WTFA-Net is shown to outperform existing methods in weak signal detection performance in vital sign monitoring, underwater acoustic detection, channel state information (CSI) of Wi-Fi and Inverse Synthetic Aperture Radar (ISAR) image generation.