P-Band Burst Signal Onset Detection Based on Time-Frequency Diagrams and RMS-Dense-YOLO Networks

In this paper, an algorithm for P-band burst signal onset detection based on signal time-frequency diagram and RMS-Dense-YOLO network is proposed. The algorithm makes full use of the feature of concentrated energy distribution of P-band burst signal preamble at fixed frequency points in the time-frequency domain, avoids the problem that traditional YOLO networks cannot match the diverse low-level features with the abstract high-level semantic features due to the different signal modulation methods, code speeds, and frequency offsets, and enhances the accuracy of the burst signal onset detection. The algorithm firstly utilizes the time-frequency diagram of the burst signal to complete the image construction, and there is an obvious difference between the energy distribution of the signal's preamble frequency points and that of the noise in the image; secondly, the RMS-Dense-YOLO network proposed in this paper contains the backbone network RMS-Backbone and the feature fusion network Dense-Neck; in the backbone network, the SPPF and the multilayer convolution structure are utilized to In the backbone network, SPPF and multilayer convolution structure are utilized to expand the sensory field and combine with the proposed RMS module to obtain high-level semantic features such as spatial relationship between the signal preamble and the noise in the signal time-frequency image; in the feature fusion network, a DenseBlock-based feature pyramid and path aggregation network cascade module is proposed, which multi-scale fuses the high-level semantic features of the signal preamble extracted from the backbone network with its own low-level information, such as texture and contour, to improve the accuracy of detecting sudden signal onset points. The experimental results show that the average detection accuracy of the proposed algorithm in this paper is improved by 22.41%, 7.80%, and 4.10% at -5dB signal-to-noise ratio (SNR) compared to the detection algorithms of signal frequency-domain ratio algorithm, signal zero-intermediate-frequency I/Q data and CNN network, and signal time-frequency image and YOLOv5s network, respectively.