A Compressive Sampling And Deep Learning-Based Algorithm For Wideband Signal Detection

Abstract

In the field of frequency-hopping signal detection, conventional methods face the dual challenges of high computational complexity and artificially designing detection statistics, which significantly limits their application in resource-constrained scenarios. To solve this problem, we proposed a frequency-hopping signal detection algorithm, stft-cs-shufflecbam, by combining time-graph compression perception and deep learning. Short-time Fourier transform (STFT) transforms frequency-hopping signals into time chart representations, and two-dimensional compression sensing technology and half-tensor product (STP) reduce the computational load by dimensioning the time chart down. A lightweight network model, ShuffleNet, is then constructed and a convolutional block attention module (CBAM) is introduced to enhance the feature extraction capability of the frequency hopping signal. Finally, the classification and identification of the compressed time chart enabled efficient detection of the frequency-hopping signal. The goal of this research is to provide a solution that achieves both accuracy and computational efficiency for wideband signal detection in resource-constrained environments such as mobile devices and embedded systems.

In conventional research, the detection accuracy of frequency-hopping signals has been improved by deep learning and the data dimension has been reduced by compression perception. However, in the existing methods, the attenuation of the frequency hopping signal in the time frequency domain cannot be fully utilized, thus increasing the computational complexity and relying on the manual design of the detection statistics. . Furthermore, some studies have attempted to combine compression perception with deep learning, but there are limitations in maintaining detection performance in low SNR environments. In the existing method, it is easy to generate the problem that the detection probability decreases due to the loss of feature information under the condition of low s/n ratio. In addition, in existing studies, the design optimization of the compression perception measurement matrix was insufficient, and it was difficult to achieve both information retention and computational efficiency in feature extraction. Therefore, innovative methods that efficiently adapt to noise environments and improve detection accuracy while reducing computational complexity are required.

The specific process of this research is to generate a time chart of the frequency-hopping signal by using TFR Includes. In experiments, when using a 512 length humming window, the stft-shufflecam method achieves a signal-to-noise ratio (SNR) of -10 At dB, it is shown that the detection probability is improved by approximately 69.5% compared with the stft-shufflecam method and 42% compared with the wt-shufflecam method. Second, by integrating the shfflenet and CBAM modules, the proposed stft-cs ShuffleCBAM model can greatly improve the detection performance while maintaining lower computational complexity. Specifically, in a low SNR environment with an SNR of -14 dB, the detection probability of stft-s-shufflecbam was 1, 0.16 for the stft-resnet50 method, and 0.9 for the cnn-st method. For actual applications, a signal acquisition system for a rp-n310 receiver and a portable radio station was constructed to verify the robustness of the algorithm. When the sample coefficient is compressed to M=64, the network model is fine-tuned to increase the detection probability from 0.58 to 0.92. In addition, the calculation complexity of stft-cs-shufflecbam was found to be significantly lower than that of hcrnn-6 or cnn-st. For example, if M=4, the time complexity is 1.4110 x 10⁷ FLOPs, and the space complexity is 2.4777 x 10⁵. It is 92.7% and 95.3% lower than hcrn-6 respectively. In this research, we propose a practical solution that effectively overcomes the limitations of conventional methods by optimizing compression perception and deep learning framework, and combines the accuracy of frequency hopping signal detection with low computational complexity.