A Deep-Learning-Based Blind Detection Method for Wideband Multisignal Sensing

Spectrum sensing plays a crucial role in fields such as sixth-generation (6G) wireless systems, Internet of Things (IoT), and cognitive radio. Blind signal detection is a crucial component of spectrum sensing. Current signal detection methods mainly focus on detection signal in a fixed single channel. Some multi-signal detection methods utilize spectrum for detecting multiple signals in broadband reception. However, the estimation of signal occupancy time exhibits poor performance. Some studies have noted that the Short-Time Fourier Transform (STFT) effectively characterizes the time-frequency occupancy properties of signals in electromagnetic space and have employed deep learning techniques to further enhance multi-signal sensing abilities. However, these methods directly apply image processing neural network models without considering the unique morphology of signals in spectrograms. In this article, we tackle the distinctive morphology of signals in spectrograms by introducing a Vertical-Horizontal Convolutional (VHConv) structure. This innovation aims to boost the feature extraction capacity of neural networks. We also introduce an attention-based decoder prediction head to enhance the receptive field of the prediction network and improve multi-signal detection performance. Based on comparative experiments, we find: 1) the implementation of the VHConv module notably improved signal detection precision and recall, achieving a notable Average Precision (AP) improvement from 29.9% to 67.7% at an 8dB Signal-to-Noise Ratio (SNR) when integrated into a convolutional neural network. 2) Utilizing a decoder prediction head networks enhanced the neural network’s capability in estimating time-frequency occupancy parameters. 3) Integrating decoder prediction head networks with Intersection over Union (IoU) loss during the training process significantly improved precision, achieving a 6.2% detection precision enhancement (from 80.3% to 86.5%) at recall rates above 80%. The experimental results demonstrate that the proposed method surpasses current methods, effectively enhancing the multi-signal detection capability of spectrum sensing systems.