Spatiotemporal Spectrum Neural Network for Low-SNR Communication Signal Detection

Detecting wireless communication signals in low signal-to-noise ratio (SNR) environments is challenging due to noise interference and frequency overlap. Traditional array processing techniques rely on direction-of-arrival (DOA) estimation and beamforming before detection and recognition, often resulting in high computational complexity and potential error accumulation. To address these challenges, this paper proposes a spatiotemporal spectrum neural network (STSNN), an end-to-end multitask learning framework that directly processes array signals, performing joint signal detection, modulation recognition, frequency band estimation, and DOA estimation in a unified model. STSNN first applies a two-dimensional Fourier transform (2D-FFT) to compute the spatiotemporal spectrum, then identifies a distinctive spatiotemporal center keypoint to enhance multitask learning efficiency. Extensive simulations and real-world experiments confirm its effectiveness, with a 16-element array achieving an F1-score of 0.989 at -15 dB SNR, significantly outperforming single-channel spectrogram-based methods (0.470). As a data-driven approach, STSNN remains robust to array element position errors, ensuring reliability in diverse deployment scenarios. By eliminating the need for explicit DOA estimation and beamforming, STSNN simplifies the processing pipeline, enhances detection robustness, and offers a scalable solution for low-SNR signal detection and estimation in radio frequency, sonar, and other array-based systems.