To address the issue of signal detection in fading channel low signal-to-noise ratio (Low-SNR) environments, we suggest combining deep learning methods with signal processing technologies and propose a Low-SNR signal detection method based on Fusion Feature Convolutional Neural Network (FFCNN). The designed FFCNN converts low-order local features of signals into high-order fusion features with stronger discriminative power through operations such as multi-layer convolution, nonlinear activation, and category mapping, enabling binary classification for signal detection. Simulation tests in Rayleigh fading channels under both Relatively Low-SNR (−10 dB, 0 dB) and Extremely Low-SNR (−20 dB, −10 dB) environments demonstrate this method’s robust generalization ability, high signal detection performance, and the FFCNN structure’s excellence in compensating for signal feature fluctuations and risisting fading interference.