Aiming at the difficulty of radar signal detection in low signal-to-noise ratio (SNR) condition with traditional methods, a stacked auto-encoder (SAE) and support vector data description (SVDD) based detection method is proposed. Firstly, the radar signal with noise is extracted by SAE to obtain the representative features. Secondly, the SVDD is trained with the extracted features to obtain a spherical discriminative boundary for classification offline. Finally, the trained SAE-SVDD used as the one-class classifier to detect the signal by minimizing both the reconstruction error and the hypersphere volume simultaneously in a real-time manner. Simulation results indicate that the proposed algorithm can extract and identify the radar signal under noise condition effectively with a good robustness. It has practical significance for improving the accuracy of radar signal detection under low SNR.