The separation and recognition of radar signals are crucial in a complex electromagnetic environment, especially multi-component radar signals. However, most existing algorithms can only recognize dual-component signals. An algorithm based on semantic segmentation is proposed to separate the signal in the time-frequency domain and classify multi-component radar signals. An improved Cohen class time-frequency distribution (CTFD) is used to represent the one-dimensional signals as time-frequency images (TFIs). A convolutional denoising autoencoder (CDAE) is established to filter the TFIs. Three semantic segmentation networks are used, a fully convolutional neural network (FCN-8s), U-Net, and DeepLab V3+. The method can separate and recognize signals simultaneously and is applied to aliased signals composed of 1-4 components. The simulation results show that the proposed method provides excellent performance for separating and recognizing multi-component signals. At a signal-to-noise ratio (SNR) of 0 dB, the accuracies of the aliased radar signals with 1-4 components are 100%, 100%, 96.67%, and 93.75%, respectively. The separation and recognition algorithm can be adapted to other signal modulation types.

雷达信号的分离和识别在复杂的电磁环境中至关重要，尤其是多分量雷达信号。然而，大多数现有算法只能识别双分量信号。提出了一种基于语义分割的算法，在时频域对信号进行分离，并对多分量雷达信号进行分类。改进的 Cohen 类时频分布 (CTFD) 用于将一维信号表示为时频图像 (TFI)。建立卷积去噪自编码器（CDAE）来过滤 TFI。使用了三个语义分割网络，全卷积神经网络 (FCN-8s)、U-Net 和 DeepLab V3+。该方法可以同时分离和识别信号，适用于由1-4个分量组成的混叠信号。仿真结果表明，该方法对多分量信号的分离和识别具有优异的性能。在信噪比 (SNR) 为 0 dB 时，具有 1-4 个分量的混叠雷达信号的精度分别为 100%、100%、96.67% 和 93.75%。分离和识别算法可以适应其他信号调制类型。