To overcome the defects of traditional multisensor systems that rely on pulse parameters for homologous data correlation, we propose a multichannel separation network that directly extracts the homologous waveforms from multisensor intercepted signals. The network consists of two main structures in parallel, which are constructed to train waveforms and spectra, respectively. In time-domain separation, we adopt one dimensional (1-D) convolution and self-attentive mechanism to extract high-dimensional important features and then construct stacked bidirectional long short-term memory (Bi-LSTM) for separation training. In the frequency-domain separation, we utilize the complex Unet structure to encode and decode the fast Fourier transformation (FFT) spectra and then reconstruct the signals with the improved complex ratio mask (IcRM). For the fusion part, the corresponding channels of the two structures are fully connected to obtain the final separation results. Experiments show that the proposed network not only has fewer parameters but also outperforms the classic networks by 1–10-dB separation signal-to-distortion ratio (SDR). It is capable of homologs separation in terms of dual sensor, multisensor, and the number of homologous signal unknowns, creating a new way for multisensor data correlation while providing strong support for the subsequent data fusion.