<p>We consider single-channel blind separation of two PSK signals with unknown parameters. This problem is particularly important for communication systems where two signals occupy the same frequency band. It is shown that the blind separation problem can be reduced to a multi-class classification problem, which can be solved using modern machine learning methods. A four-layer neural network is proposed to solve this problem. The effectiveness of this method has been verified using BPSK and QPSK signals under various amplitude ratios and noise levels. It is also shown that in the absence of one of the signals, the proposed separation method effectively demodulates the present signal.</p>