Automatic modulation classification is a classification problem, and with the help of deep learning achieves outstanding results in various classification tasks. This paper proposes a method based on deep learning combined with convolutional neural networks (CNNs) trained on its own dataset to achieve higher accuracy. In this work, we establish the connection between CNN and signal modulation. The CNN is trained on samples consisting of in-phase and quadrature signal components. In-phase and quadrature patterns are needed to distinguish between modulation modes, which are relatively easy to identify. We selected network parameters to achieve higher recognition accuracy. CNN based on constellation diagrams was able to distinguish between modulation types such as BPSK, QPSK, 8PSK, 16PSK, 32PSK, 64PSK, demonstrating the ability to classify signals even with low signal-to-noise. Several experiments have been conducted to test the CNN network. At 6 dB SNR, most signals achieve accuracy levels greater than 99.9%, with the exception of QPSK (94.5%).