Abstract: ‍ ‍As one of the core technologies of non-cooperative communication， blind modulation recognition technology for wireless communication signals plays a crucial role in improving spectrum utilization efficiency and the demodulation of unknown signals. In addition， non-cooperative communication experiences problems such as an unknown electromagnetic environment， serious noise interference， and a low signal-to-noise ratio， which make it difficult to blindly modulate and identify unknown signals. In order to solve the problem of subcarrier blind modulation recognition of multiple-input multiple-output orthogonal frequency division multiplexing signals in non-cooperative communication at a low signal-to-noise ratio， this study used a convolutional neural network （CNN） and long short-term memory （LSTM） network to build a one-dimensional CNN-LSTM network for blind modulation identification. Because of the strong feature-expression ability of I/Q data， the algorithm used I/Q data as the first input feature and directly entered it into the network. In order to compensate for the interference of noise on I/Q data， a cyclic spectrum with strong noise immunity was also selected as another input feature. In order to further improve the noise immunity of the cyclic spectrum， a cyclic spectrum slice accumulation sequence with better noise immunity was used as the second input feature. Simulation results showed that the proposed method could recognize the ｛BPSK， QPSK， 8PSK， 16QAM， 32QAM， 128QAM｝ modulation mode under a signal-to-noise ratio of 2 dB， and the recognition accuracy reached 98%.