Aiming at the problem of poor detection accuracy of OFDM signals caused by channel time-varying and non-stationary under dual-selection fading characteristics, a signal detection scheme based on CNN-GRU (Convolutional Neural Network-Gated Recurrent Unit Neural Networks, CGNN) is proposed. First, use the channel model to generate data to fully mine the prior knowledge of the channel; then use the one-dimensional convolutional neural network in offline training to reduce the dimensionality and feature extraction of the original signal, and use the memory characteristics of the gated loop unit to restore the fading signal; finally In order to reduce the interference caused by the severely fading sub-carriers, an attention mechanism is added to the network training, and weights are assigned to each sub-carrier, so as to perform differentiated training. The simulation results show that the error performance of the detection method proposed in this paper is significantly improved. In a flat fading channel, CGNN can obtain an error performance gain of 0.3dB~1dB. In a frequency selective fading channel, CGNN can obtain an error performance gain of 2dB~5dB, and it has strong robustness.