Automatic Modulation Recognition of Short-time Underwater Acoustic Communication Signal Based on GRU and ResNet

‍ ‍Underwater acoustic communication signal modulation recognition has long been notorious for its low accuracy and long input signal time. To tackle this challenge， a short-term underwater acoustic communication signal modulation recognition method based on gate recurrent unit （GRU） and residual network （ResNet） was proposed of in. First， we designed a new output structure of the GRU layer， extracted the feature image from the short-term communication data， and used it in the subsequent network for further feature extraction. Secondly， the original neural residual unit was improved. This step not only removed the restriction that the number of convolution kernel channels in the unit must be the same as the number of input channels but also retained the function of the residual unit to avoid gradient disappearance. Subsequently， the improved residual unit was used to form the ResNet network layer. This layer used the feature image of the GRU layer as input for high-dimensional and deep-level mining. After that， we used the redesigned GRU and ResNet to form a fusion neural network. Finally， in the simulation， the data of single-path channel， deep sea channel， and shallow sea channel were generated， respectively， as collected data sets to verify the modulation recognition performance of the network， and the real underwater acoustic communication data were collected in the actual sea test as well to ascertain the performance of the network. The proposed network in this paper were compared with the traditional neural networks such as visual geometry group network （VGG）， long short-term memory （LSTM）， gated neural unit， convolutional long short-term memory network （CNN-LSTM）. The simulation results demonstrate that the recognition rate of the proposed network structure can reach 100% when the signal-to-noise ratio is 8 dB or higher under the time-invariant underwater acoustic channel， and the modulation recognition under the time-varying underwater acoustic channel outperforms the above-mentioned traditional network. The sea test shows that the modulation recognition accuracy rate of the network in the natural underwater acoustic channel has a more significant advantage than those traditional neural network mentioned earlier. This is supported by the fact that the proposed network has a correct rate of 65.17%， 11.06% higher than the second-highest GRU， 17.04% higher than LSTM， 21.52% higher than CNN- LSTM and 41.7% higher than VGG.