Automatic Modulation Recognition Method Based on Hybrid Model of Convolutional Neural Networks and Gated Recurrent Units

Modulation signal recognition based on feed‐forward attention mechanism

“…The architecture of our proposed RLADNN is shown in Figure 1. Our model starts from a CNN with a convolutional kernel size of (2,8), which consists of two convolutional layers, a residual structure layer, two LSTM layers, and an Attention layer, and the whole is mainly presented as a tandem arrangement. The superiority of CNN in extracting spatial features, and generating high-level information makes CNN widely used in the field of image recognition.…”

“…The ratio is 6:2:2. In this letter, the publicly available dataset RadioML2016.10a from [3] is used to evaluate the DL models, we chose the following typical DL-based AMR algorithms as benchmarks, which contain various types of mainstream frameworks for DL and each has its own advantages (CNN-LSTM [9], Complex-CNN [15], improved convolutional neural network (CNN)-based automatic modulation classification network (IC-AMCNet) [16], Convolutional Long-Short Term Deep Neural Network (CLDNN) [17], Convolutional Neural Networks and Long-Short-Term Memory Networks (RCTLNet) [18], and CNN-GRU hybrid network model (CGRNet) [19]). As shown in Figure 2, even compared with the current state-of-the-art model, the recognition rate of the model proposed in this paper has improved significantly over the entire SNR range, and when Attention is added, it results in a more substantial performance improvement at [−6:18] SNR based on the original model, but with only a small increase in time and number of parameters.…”

“…The global average pooling (GAP) method used by Hao et al. [8] to reduce the dimensionality of the high‐dimensional feature maps makes the model more simplified, requires a smaller number of parameters, and has a higher training speed. Zhang et al.…”

“…Zhou et al [7] proposed an automatic modulation classification relation network (AMCRN) structure to improve the performance with few samples, which uses a onedimensional convolution to effectively reduce the dimensionality, reduce redundancy, and remove some noise. The global average pooling (GAP) method used by Hao et al [8] to reduce the dimensionality of the highdimensional feature maps makes the model more simplified, requires a smaller number of parameters, and has a higher training speed. Zhang et al [9] used a dual-stream structure of CNN and LSTM in tandem to fuse the features obtained from different branches by feeding the inphase/quadrature components and amplitude/phase of the signal into the network simultaneously, respectively.…”

Radio Modulation Classification Optimization Using Combinatorial Deep Learning Technique

“…S. Chang et al [14] combined CNN with Bi GRU networks achieving an accuracy of 84% at high SNR. Q. Duan et al [18] used combination of CNN with BiLSTM and Attention models achieving an accuracy of 93% at high SNR. J. Xu et al [11] combined CNN and LSTM with FC models achieving an accuracy of 90% at high SNR.…”