Abstract: Modulation recognition represents one of the key fields in the wireless communications field, as it is widely used in both criminal and peaceful communication forms, following the application area. Nonetheless, things get hard when there is a drop in the signal quality, especially when we deal with the low SNR (across -10dB), providing us with such poor results. This is why we offer an enlistment of the Modulation Recognition System using a Denoising Bidirectional Recurrent Neural Network (DBRNN). The initial layer examines signal restoration, which is the role of the DBRNN (Denoising Bidirectional RNN), and the goal is to reduce the effect of noise. Next, an autoencoder is used as the second phase of the system, where the signal is decomposed into the information without noise. The procedure in question involves transmission and then finally, a noiseless deletion. Finally, we use three stages of the Convolutional Neural Network (CNN) modules for real classification of the noise-reduced output of the autoencoder unit. By means of the application of our method, we can look for the modulations, spectral features or speech patterns within the noisy audio signals, which are very useful for cognitive radio systems and many other applications. The detection algorithm simulations that we conducted evidenced that our solution was showing a growing tendency to increase the efficiency of the recognition accuracy. Subsequent inspection of the diverse design with various structural groups and frameworks of the parameters helps in achieving an accuracy of 85% for classification, even for -20 dB signals , thus proving its sturdiness and efficiency.

Keywords: autoencoder, CNN, Denoising Bidirectional Recurrent Neural Network, low SNR, modulation recognition, noisy audio

Pages: 309-316