Automatic Classification of Modulation Types in Digital Communication Signals with Probabilistic Neural Networks

In recent days, the Automatic Classification of Modulations (ACM) is one of the major considerable research attention techniques in Digital Communication (DC) signals, which is used to design the intelligent transceivers. Previous researchers have suggested various traditional feature-based methods including Higher-Order Cumulants (HOC) have been utilized for ACM due to low computational complexity but these are sensitive to noise and interference. To overcome these issues, Probabilistic Neural Network (PNN) is proposed due to faster adaptability. Initially, the data is considered from Real-world Wireless Communication (RWC) Dataset and further preprocessed with Spectral Subtraction (SS). The SS preprocessing technique effectively reduces the stationary noise without significantly destructing the signal. After that, Mel-Frequency Cepstral Coefficients (MFCCs) extracts features from the preprocessed signals by mel-frequency wrapping to represent the signals spectral characteristics. Finally, the classification of modulations is performed with proposed PNN by computing distance between input signals and modulation patterns to select modulation type with highest probability. From the results, the proposed PNN established better results when compared to existing threshold autoencoder denoiser convolutional neural network (TADCNN) by providing accuracy as 88.90% respectively.