Methods of Spectrograms' Processing for Deep Neural Networks for the Real-Time Recognition of Radio Signal Modulation

Automatic recognition of radio signal modulation is a key factor in processing radio signals in various technical applications, such as radio communications, radar, direction finding, and others. In this study, methods for automatic modulation recognition applicable to embedded and edge hardware platforms with limited performance and power constraints are considered. The foundation of the proposed approach is the application of a conventional short-time Fourier transform to obtain a two-dimensional time-frequency spectrogram, subsequently, various combinations of spectrogram output parameters and their transformation functions are studied to generate the neural network's input features. Our initial assumption has been that traditional modulation recognition, which relies solely on the magnitude of the spectrogram, results in an inherent and undefined reduction of the essential information present in the original complex samples. The initial output complex parameters are introduced into the processing pipeline; these include the real and imaginary parts, as well as derived parameters such as magnitude and phase. An optimised CNN class neural network, tailored as an optimal architecture for processing two-dimensional spectrogram inputs, is constructed. By adjusting the neural network's parameters, the dataset is processed to obtain output quality metrics. A comparison is presented, detailing the accuracy and efficiency of radio signal modulation processing across all proposed spectrogram output parameter representation methods. Our research shows that introducing a combination of the real, imaginary, amplitude, and phase spectrogram parameters significantly improves neural network recognition, achieving 96 % accuracy, which is 25 % higher than conventional magnitude-only input. The observed result is invariant to the type of neural network utilised, introducing only a minor increase in network complexity and computational burden.