Multi-In-Multi-Out Neural Network for Joint DOA Estimation and Automatic Modulation Classification

Direction of arrival (DOA) estimation and automatic modulation classification (AMC) of radio frequency (RF) signals are two crucial tasks in electronic intelligence systems. These two tasks are traditionally performed in separate individual processes that result in slow latency and computational complexity. In order to mitigate the mentioned issue, a multi-in-multi-out deep neural network (namely MIMONet), which has three inputs and two outputs, is proposed in this paper for joint DOA estimation and AMC applied for uniform circular array. The three inputs are designated for raw in-phase and quadrature-phase signals, Fourier transform data, and covariance matrix. The two outputs are assigned in turn for DOA estimation and AMC. The MIMONet model is analyzed with different hyperparameter options to find the best performance trade-off between DOA estimation and AMC accuracy, computational complexity, and execution time. As a result, the MIMONet model of 32 filters with a size of 3×3 has achieved the best performance with AMC accuracy higher than 95%, root mean square error of DOA estimation below 0.1°, and execution time of 0.74±0.02 ms for SNRs greater than 10 dB. In comparison, the proposed model has outperformed some other state-of-the-art models in the same experimental scenario.