An Efficient Deep Neural Network Structure for RF Power Amplifier Linearization

There has been a strong interest in using deep neural networks (DNNs) for modeling the power amplifier (PA) non-linearity and designing the digital pre-distortion (DPD) circuit. Most DNNs only accept real-valued inputs since the baseband signal has in-phase and quadrature (I/Q) components. As a result, their entire structures can be highly complex. In this paper, we are interested in reducing the complexity of such structures by exploiting both the envelope-dependent terms and residual learning. To acquire such an efficient structure, we propose a novel methodology executed over two consecutive steps; at first, we estimate the best input combinations to a shallow NN that allows it to achieve a threshold value of NMSE. Then, we exploit these combinations as inputs to our proposed structure and increase the network depth until we obtain our system's actual requirements. Finally, our optimized structure (ODNN) performance has been evaluated using MATLAB simulation and real measurements. For a 15 MHz test signal, ODNN achieves lower NMSE than conventional DNN by 2.13 dB and 3.08 dB for Doherty PA behavioral modeling and its DPD design, respectively. For a broader 40 MHz test signal, ODNN achieves lower NMSE by 0.94 dB and 1.94 dB. Moreover, in all previous scenarios, ODNN reduces the complexity of DNN by 26.40%.