Approach for Extreme Learning Machine-Based Microwave Power Device Modeling

Abstract

The nonlinear representation of active devices plays an important role in microwave circuit design. Whereas, it takes a long time to extract a large amount of large signal data, and the problem of memory resource and CPU occupancy becomes significant. In order to address the problems in traditional large-signal modeling methods, in this paper an X-parameter modeling method for microwave power devices based on extreme learning machine (ELM) is proposed. To demonstrate the effectiveness of this method, a double layer back propagation (BP) neural network model is established. Then, harmonic balance simulations are used to verify the accuracy of these two models. After comparisons, it is proved that the three harmonic errors of double layer BP neural network model are 9.525dBm, 1.309dBm and 14.593dBm, respectively, and the three harmonic errors of ELM model are 0.673 dBm, 0.314 dBm, 3.09 dBm, respectively. Furthermore, the three harmonic modulus errors of double layer BP neural network model are 0.031, 0.002, 7.665e-4, respectively, and the errors of ELM model are 0.005, 0.001, 8.38e-5, respectively. Finally, in order to verify the accuracy of the predicted model in circuit design, the predicted X-parameter is used in the design of power amplifier. Moreover, the errors of the double layer BP neural network prediction model at 2.5 GHz, 5 GHz and 7.5 GHz are 1.142 dBm, 1.436 dBm and 2.294 dBm, respectively. The output power error of the ELM model at 2.5 GHz, 5 GHz and 7.5 GHz are 0.089 dBm, 0.311 dBm and 0.309 dBm, respectively. These experimental results demonstrate that the established ELM model is an efficient and valid approach for modeling GaN high electron mobility transistor types of nonlinear microwave devices.