28 August 2024 in IEEE Antennas and Wireless Propagation Letters
In this letter, an attention-based residual network (ResNet) is proposed to predict the radiation pattern of phased array antenna (PAA). The proposed model consists of two input and preprocessing modules, a ResNet-attention module and a multi-layer perception (MLP) module. It enables simultaneous prediction of the radiation patterns for PAAs with multiple arrangements and frequencies, which supports the design of PAA and ensures the optimal performance of the equipment in the same electromagnetic environment (EME). Various antenna arrays are utilized to verify the effectiveness and superiority of the proposed model. Discussions also encompass the model's performance under a large array and its generalization capabilities.
29 May 2024 in IEEE Transactions on Electromagnetic Compatibility
Radio frequency power amplifier (PA) is an important part of the transmitter system, which can drive numerous output devices. However, the nonlinear characteristics of PA will cause serious harmonic interference, which leads to electromagnetic interference (EMI) problems. In this article, the nonlinear characteristics and the memory effect of PA are analyzed. The strong nonlinearity region and the weak nonlinearity region are divided according to the strength of the nonlinearity. For the strong nonlinearity, an encoder–decoder-based (E-D-based) artificial neural network model is proposed to predict the harmonic interference of PA. To promote the prediction of high-order harmonics when the input signal is small, a multilayer perceptron model is used for the weak nonlinearity region. The models can effectively predict the first five harmonics of PA, in which the mean absolute error of the fundamental wave is about 0.1 dB, the one of the second-order and the third-order harmonics is about 0.5 dB. Since transfer learning (TL) can simplify the training of the model based on the similarity of different tasks, TL based on model transfer is used to predict the harmonic interference of other PAs according to the existing models. The amount of data required for the modeling of PA can be greatly reduced and the accuracy of prediction can be guaranteed by applying TL. Ultimately, the proposed method can predict the harmonic interference rapidly and accurately according to the known excitation signal so that corresponding measures can be taken to avoid the influence of radiated spurious emission on the use of the sensitive receiving devices in the same electromagnetic environment.
10 October 2024 in IEEE Transactions on Electromagnetic Compatibility
Wireless communication systems have become the cornerstone of modern society. However, the escalating issue of electromagnetic interference poses a significant challenge to the quality and stability of communications. This article aims to assess and analyze the electromagnetic susceptibility of wireless communication systems for enhancing communication performance. This article proposes an advanced predictive model, multifaceted interference susceptibility threshold prediction (MIST-P) via enhanced multitask learning, which integrates multitask learning (MTL) and data augmentation to accurately predict the susceptibility threshold of communication systems in complex electromagnetic environments. By assessing the preformance of four different single task networks, this article adopts the best-performing deep neural network with Xavier initialization as the basis for MTL framework. The MIST-P model is capable of concurrently learning the system's susceptibility under various interference signals and capturing cross-task common features through a shared underlying network structure. This approach improves the generalization ability and prediction accuracy, keeping the prediction error within 0.24 dB. The measured data shows that the MIST-P model not only conserves computational resources, but also holds theoretical and practical implications for the management and optimization of interference in communication systems, providing an effective strategy for the design of future communication systems.
26 September 2025 in IEEE Transactions on Power Electronics
Gallium nitride (GaN) Schottky barrier diodes (SBDs), as essential devices in power-electronic systems, are particularly vulnerable to high-power microwave (HPM) pulses. HPM pulses can induce electro–thermal multiphysics coupling effects in the GaN SBDs, potentially leading to device damage or failure. Dislocation density, as a critical parameter of GaN materials, has a significant effect on the performance of GaN SBDs. In this article, we present an electro–thermal multiphysics coupling analysis of GaN SBDs under HPM pulses, with particular emphasis on the influence of dislocation density on this coupling. Furthermore, by integrating deep learning techniques with technology computer-aided design (TCAD) simulations, we propose a deep multihead attention residual network (DMARN) model that enables the rapid and accurate prediction of maximum electric field strength and temperature within GaN SBDs under varying dislocation density configurations and HPM pulse parameters. Through the validation of ablation and comparative experiments, an additional test set, and hardware experiments, the DMARN model demonstrates outstanding performance and generalization ability.
22 April 2025 in IEEE Sensors Journal
The localization and power estimation of radio frequency interference sources is increasingly becoming an important topic in radio frequency sensing. This article proposes a novel two-step approach based on machine learning, which employs a non-convex sparse recovery algorithm and a fusion neural network specifically designed for the localization and power estimation of multiple RF sources. In the first step, the improved non-convex sparse recovery algorithm for the direction of arrival (DOA) and received signal strength (RSS) joint estimation is proposed based on an iterative maximally sparse convex algorithm, which offers enhanced DOA and RSS estimation accuracy. In the second step, a DOA and RSS fusion artificial neural network is constructed to accurately determine the location and power of the sources by inputting the DOA and RSS estimated in the first step. Simulation results show that both the first step and the second step perform better when used alone or in combination. The performance of the proposed two-step approach is verified through a sensing experiment involving two sources and four sensor arrays with an unmanned vehicle-based testing system.