Determining the Highly Susceptible Waveform for Electromagnetic Susceptibility Testing via Reinforcement Learning
Jingxuan Chen, Guangzhi Chen, Junhao Zhang, Donglin Su · IEEE Transactions on Electromagnetic Compatibility · 2024
Electromagnetic susceptibility (EMS) testing is critical for ensuring the electromagnetic compatibility of electronic equipment. While current standards employ fixed waveforms for tests, equipment under test (EUT) often shows high susceptibility to certain waveforms. This article proposes an EMS test method based on reinforcement learning, termed EMS-RL, which can determine the highly susceptible waveform for the EUT. The related concepts of reinforcement learning are introduced to the EMS test in the architecture of the EMS-RL. It considers the RF signal source as an agent, which modifies the modulation state of waveforms through actions to obtain rewards. The rewards are defined as the EUT's negative susceptibility threshold. EMS-RL applies a twin delayed deep deterministic policy gradient network and a reward function designed with reward-shaping methods to determine the highly susceptible waveform more stably and efficiently. An automated closed-loop EMS test system is constructed to verify the proposed method on three different EUTs with bulk current injection test and direct power injection test. Compared to standard and genetic algorithm-determined waveforms, the EMS-RL-determined waveform excites similar susceptibility with lower carrier power and faster testing time.