Adaptive Measurement for High-Speed Electromagnetic Tomography via Deep Reinforcement Learning

Electromagnetic tomography (EMT) is an advanced non-invasive imaging technique that is sensitive to conductive and magnetic materials, offering significant potential for metal non-destructive testing. To overcome the challenges associated with measuring a large number of electromagnetic features in high-speed EMT imaging, we propose a novel adaptive measurement method based on deep reinforcement learning. The AMEMT-Net decision model is designed to dynamically optimize the measurement sequence and prioritize the key features that most significantly contribute to the imaging process. Since EMT’s electromagnetic features in EMT are not entirely independent, the proposed method enables high-quality imaging without requiring all features, thus improving detection speed. Using a 24-coil planar EMT system as a case study, we evaluated the method’s performance in metal defect imaging through electromagnetic finite element simulations and FPGA-based practical experiments. The results show that, when using the Tikhonov image reconstruction method, AMEMT-Net requires only 10% of the electromagnetic features to achieve over 85% of the correlation coefficient compared to using all features. This reduces the EMT imaging time to just 21% of that required by traditional methods and makes image reconstruction less sensitive to noise. This study demonstrates the effectiveness of the proposed adaptive measurement method, providing an innovative approach for the development of high-speed EMT.