Leveraging Ultrasound and Neural Networks for Non-Invasive Power Converter Efficiency Estimation

Measuring the efficiency of power converters traditionally involves invasive sensors or computationally heavy indirect methods. We propose as an emerging topic, a novel, contactless approach using ultrasound measurements to estimate power converter efficiency. Our method employs a multitask CNN-based neural network that simultaneously compresses and analyzes high-frequency ultrasound data. The network utilizes an AutoEncoder for efficient data compression, preserving essential features while reducing storage requirements. A custom loss function based on the characteristics of the signals optimizes both signal reconstruction and efficiency prediction, ensuring precise and reliable results. Experimental validation on a custom-built test bench demonstrates the effectiveness of this technique with a root mean squared error of 0.78%, being the first known exploitation of ultrasound measurements for efficiency estimation in power converters.