Innovative approach to reducing energy consumption in ultrasonic tomography thanks to machine learning

A research team composed of experts from the Netrix S.A. Research and Development Center and Lublin universities has developed an innovative ultrasonic imaging method, the aim of which is to significantly reduce energy consumption while maintaining high precision of object detection. Using the potential of machine learning, scientists have proven that it is possible to reduce the number of measurement sensors even several times – from typical 8-16 to just 2-3 – without significant loss of measurement quality.

The new approach meets the growing demand for energy-efficient diagnostic systems, especially in industrial and medical applications, where long-term monitoring requires low-power solutions. The key element of the project was the use of algorithms such as Extra Trees and k-NN, which enabled detection and localization of objects with an accuracy exceeding 96%, even with the minimum number of measurement channels.

Reducing the number of sensors from 16 to 3 allowed for a four-fold reduction in energy consumption, while the configuration based on two sensors resulted in even seven-fold savings. Despite such a significant reduction in the measurement infrastructure, the system demonstrates high efficiency in locating objects. Classification algorithms achieve an accuracy of predicting the number of objects at the level of 99% with three sensors, while regression models such as Gaussian Process Regression allow for determining coordinates with a very low mean square error – at the level of 0.68 for the minimum configuration.

The developed technology has potential applications in many areas. It can be used in industry for pipeline inspection and detection of discontinuities in materials, in medicine as an element supporting diagnostic imaging with minimal power consumption, as well as in remote monitoring systems, where energy constraints pose a significant challenge.

The research results show that integrating ultrasound tomography with machine learning algorithms not only enables miniaturization and optimization of imaging systems, but also opens the way to their wider application in low-power environments. The next steps in the development of this technology may include the use of deep neural networks and optimization for real-time operation, which will further increase its usefulness in industrial and clinical practice.

The full publication describing the research results is available at: