Real-Time Fouling Tomography with Neural Networks

What is it about?

Many industrial systems, such as pipes in energy, chemical, or water plants, gradually collect unwanted material on their inner surfaces. This accumulation, called fouling, can reduce efficiency, increase costs, and even cause system failures. However, detecting where this fouling is located inside a closed structure is very difficult without opening it. In our work, we present a new way to see inside such systems using ultrasonic waves in a non-invasive way. We send ultrasonic waves through the structure's surface and analyze how the signals change as they travel. These changes contain information about where fouling is present and how the fouling spreads. Instead of relying on expensive measurements and costly annotation of such measurements, our method learns from simulated data based on probabilistic methods and physical laws. We combine this with a neural network that can reconstruct a detailed map of fouling from just a few sensor readings in milliseconds.

Featured Image

Photo by Quinten de Graaf on Unsplash

Photo by Quinten de Graaf on Unsplash

Why is it important?

Our work is addressing the challenge of understanding what is happening inside complex systems without direct access, costly measurements, large labeled datasets, or computationally expensive algorithms. In the context of pipelines, this enables real-time detection of fouling using only a few sensors, which can reduce maintenance costs and prevent failures. Importantly, the contribution goes beyond pipelines. The core idea, learning from simulated data and using neural networks to solve inverse problems, can be applied to many other domains where direct measurements are difficult or impossible.