Browsing by Author "Rajani, Chang"

Fouling is a large scale problem in industrial equipment such as heat exchangers or pipes, used in factories, ships, airplanes, etc. Traditionally, such equipment is cleaned using sandblasting, chemicals or mechanical methods, all of which require halting the process, which is costly. Recently, high-power ultrasound has become a viable option to these methods. In ultrasonic cleaning ultrasound is projected into the equipment from the outside, which means that the equipment does not need to be halted to perform cleaning. While the cleaning itself is not invasive in nature, in most cases vision cannot be used to determine whether cleaning is actually necessary or not. What remains is to have such a method that is also non-invasive. It is possible to use ultrasound as a kind of a radar to detect whether or not fouling is present, and this has been attempted in previous literature. However, until now, such methods have required extensive manual calculation and knowledge of the physical properties of the setup. We present the first ever system to concurrently clean and detect industrial fouling using ultrasound and deep learning. Our method does not rely on specific properties of the equipment, allowing it to generalize to large industrial processes where it is not practical to calculate or simulate the cleaning scenario. To this end, we extend existing literature on semi-supervised learning by presenting algorithms used to learn from a monotonic process, and model the high-dimensional signal data using a convolutional neural network that is highly robust to temporal variance. This thesis presents the machine learning solution behind the system, and the cleaning components are provided by Altum Technologies. Further, we explore methods to detect and counter the so-called domain shift that occurs when experimenting in the physical world, and provide experimental evidence that our methods work in practice.