PI: Pingbo Tang

Co-PI(s): Erica Cochran Hameen

University: Carnegie Mellon University

The proposed PITA project aims at developing a quick, remote, and reliable way to diagnose water treatment plant filters’ structural integrity and potential backwash problems. As parts of the critical water infrastructure, water filters need regular inspections to identify uneven, curved, or misaligned parts of gravel support and pipes inside the filter. These geometric defects in the water filter can cause choppy water to flow through the filtration layers and produce water quality that cannot meet the standards. They occur due to improper “backwash” operations that use poorly controlled high-speed water flow from pipes at the bottom of a filter to clean the filtration media. As these geometric defects are under the top layer filtration media, conventional inspection methods have to punch or excavate top layers. At multiple locations, punches need at least two workers to spend hours inside the water filter, while they could miss defects between sparsely sampled locations.

The project team and AQUA propose to examine the use of 3D laser scanning technology for carrying out non-contact inspections to reduce the time, costs, and errors in conventional field punches of filtration media. The hypothesis is that the filtration media’s exposed surface could have uneven geometric changes because of subsurface defects and uneven water flows. The project team plans to collect training samples from ten water filters and train a machine learning model that captures the correlations between anomalous geometric change patterns on the surface and the hidden geometric defects. This trained model can predict underlying defects based on new surface samples. Data augmentation and transfer learning approaches will help overcome the challenges of producing a reliable machine learning model with limited training samples from ten water filters.