Sophisticated image recognition helps reveal what is hiding under your microscope
While Internet of Things (IoT)-enabled smart sensors are becoming more relied upon for minute-by-minute operation of today’s water reclamation facilities, the microscopic evaluation of activated sludge is still a useful tool, especially for troubleshooting solids separation issues. Excessive growth of filamentous bacteria (filaments) is a common cause of poor settling and the morphologies of most filaments are well characterized. Most are associated with either specific substrates such as organic acids, sulfides, or FOG or certain selective pressures such as low DO, low F/M, or a deficiency of either nitrogen or phosphorus. If the most abundant filaments are identified, their causes can be investigated and corrective operational actions can be implemented.
Determining filamentous abundance is relatively simple, but specific training is required to properly identify filaments. Samples can be mailed to consultants or experts can be called on site, but this can take days to organize and arranging for samples to be mailed can sometimes take weeks.
Microscopic Evaluation Made Easy
To simplify and expedite this process, Novozymes has developed Plant Assistant, a web-based application that can identify and score common wastewater filaments from a microscopic image.
“Previously, people would send us samples from all around the world, which is both expensive and time consuming,” said Chris Flannery, technical service manager for Novozymes.
With Plant Assistant, users simply need to obtain a microscopic image of the wastewater microbiology and upload it to the app from their preferred device. From the image, Plant Assistant instantly determines the most likely filament, scores the abundance, and prescribes a course of action based on the types of filaments found.
Data is Key For Accuracy
Data scientists and biological experts have collaborated on the app development, while using thousands of different microscopic images to train the algorithm to identify filaments and differentiate around characteristics such as branching, cell shape, and diameter.
“Any object a human can recognize an algorithm can do as well. It just requires enough training data — in this case, images. And plenty of them,” said Mathias Gruber, senior data scientist and AI Lead at Novozymes. The more sophisticated the object is, the more training data is needed to achieve a highly accurate match. Filaments, for instance, move in three dimensions while images only capture two. Furthermore, there are differences in microscope type and quality, magnification degree and experience of the user, which all adds complexity. Reaching a critical mass of data that takes all these variations into account has been the key goal for the development team.
Currently, Plant Assistant can identify eight filaments with more than 90 percent certainty. Plans are underway to expand Plant Assistant’s algorithm to identify approximately 20 filaments as well as characterize floc particles and bulk water, and identify higher life forms. WW