LEARNING IN THE DEEP

LEARNING IN THE DEEP: MACHINE LEARNING FOR CONDITION ASSESSMENTS

From Siri, Apple’s personal digital assistant, booking in your schedule, to self-driving cars realising that a pedestrian
is about to cross the road; working behind the scenes, machine learning is changing diverse aspects of our lives, making
daily tasks easier and faster.
Although machine learning can be applied to a wide range of applications, Abyss Solutions is using this powerful
artificial intelligence tool to innovate the way that condition assessment is undertaken.

What is machine learning for condition assessment?
With the primary focus on image classification, Abyss’ team has designed computer vision and deep learning algorithms
able to find patterns and features in large-scale data, on the terabyte level. These patterns could be cracks on a dam
wall or corrosion on different parts of a bridge, and thanks to this technology, they can be automatically identified in the
whole data set.

As a result, Abyss provides an accurate data analysis in minimal time, facilitating the generation of
cost-effective condition assessments!

Abyss’s case study
Abyss was recently faced with a challenge after the completion of more than 30km canal inspection. With 10
operation days and hundreds of hours of videos collected, generating approximately 1.5TB of output, the sheer
volume of data was overwhelming.

Let’s assume that all the collected data is going to be manually processed and analysed by engineers. Let’s assume
also that they would take around 10 seconds per image to find a fault/critical features.

How long would it take to access all the data?
It would require at least 72 working days!
Despite all the time spent, this manual data analysis is also subject to human fatigue and error.

Abyss’s Machine Learning Algorithms
Abyss team has developed algorithms that identify critical features automatically, resulting in a more accurate condition
assessment in less time.

Check out some results of this development, there are quite impressive!

An engineer feeds a small portion of the collected data into the system and in just 5 minutes,
the computer learns to detect certain patterns present in the data!
Abyss’ machine learning algorithms take
approximately 0.1 seconds to identify and locate  
faults in each image, with over 95% accuracy.

Do you remember the 72 days that would be expended to analyse all the collected data?
With Abyss’ machine learning algorithm, the results came in less than one day!
This enables Abyss to rapidly generate a ‘google earth’ view of hot spots along the 30 km canal.