The Challenge

LINC with Sellafield Ltd is a scheme that encourages SMEs at local and national level to collaborate and deliver innovative solutions to support the mission at Sellafield. LINC challenge 42 was titled ‘Image Processing for Assessing Package Integrity using Machine Learning’ and set the challenge as follows:

“The long-term storage of nuclear waste is at the heart of the nuclear industry in the UK. As a requirement this material must be examined on a regular basis, which generates a vast amount of data to be reviewed. Presently this is done manually and takes a lot of time and is vulnerable to human error.”

Cerberus Nuclear’s Data Science team proposed the development of software containing a trained machine learning computer vision model that would be capable of automatically recognising issues that could affect integrity of the package. Our team use the ‘Agile’ design methodology, which incorporates software testing in short, focused development cycles; ideal for the project.

We were delighted when Sellafield Ltd chose our solution ahead of some tough competition.

ANDI: Automated Nuclear Damage Inspection

ANDI is a high-quality user-friendly software program that utilises computer vision machine learning for automated identification of damage from externally supplied video.

The software allows detailed examination of product can inspection videos, automatically identifying damage such as scratches, dents and corrosion. The neural network within ANDI uses a cutting-edge R-CNN approach for image analysis and was trained using previous examples of damage.

Damage identified is highlighted within an embedded video player, which allows users to quickly skip to areas of interest and examine results frame by frame to inspect the exact moment(s) that damage has been detected. The confidence level of identified damage can be customised by the user with damage highlights switched on or off to assist with detailed inspection.

An integrated inspection report system was incorporated into the software to allow users to make notes and log frames for easy follow up review. The software also allows the processing of multiple batches of inspection videos with minimal user interaction. This allows the review of multiple processed results within a single session.

The algorithm for the Sellafield challenge uses an extension of the R-CNN called the Mask R-CNN. The R-CNN algorithm (Region based Convolutional Neural Network) can detect and classify objects within images, it focuses on variations of colour, texture and scale within an image to form a region.

“I really like the look and feel of the software and I’m impressed how well the neural network is identifying the key elements of damage, it’s very good!”

Gareth Myers, Technical Researcher, Project Lead, Sellafield Ltd

“We are delighted that Cerberus Nuclear helped make a difference at Sellafield Ltd. The Data Science team have delivered a great solution, bringing modern techniques to the nuclear industry.”

Daniel Cork, Director, Cerberus Nuclear

Sellafield Ltd are currently using ANDI to enhance their damage inspection work flow which, prior to using the software, had taken many man-hours to identify, categorise and log.

Cerberus Nuclear are proud to announce ANDI has recently been a key feature for inTechBrew. inTechBrew promotes the latest high Technology Readiness Level (TRL) nuclear industry innovations across UK and Europe.

Object Recognition

Machine Learning
Computer Vision

Cerberus Nuclear’s Data Science team built upon previous experience when developing the computer vision algorithm used for ANDI.

Previously, the team developed a custom computer vision object identification algorithm to identify cars and lorries on a motorway using R-CNN object detection method.

The algorithm identified the number of objects, object type (car, lorry, etc.), object colour, object speed and confidence in match. In addition, a report was generated that summarised the information gathered over time. The development was to demonstrate validity of use for stopped car identification as well as traffic monitoring purposes.

The Sellafield Ltd LINC challenge aligned well with the previous development work already performed and paved the way for the creation of ANDI. Additional technical challenges such as variable lighting, frame blur and reflections had to be overcome as well as creating a custom user-friendly interface that met with Sellafield Ltd requirements. The processing time for the computer vision algorithm was also enhanced.