Recycling: Automating the sorting and separation of e-waste.
My name is Ben and I’m an intern at Cambridge Consultants. Each year all the technology scholars and interns at Cambridge Consultants - read more about the program - are given the opportunity to run a project by ourselves, while being mentored by senior engineers and business developers.
Ideas for different projects are pitched by sponsors, and we choose the one we want to work on. I immediately chose an e-waste recycling project. The current processes for e-waste recycling are inefficient, wasting resources and money, read more about this here.
Our project brief was to come up with a system that could improve the profitability and efficiency of the e-waste recycling industry through accurate grading of Printed Circuit Boards (PCBs). This lead us to our project specification which was to use machine vision and machine learning as a core technology, and build a piece of equipment that could predict the value of PCBs and categorise them accordingly.
Working in a team of software engineers, my job was to produce an algorithm that approximates the embedded material value of a waste PCB. Part of this algorithm involved using machine learning to pick out individual components from a single photo of a PCB. Machine learning can be used to identify patterns in data, without rules or specific instructions. This proved to be an enabling technology for this project, as we were able to train a neural network to identify individual electronic components within an image of a PCB.
As a software intern, I did not have experience in developing machine learning algorithms, but I was able to draw on some great expertise within Cambridge Consultants. Since there were no existing solutions for recognising discrete electronic components on a PCB with a machine vision system, this challenge required detailed planning to define a novel approach.
Some of our algorithms and analytics specialists were able to help me define an initial strategy, and I was given access to one of our high-performance compute (HPC) servers. When training the machine learning model, access to this HPC proved to be invaluable as I could iterate through different datasets and training parameters in just hours rather than days. We settled on an approach to detect components using a model with Faster R-CNN architecture, pre-trained then augmented with our own data. Preparing training data for the neural network took some time, but we found ways to auto-generate extra data and augment our current data, making the process much easier.
It was a really valuable learning experience contributing to the project and being able able to work in a motivated team of different disciplines. Overall I really enjoyed the opportunity to develop a machine learning system for an unmet need. Being able to draw on Cambridge Consultants’ existing capabilities allowed me to learn about and create algorithms at a low level, whilst having guidance at a strategic level.
The work on the project also lead to a white paper being written on the subject of how automation could improve the pre-processing of e-waste. A copy of the whitepaper can be downloaded here.