Machine learning (ML) has become one of the most prominent industry buzzwords of the last decade, with almost half of Australian companies now experimenting with the technology.
Within rail, ML has been particularly transformative, with dozens of applications in maintenance, asset management, and other safety critical operations.
However, while many may recognise the value of ML within rail, fewer understand the mechanism by which it works. This makes it hard to imagine future applications, leaving much of the technology’s potential untapped.
Within rail asset management, this is no small matter, with ML tipped to save individual companies millions in maintenance costs, and markedly reduce safety incidents, by flagging asset defects earlier.
To address this, Jaya Negi of Arup is lifting the bonnet on her own machine learning research, in a bid to give rail businesses insight and confidence before deploying ML tools. Negi’s research analysed deep learning tools within civil engineering, specifically, to optimise image object defect detection in concrete structures.
“I want to debunk the myth that ML is some kind of witchcraft and drive home the message that it is something everyone can use,” she said, ahead of the ARA NZ Rail Conference. “In simple terms, ML is where algorithms learn how to perform tasks by generalising from examples. In our case, that meant feeding the technology with hundreds of images of structures and ‘teaching’ it how to recognise defects.”
To do this, Ms Negi and her team went through a three stage process. “Our first step was to understand in-depth the functionalities of a Convolutional Neural Network (CNN), including activation layers, pooling layers and their mechanisms. Second was to define the algorithm most appropriate for our project.
“We settled on a ‘bounding box object detection’, which has been widely used in autonomous vehicle applications. We then collated a total of 714 images of various structures throughout Australia and defined appropriate labelling tools to prepare the training data set.”
With these stages complete, Negi and team hope to initiate the next phase of this research – testing the algorithm – to determine the exact location and type of defect within a concrete structure.
“Currently researched tools can detect defects with a confidence level of more than 90 percent,” she said. “This is significantly more efficient and effective than the human brain, meaning we can flag structural defects more quickly and save significant sums in maintenance costs. It also means we don’t have to send engineers out to dangerous sites to perform inspections.”
While the process was time-consuming and thorough, it was simpler than Ms Negi had expected. “You certainly don’t need a team of specialists with PhDs to understand machine learning. It’s something that everyone can do with a bit of insight. And it’s certainly worth the effort.”
Indeed, benefits of using deep learning in image recognition include time and cost savings in the collection and analysis of data, incentives for long-term assessment through predictive analysis, and automatic regulation of databases.
Although Ms Negi’s research focused solely on concrete structures, she believes transferring these learnings to other rail infrastructure projects is a straightforward process. “With an up-to-date database, you can easily record asset deterioration and implement predictive maintenance protocols with a deep learning tool like this.”
Jaya Negi is due to present at the ARA NZ Rail Conference where she will discuss the technicalities of machine learning, and how to use it for rail asset management. Covering appropriate algorithms and labelling techniques, she will detail the benefits and next steps for rail businesses to use ML for automatic asset defect detection.