One subset of the internet of things – the industrial IoT – adds new capabilities to operational technology including remote management and operational analytics, but the number-one value-add so far has been predictive maintenance.
Combining machine learning and artificial intelligence (AI) with the deep pool of data generated by the flood of newly connected devices offers the opportunity to more deeply understand the way complex systems work and interact with each other.
RELATED:
And that can promote predictive maintenance – with the ability to pinoint when components of industrial equipment are likely to fail so they can be replaced or repaired before they do, thereby avoiding more costly damage and downtime.
Fine tuning IIoT predictive maintenance models
According to Wael Elrifai, senior director of sales engineering and data science at Hitachi Vantara – the company’s IoT arm – one of the complexities of predictive maintenance is that AI-produced models for system behavior have to change over time. He used the example of a Hitachi Vantara railway customer with a 27½-year maintenance contract to illustrate the issue.
As train parts age, they respond to stresses differently than they do when they’re new. Because of that, maintenance schedules should be adjusted over time to take into consideration changing failure rates. These schedules can be generated with models that are the output of machine learning, he says.
There’s a “bathtub curve” to equipment failure, Elrifai said. At the beginning of its service life, there are frequent failures, but maintenance processes get figured out as time passes, so failures become much rarer. “And then, of course, end-of-life – it starts to fail a lot again,” said Elrifai.
This type of AI-produced model can be created for other industries as well, and Hitachi has just released a platform called Lumeda that pulls in IIoT data that data scientists can use to adjust their machine-learning models more precisely. “It’s all about being able to monitor machine-learning-model accuracy after a model goes into production,” said Arik Pelkey, senior director of product marketing.
One example is a chemical-manufacturing process. Lumada creates a centralized data pool on which data scientists can experiment, so the process of testing different models against each other means that the company can change its inputs and get a more accurate prediction of what’s going to happen to the chemicals at the other end of the production line.
Elrifai and Pelkey said that the biggest impact that evolving machine-learning-model management will have will be on low-margin, high-capital businesses, like heavy industry and transportation.
