SAN DIEGO—The future of manufacturing is “an awful lot like the fulfillment centers of today,” Sarah Cooper, general manager for IoT solutions at Amazon Web Services (AWS), said last month at Industry of Things World here. “Kiva robots running around, big pickers, vertical stacks. It’s really cool how many of these areas are coming together. And that even though these are very different problems we’re trying to solve, … we see a lot of the same fundamental technology pieces in our own fulfillment center looking at route optimization for robotics or route optimization and planning for goods throughout the system.”
A lot of manufacturers feel they are behind in robotics, automation data analysis and especially machine learning, and that makes sense, she said.
“One of the core reasons, is irony: Machine learning on machine data is extremely difficult—for a couple of reasons,” Cooper said. “One, in order to be able to build equipment that can go onto almost any manufacturing floor …, you’ve got to build equipment that sends very simple data that can get recombined in all sorts of wonderful ways, depending on how it’s been deployed and the current process it’s running.
“Another big challenge is the data is crap,” she added. “Even the expensive sensors periodically fail. And when they fail, they fail beautifully. Right? Those thermocouples that go to like 10,000 degrees. And then they go negative, and then they come back. You’re asking, ‘OK, was the reset successful? Is this a real value?’ And then it starts to drift a little bit. And then you think, ‘Well, that could still be information. Maybe it’s not that the sensor went bad.’ So, you’ve got to figure out how to make your analytics responsive to the crappy data.”
When cleaning the data, it’s important to remember that with certain machine learning algorithms “if you remove outliers, you’re removing information,” Cooper said. So it’s important to bring together tools that are sometimes sold separately.
Manufacturers cannot rely solely on data scientists to think about the automation needed to get to predictive maintenance. Data scientists are not trained to think that way.
“They’re in research roles, which means they typically go in and solve the problem once,” she said.
To get to the point where you have a predictive maintenance model running against 5,000 compressors on five-minute rotating intervals, all of the data crunching “needs to be in one tool that has been abstracted so you don’t have to care about what an underlying data structure,” Cooper said.
Makers of beer have learned how to do in-line quality control that led to continuous improvement, the elimination of manual-measurement stages and dramatically increased output, she said.
“That’s the same thing we’re doing with automotive OEMs,” Cooper said. “We’re asking, ‘How do we not pull something off the line? How do we do rework in line? Where should machine learning happen?” We have this conversation 15 times a day: If I talk to 15 customers, I have this conversation 15 times a day. You’ve got to run your workloads where the data is.”
Everyone calls Amazon Web Services a cloud company,” she added. “Yeah, we do cloud. We also do a bunch on the edge.”
One machine-learning-enabling enabling technology to pay attention to in the future is acoustics, Cooper said.
“Acoustics are taking over the world,” she said, showing the audience in San Diego an example from Siemens, in its MindSphere environment. “They worked with their customer Ham-Let, valves, using a microphone on that valve to listen to the ‘ssssss’—to detect how much material had moved through it,” as well as any leaks, Cooper said.
“Again, machine learning was running on this valve,” she said. “Super cool.”
As manufacturers explore machine learning, they would do well to keep in mind all the ways it is useful, Cooper said, including anomaly detection, language processing, object identification, prediction and forecasting and route optimization.