One way of looking at IoT deployments is this – a large array of not-particularly-sophisticated endpoints, mindlessly sending individual data points like temperature and pressure levels to either an edge device somewhere on a factory floor, or all the way out to a cloud back-end or data center.
And that’s largely correct, in many cases, but it’s increasingly not the whole story – IoT endpoints are getting closer and closer to the ability to do their own analysis, leading to simpler architectures and more responsive systems. It’s not the right fit for every use case, but there are types of IoT implementation that are already putting the responsibility for the customizing their own metrics on the devices themselves, and more that could be a fit for such an architecture.
There are three main areas where letting the endpoint do its own data analysis – in whole or in part – is becoming increasingly common – smart cities, industrial settings and transportation.
IoT in smart cities
In smart cities smart cameras can do certain kinds of analysis right there on the device, helping planners understand pedestrian and motorized traffic patterns.
The difference between doing analytics completely on an endpoint device or partially on a device is an important one, according to Gartner research vice president Mark Hung. At the core, the analytics done by IoT implementations is about machine learning and artificial intelligence, letting systems take data provided by smart endpoints and fashion it into actionable insights about reliability, performance, and other line-of-business information automatically.
Applying the lessons learned from sophisticated ML is easy enough, even for relatively constrained devices, but some parts of the ML process are much too computationally rigorous to happen at most endpoints. This means that the endpoints themselves don’t change their instructions, but that they provide information that can be used by a more powerful back-end to customize a given IoT implementation on a per-endpoint basis.
The case of video analytics for smart city applications like traffic monitoring – using a system where the cameras themselves track pedestrians and motorists, then score that data against a centrally-created AI model – is an instructive one.
Every intersection is different, so trying to push an identical rubric for making sense of different traffic patterns and volumes to the cameras monitoring every intersection isn’t going to work. Each intersection needs its own rubric. Yet the AI training that’s needed to come up with them requires heavier computational lifting than the cameras alone can provide, so it has to be done somewhere in the back end. The cameras themselves can apply the lessons learned by the AI model, but they need more powerful hardware to intelligently change the instructions they’re given.
“So, to come to some preliminary analytics at the endpoint and then send that back out for further training, you have kind of a federated learning [system],” Hung said.
Another key area for endpoint-based IoT analytics is the industrial and manufacturing sector. Joe Biron is the CTO of PTC, a Needham, Mass.,-based software company that makes ThingWorx, an industrial IoT software platform. Biron said that PTC’s been trying to get intelligence into industrial machinery for about a decade now, with the idea being to help companies save money via predictive maintenance and other automated management and operational applications.
“Ten years ago, the state of technology for doing proactive and predictive failure detection … wasn’t exactly a life-changing kind of technology,” Biron said. It was largely a human-intelligence-driven process that relied on a technical specialist’s intimate knowledge of how the industrial components worked. Based on that knowledge, rules for detecting the parameters that predict impending failure could be hard-coded into even the “dumbest” of endpoints.
The real challenge comes when there’s no one person familiar with the critical confluence of indicators that predicts a problem in the offing. For this, you need machine learning, and more specifically, a machine-learning model that can score data inputs vs. outcomes and sift out which data points are the most important to making the predictions. That’s computationally expensive, according to Biron, limiting its ability to be handled on endpoints.
“Once you’ve created the model, however, now you’ve got something very lightweight to score against; now this model can be fed real-time or near real-time or micro