Security and networking were industrial IoT's top challenges. Now there's a third: Practical AI
Useful machine learning was promised for factories and similar installations. In reality, the tech is still some way off
Sponsored Some would have us believe that the whole of Internet of Things will soon be artificially intelligent. Not only do we not think that's true, we think the Industrial Internet of Things (IIoT) – the part that does the meaningful work – will take longer than the rest of the connected device industry to acquire those AI features.
IDC in November 2016 made a lofty prediction: some form of AI would make its way into all IoT deployments by this year. “AI, Everywhere,” IDC proclaimed. “By 2019, 40 per cent of all digital transformation initiatives, and 100 per cent of all effective IoT efforts, will be supported by cognitive/AI capabilities.”
A Skynet crafted from smart kettles and door locks by 2019? There’s a scary thought. IDC cited a new acronym for AI-powered IoT – AIoT – and for its 2017 North America IoT Analytics and Information Management Survey, IDC reckoned 31 per cent of organizations that deployed IoT analytics were using machine learning in it.
Today IDC is contrite, admitting the prediction wasn't feasible. “The earlier prediction was too optimistic,” an IDC spokesperson told us.
The rise in consumer-based “smart” AI no doubt fueled IDC. While a sprinkle of Siri can make your smart speaker “cognitive," getting it to remind you to pick up groceries with the power of your voice, the world of industrial devices has a different set of standards for AI.
AI in IIoT doesn’t mean neat-but-limited voice recognition tricks or home automation. It means systems running applications that can tell you when they will fail, and when they will need maintenance, based on a careful reading of sensor data via analytics and machine-learning models. The underlying mechanics of that are more difficult than some analysts would have us believe.
No wonder that a 2017 Cisco report found just 26 per cent of manufacturing companies considered at least one IoT initiative a complete success. Overall, 60 per cent of respondents said that IoT initiatives looked good on paper but proved more difficult than anyone imagined. A Bain & Company survey of 600 high-tech execs found despite ambitious long-term goals, customers have struggled to implement IoT and now believe it will take longer to reach scale than expected.
So what’s holding them back? Security and infrastructure continue to be the main challenges, but if we’re talking intelligent IIoT, then surely we have a new infrastructure problem to chuck on the pile.
IIoT security topped the list of worries for those considering IIoT according to Bain & Co – a concern for around 40 per cent.
One concern is network complexity. Understanding how each individual device connects to the network and to the broader internet is crucial. What happens if you fail to secure a vulnerable network service or route, used by one of your thousands of IIoT sensors to reach the outside world, leaving your industrial network open to attackers? What if an IIoT device is accessible via a link to an internal network segment that you didn’t notice? This system complexity could lead to accidental exposures.
Another worry is device complexity. IIoT devices could number tens of thousands on a network. You must ensure that each is secure, so that attackers can’t compromise or impersonate them. That means unique digital keys on each device. Configuring tens of thousands of devices with their own unique digital keys and then managing those device IDs on the network is a daunting task.
You must also, of course, keep devices, and the equipment and services that glue them all together, patched and up to date, software and firmware-wise.
One of the biggest means of attacking anything running software is exploiting a known, but unpatched, vulnerability. The difficulty in patching IIoT systems out in the field topped the list of concerns for 56 per cent of techies participating in the SANS Institute's 2018 Industrial IoT Security Survey. The Ponemon Institute, meanwhile, surveyed 3,000 cybersecurity pros last year and found that 57 per cent of network intruders exploited a known vulnerability the victim hadn't patched.
“The staff at plants are already overwhelmed with security hygiene tasks for existing assets,” SANS Institute explained. “There is no bandwidth for coordinating security patches from a multitude of different OEMs. Likewise, few plants have the kind of secure remote access needed to enable direct management by the OEMs.”
Networks and connectivity
Networks are the other major IIoT challenge – architecting and managing them. With so many potential devices on the network sending telemetry data and potentially receiving control data from a central server, companies face a challenge planning network capacity and managing peaks and troughs in traffic.
Not only is the nature of the traffic a challenge, the terrain it must traverse also poses network challenges. Data has to get from embedded sensors in industrial equipment through a gateway to a wide area network (WAN), and then on to a managed cloud server. There are challenges at every stage. In a plant, connectivity modules may have to cope with electrical interference, and the equipment may be remote thereby making it difficult to get data to the gateway and then out to a cloud-based server.
There are ways around this, but they add to the complexity. You can mix in cellular and satellite technology, though IIoT design teams must plan for outages and latency when it comes architecting this kind of set up. One potential way around this is to build a mesh network, where multiple nodes in a local network communicate wirelessly with each other, so that if one fails a connection can still persist. Another for wide-area communications is session migration, in which a team designs a gateway with multiple connections including satellite and cellular. Devices can then move between them should one experience packet loss or jitter. You could adopt a model of edge computing, where the compute power to process IIoT-generated data is close to the source of that data – a model that reduces the need to send data to the network’s core for processing, thereby sidestepping any latency and connectivity.
High-frequency short-range communications could help companies put edge-based computing where it’s needed in challenging RF environments while sending information to the cloud over a low-power, wide-area communications link, for example. Then you have 5G. Industry wonks hope that the emerging technology will help create more reliable, low-latency IIoT networks by marrying high- and low-frequency spectrum communications to suit specific connectivity requirements – though, it’s still early days for 5G.
Managing such a complex network means needing to understand performance and resilience in real-time so you can spot connectivity problems before they become an issue. Here, a network management tool like Paessler’s PRTG Network Monitor can monitor events across a variety of links.
One of the drivers of IIoT is data – that the data flowing from your devices can be harvested and analysed. Data from industrial devices will, it’s hoped, help you run things such as plants more effectively. Industrial washing machines will communicate their detergent levels and warn when they need refilling. Data from connected fleets will communicate fuel usage levels to help companies plan routes and reduce petroleum costs. Dealing with data volume is one of the biggest challenges facing organizations with IoT projects, industrial or otherwise.
In its research, Cisco found many respondents already unable to cope with the flow of real-time data from sensors, which doesn’t bode well for the explosion of data in the future. They must not only build the low-level network infrastructure to support these data flows, but they must have the processing power to cope with streaming information. For back-end analytics processing, this typically means distributed parallel data processing using something like Hadoop clusters – but deploying Hadoop is no ordinary technology decision. It’s a platform investment that requires commitment and investment, factors that will complicate its addition to your technology infrastructure.
Machine learning – the new infrastructure challenge
It’s here the infrastructure challenge starts to segue into artificial intelligence. Here, technologies under the AI umbrella, such as machine learning, are being employed to feed systems vast quantities of data so they can begin to think for themselves. By modeling historical data, the hope is that manufacturers, energy companies and other industrial players can make useful predictions or spot patterns in their telemetry that could prevent outages, automate manual activities, and increase throughput.
Thus, we see predictive maintenance cited by vendors trying to push machine-learning models in industrial environments. If we can tell when a motor will fail by running sensor vibrations through a machine learning model, they say, we can save unscheduled downtime. Everyone’s a winner. Another potential application is industrial quality control. Companies are already using image recognition to sort good cucumbers from bad as they roll off a conveyor belt.
But even – yes – IDC admits in its 2019 IoT predictions document that early attempts at machine learning often aren’t that impressive, illustrating the challenges facing those who build AI-powered IIoT. “Often the initial runs on ML have low causality, which means teams abandon the ML efforts or have no ability to act on the prediction,” it says.
Companies don’t just plug in a sensor and expect its data to change the way a factory operates. They must repeatedly retrain the software, massaging the information and the machine learning algorithm that consumes it to get better results each time. Only after many tries will they get results that match their expectations.
Supporting AI and ML means infrastructure investment. Machine Learning with high volumes of IIoT data takes lots of storage and computing power, companies may need multi-tiered storage to cope with the different stages of machine learning. You might store raw IIoT data on a lower-speed medium before moving the information for training to high-speed storage such as flash for training purposes.
IIoT is a promised land. Some people came down from the mountain, handed you a couple of white papers, and told you everything will be great. The fact one of those, IDC, has stepped back is a clear indicator that some real challenges continue to bog down IIoT’s rollout. Security and network connectivity are the primary infrastructure challenges but, taking IDC’s original contention to its fullest, we have another infrastructure hurdle slowing down progress: machine learning. It will take a long time and a lot more work before you get anywhere near AIoT, or whatever it becomes known as.
Sponsored by Paessler.