Internet of Things (IoT)
Showing results for 
Search instead for 
Did you mean: 

Bringing Artifical Intelligence (AI) and Machine Learning (ML) to the industrial edge



Mike Shaw profile pic.jpgBy Mike Shaw
Director Strategic Marketing, HPE


When I was a student some thirty years ago, I spent my vacations working for a company that developed industrial control systems. They could actually be regarded as rudimentary industrial IoT systems -- we would take readings from sensors located inside the machines, parse them through  simple processing scripts, and then send the outputs over a wire to control equipment on the factory floor. For example, I made a system that controlled a furnace, based on inputs such as temperature and time.

How times have changed! Today’s industrial IoT systems are far more capable on a number of dimensions:

  • Sensors are much more accurate, and can measure different types of inputs, including pressure, friction, vibration, and appearance.
  • Communications greater capacity and reliability, enabling wired and wireless connectivity over longer distances.
  • The ability to handle all of the data flowing in from the edge has massively improved.

This last point does not only refer to faster processing, larger storage volumes, and more bandwidth. It also refers to the increased ability to derive insights and take immediate action based on them automatically with Artificial intelligence (AI). This latent technology trend has been re-ignited by the explosion of sensors’ data – Industrial IoT data - and  has started to transform operations at the edge, leading to new and exciting possibilities.

For instance, maintaining a small parts inventory is crucial to companies that use lean production techniques and just-in-time manufacturing. AI based on mathematical formulas and warehouse inputs can help automate restocking and supply-chain workflows. Optimizing inventories is a huge cost saver, by ensuring that there are no parts shortages or write-offs that result from stocking parts that are never used.

According to a recent IDC report, in 2019 43% of IoT processing will take place on the edge. Further, AI will be part of all IoT effective implementations by 2019, the report says.

Programmed Artificial Intelligence vs. automated Machine LearningGettyImages-667612239_super_800_0_72_RGB.jpg

Let’s take a step back and examine what “artificial intelligence” means in an industrial setting. Inventory optimization based on probability functions is a basic example. You may also be familiar with programmed AI, in which a milling machine, robotic arm, or mining vehicle takes action based on pre-programmed routines, basic inputs, and rules-based decision trees. Once certain conditions are met, the process may be repeated, or a new process may be launched.

There are limitations to programmed AIs, starting with the fact that it’s very hard to program an AI to handle complex tasks with lots of inputs. If a robotic arm next to a conveyer belt picks up a defective part, or a part that’s upside-down, how will it “know” that there’s a problem, and respond accordingly? Or, if a person, a large piece of machinery, or a small boulder sits in the path of an autonomous mining vehicle on a pre-programmed route, how will it recognize the obstacle and either stop or drive around it?

Human developers might spend several weeks to program a simple routine, but as requirements and inputs increase, the development timeline extends into the months or even years. Some scenarios simply cannot be programmed, because they are just too complex for humans to understand, much less develop algorithms that can process them.

This is where machine learning (ML) comes in. An application of AI, machine learning enables computers to be taught--or even self-learn--how to recognize things, whether it’s a defective part or a person standing in the middle of the track. Machine learning requires a lot of training data to create a model--thousands or even millions of examples, depending on the task at hand. But once an ML algorithm has passed the initial training phase, it can use new data to update its models.

Another benefit to machine learning is because the learning continues as it encounters new examples, ML systems get better and better over time. This ML approach is known reinforcement learning. Moreover, the learning need not be actively managed by a human.GettyImages-569263615_800_0_72_RGB.jpg


Recently, Carnegie Mellon researchers developed a poker bot designed to beat human players in Texas Heads-Up No Limit poker. For AI developers, this is a complex game--unlike the games of chess or Go, poker players don’t know their opponents’ positions. Guessing and game theory is a central part of the game.

Last year, when the bot played against four top players, it only beat one of the players and lost $700,000. This year, it beat all four players and won $1.7 million. In the intervening period, the researchers didn’t do any additional programming on the ML system, which runs on HPE high-performance computing hardware. Rather, the system continued to learn on its own, by playing against itself and learning from the matches.

Reinforcement learning is not only used for poker. In fact, the Carnegie-Mellon poker bot can also be used as a negotiator or a bidder in an auction, situations in which an opponent’s positions are not well known. This approach also has huge implications for industrial ML. Applications will get better over time as ML systems are exposed to more and more data.


Video Analytics: Machine Learning in Industrial IoTManual QA testingManual QA testing

Consider the QA processes at the end of an assembly line to determine that a product has been properly finished. Of course you can have a human being standing there, and making sure that the component is placed into the box in the right way, and the flaps are completely closed and properly sealed. But it’s repetitive and prone to error if the inspector looks away or gets tired.

Wouldn’t it be better to reassign the worker to handle something that requires more skill, and have an AI handle packaging QA? With ML, that is possible, once the system has been taught to recognize what is a good quality box and what constitutes a bad packing job. The will infer with a certainty of 89% that a particular box is not well packaged. The box in question will be taken out of the line for a human inspector to examine, who will make the final determination to pass QA or send the item back for repackaging.

Over time, the QA system will get better and better at detecting badly packed boxes. Quality will increase, while less human time will be required to make final checks.

How to get started with AI and ML in Industrial IoT

This is where the value of partnerships comes into play. HPE has established partnerships with companies that can bring platform and application know-how. They include GE Predix, a platform for developing industrial applications and analytics, as well as PTC’s software and agents for anomaly detection, predictive analytics, and IoT connectivity. HPE’s service arm, HPE Pointnext, can also advise companies interested in implementing ML solutions.

If you’re interested in learning more about machine learning and other AI applications in IoT, I will be speaking at HPE Discover in Madrid. Here are a few of my sessions:

Also, here are some demos showcasing AI and ML applications at the event:

I hope to see you at the event!

Related reads:


Follow me on Social media:

twitter.gif @mike_j_shaw

linkedin.gif Mike Shaw


HPE Edgeline Converged Edge Systems are powered by Intel Xeon.

Empowering the Digital Enterprise to be more efficient and innovative through data-driven insights from the Internet of Things (IoT)
0 Kudos
About the Author


Jan 30-31, 2018
Expert Days - 2018
Visit this forum and get the schedules for online HPE Expert Days where you can talk to HPE product experts, R&D and support team members and get answ...
Read more
See posts for dates
HPE Webinars - 2018
Find out about this year's live broadcasts and on-demand webinars.
Read more
View all