The next generation of internet of things platforms could be one that allows things to become thinking, learning objects. Imagine that your smartwatch could not only predict when you might be ripe for a heart attack, it could also sense when a hacker was trying to access your personal data.
The way to augment things with a “brain” is to enhance them with artificial intelligence (AI). Let’s call this AIoT, the artificial intelligence of things.
This year has shown peak investment in AI, with startups in the U.S. alone having raised $1.5 billion, and I’m sure we will see the fruits of those investments in our daily lives very soon. To imagine where AI will play a role we need to understand what AI is — and what it is not.
AI is an algorithm powered by statistical models allowing the AI to “learn” through feedback loops. So rather than deterministic models where an algorithm uses predefined rules upon which to base its decisions, other models are applied.
For example, Google makes use of a technique that’s called deep learning; much of the work in this area is inspired by how the human brain works. Those models are no longer deterministic and, as such, could mean that how an AI comes to a certain decision might become opaque. This could give rise to unforeseen situations; witness Microsoft’s AI chatbot that learned to be racist within hours through analyzing twitter feeds.
Will AI become all-knowing? The current AI’s will certainly not, they are trained on specific domains and will not be able to apply that knowledge in other contexts. For example, a recent botnet attack crashed several high-profile websites by infiltrating things such as connected DVRs and cameras. Had they been augmented by AI, the things could have sensed a traffic overload and shut them down.
So where will AI augment IoT? The most likely area will be in manufacturing, an industry that is already spending heavily on IoT. The use case that manufacturing is attacking with AI is predominantly predictive maintenance. The form of AI they are doing this with is called machine learning.
Manufacturers are chasing predictive maintenance because there are some real and tangible benefits; the low-hanging fruit is increased uptime and less unplanned downtime, allowing organizations to lower the cost of maintenance and repair.
But there is more at stake. Having those capabilities will allow manufacturers to adopt new business models to better compete in the marketplace. For example, in some areas there is a need to move from capital intensive investment to more operational investments, from Capex to Opex. So instead of offering a machine for a fixed price, a machine is rented and paid for only when it is used. IoT will allow the monitoring of usage.
A side effect to this is that the manufacturer would not be paid when the machine breaks down, so uptime is in his direct interest, likewise is the lifetime of the goods. If the lifetime can be extended then the margin on the rent will go up. Having predictive maintenance capabilities are essential to reaching those goals.
If predictive maintenance is so important, why isn’t there a full adoption going on yet? Well, there are some steep hurdles. The lack of reliable sensors for monitoring performance and behavior of machines is one, the challenges of getting reliable connectivity into shop floor operations another. Both are prerequisites to collecting the data that is necessary to test the statistical models.
Then there is a lack of statistical models that can predict behavior, largely because of a shortage of data scientists that can build and test those models. And the real world is complex; machines are shipped all over and work under different conditions. For example, the vibration of a machine will be influenced based on the type of floor it stands on; a wooden floor will influence the measurements differently than concrete.
Manufacturers often make many different machines, in different versions and models. Those machines often are constructed on parts that were ordered through different vendors and suppliers. Although designers tend to set the quality to certain standards, spare parts delivered by third-party vendors might behave slightly differently, undercutting the models in unseen ways.
We are certain that deterministic models will be insufficient to deal with the above situations effectively and that the only way forward to tackle these challenges will be AI-inspired, real-time analysis approaches.
All IoT Agenda network contributors are responsible for the content and accuracy of their posts. Opinions are of the writers and do not necessarily convey the thoughts of IoT Agenda.