Project will transfer AI from the cloud to the IoT device

Digitization of society is one of the prerequisites for achieving the climate goal of 70 percent CO2 reduction by 2030. And the small sensors (IoT devices) installed in e.g. buildings, heating systems, and treatment plants will play an important role in managing energy consumption, heat, indoor climate, etc.

In a new project 'Embedded AI' - supported by the national research center DIREC - researchers together with industry will investigate how to develop AI (artificial intelligence) that can be implemented in IoT devices so that they can do more themselves. Today, sensors are dependent on AI algorithms on cloud platforms or decentralized networks (Edge Computing), where data and commands are sent via internet / wireless networks.

"It is quite obvious that you will not be able to do the same as with the cloud and edge, but it will cost less, use less energy and be able to react faster. It will also increase security and privacy because data can be kept where it is collected. So there are many benefits to embedded AI, says the project manager, Professor, Section Manager, and Deputy Director at DTU Compute Jan Madsen.

In the project, DTU, Aarhus University, the University of Copenhagen, and CBS collaborate with the pump manufacturer Grundfos Holding, the engine and machine manufacturer MAN Energy Solution, the window manufacturer VELUX, and the technology company Indesmatech.

Move AI from large platforms to small ones
During the three years, the partners will work on specific issues within the four industry partners. They are strong representatives of companies that will be able to strengthen competitiveness by knowing the right tools and platforms to leverage embedded AI (eAI) in their products.

The project will examine the process of going from large platforms to small ones, explore suitable tool platforms, check what opportunities new types of chip provide for embedded AI and map out how embedded AI will be able to change the business models for companies.

“A lot is happening within research in embedded AI. The difference between that and our project is that we work closely with the companies and their issues and their visions of where they want to go with embedded AI. We will find something specific for each case, but we will also identify what is generic and applies across companies,” says Jan Madsen.

Grundfos is experiencing a knowledge gap
The idea for the DIREC project has come through network meetings, where research institutions and industry talk about future competencies and technology needs. Here, Thorkild Kvisgaard, Head of Electronics, Director Technology Innovation at Grundfos, has participated.

He says the company sees a clear need to be able to move some of the artificial intelligence from the large platforms that run on mainframe computers, etc., down and run in more embedded devices (AIoT), even though it will be very resource-limited platforms to work on. Because you can save energy, and you avoid having to send data over the Internet and be dependent on the Internet and cloud solutions that run outside your own control.

“It will, of course, turn out that you can not do quite as much on platforms with limited resources, but we do not know those limits today. And maybe we can do a lot more than we think. If we work with something that is not time-critical, it does not matter that the embedded AI has to spend several minutes figuring something out if it is a slow and complex process,” says Thorkild Kvisgaard.

“At Grundfos, we have experimented with the technology ourselves, but we are experiencing a gap between what data science experts work with on large cloud platforms and what IoT programmers work with. So we hope that the project will also create a better understanding of each other's work areas.”

Chip becomes crucial
The industry partner Indesmatech acts as both the local office for chip manufacturers, facilitates various development projects with new technology and helps companies to develop technology.

The company is looking forward to clarifying the possibilities when working with Embedded AI algorithms, explains Co-founder of Indesmatech Rune Domsten:

“What is interesting about the Embedded AI project, in addition to the software used for AI, is to investigate which chip and hardware platforms to execute on and use in the different situations. Because the battery consumption in sensors really depends on which chip you use, and it can be a question of whether the battery lasts for e.g. five or ten years.”

Although the industry partners in the DIREC project like large companies are already working on AI, the project could also have great significance for especially small companies that lag behind with artificial intelligence, says Project Manager Jan Madsen:

“While it may seem rather uninteresting in research to develop small AI algorithms, there are actually major research challenges in developing efficient architectures and methods that can be used in smaller and resource-limited sensors / IoT devices. It can also be what gets a small business started using AI for complex tasks and processes.”