Husqvarna brings gen AI to the factory floor

“Troubleshooting has always been a problem in manufacturing,” says Jonathan Wickström, manufacturing digitalization lead at Husqvarna.

Headquartered in Stockholm, Sweden, Husqvarna has been around since 1689 when it manufactured muskets. It produced its last firearm 300 years later and today makes outdoor power products such as robotic lawn mowers, chainsaws, and trimmers, as well as equipment and diamond tools for the construction and stone industries.

Today, the company, working closely with partner Microsoft, is leveraging gen AI to reduce unplanned downtime on the factory floor when technicians and operators must troubleshoot issues with equipment. Enter the AI Factory Companion, a gen AI copilot that helps technicians and operators diagnose and resolve issues with machinery. They can describe the symptoms a machine is experiencing, and the companion can suggest tests to validate the issue and suggest possible solutions.

“When a machine breaks, it could be for a lot of different reasons,” he says. “Historically, that’s been solved through documentation.”

Technicians would pore through user manuals and maintenance histories to try to understand the issue. The manuals were often paper or in PDF form, and maintenance history could be logged in operations, where Wickström sits, or in other silos.

“We have connected machinery,” Wickström says. “Sometimes our machinery stops because we have an IT outage, and all of those outages are documented in our IT incident system.”

Wickström notes that operational personnel with more than 20 years of experience might know what the issue with a machine could be straight away, or would at least know where to look. “A newer person who comes onto the shop floor wouldn’t have that depth of knowledge,” he says.

Moreover, Husqvarna’s shops run 24/7. Extremely knowledgeable technicians who spend most of their time building new solutions are available during the day, but not during the night shift, which could lead to extended periods of downtime.

Powered by Microsoft Copilot, the AI Factory Companion has access to data generated by machinery on the factory floor, and a knowledge base of manuals and documentation of maintenance history regardless of the departmental silo it sits in. The companion can also access the knowledge base using retrieval augmented generation (RAG).

And not only can users turn to the companion for help, Wickström and his team use the capabilities to make machinery on the factory floor proactive too. If a piece of equipment issues an alarm that something is wrong, for instance, that alarm can trigger the companion to take the initiative, and query possible diagnoses and solutions that it can then provide proactively to an operator.

Simplifying complexities

Husqvarna has been on a digital transformation journey for years, seeking to improve the efficiency of its operations, so the AI Factory Companion is just one piece of the journey. The challenge for the company, and many manufacturers, is that legacy industrial equipment can be in place for decades, and getting data out of that machinery and integrating it with modern systems is no easy task. The need for low latency on the factory floor means some of the data infrastructure needs to sit locally, at the edge. But it also needs resource capacity of the cloud.

Husqvarna’s answer has been to leverage Microsoft IoT Operations as the central hub for collecting and processing data from its industrial equipment. IoT Operations is meant to help organizations transform their physical operations via Microsoft’s adaptive cloud approach, which unifies siloed teams, distributed sites, and systems into a single operations, security, application, and data model.

“Microsoft’s adaptive cloud approach is about bringing cloud services to places they haven’t been before,” says Bernardo Caldas, CVP of Azure edge product management at Microsoft. “Implementing AI in edge environments like factories can be very difficult, as many of these companies grapple with the complexities of data management in those environments, and the integration with existing systems where the data comes from.”

Traditionally, manufacturing data sits in multiple layers:

• Level 0: physical processes from devices like sensors and actuators
• Level 1: local control stations like programmable logic controls (PLCs) and remote terminal units (RTUs) that manage and control the physical processes
• Level 2: supervisory control and data acquisition (SCADA) systems that monitor and control machines and processes
• Level 3: manufacturing execution systems (MES) that manage the production lifecycle, including production scheduling, material tracking, and quality control
• Level 4: ERP systems that manage the overall organization’s operations

The AI Factory Companion leverages data from the senor level, but the full vision is to transform operations by leveraging data from every level.

Wickström says that part of Husqvarna’s transformation journey has been allowing it to get to a place where it can measure downtime in a meaningful way, but it’s difficult to provide exact metrics regarding how much more efficient the AI Factory Companion is making the company.

“We know the technicians and maintainers are sometimes spending two hours or more to fix stoppages,” says Daniel Johansson, manager of manufacturing digitalization and global operations at Husqvarna. “In some cases, we’ll effect that by a small percentage, but in other cases, we’ll reduce the time by 50 to 60%.”

Work in progress

Wickström explains that the companion is only in its early stages, and he expects its utility to increase dramatically as the knowledge base expands.

“We can’t answer every question yet, but we can address the simple ones, like 20% of the full bucket, and in some cases, we can speed up the time to resolution very quickly,” he says. “If we don’t have the data, we can’t do anything, and it’ll be like that for some time. But now we can improve the common knowledge base instead of just filling in another PDF that no one will look at.”

He adds that the real magic is in good search — finding the documents most relevant to answering a user’s question. That, in turn, requires mastering traditional technologies and combining them with LLMs. For instance, he says, if a user asks the companion, “What was the latest thing that happened to this machine?” that’s actually difficult for gen AI to resolve because “latest” is a complex concept. But a SQL query can easily sort by date and provide the top result.