Stellantis Goiana: AI on Welding Lines Wins Over Operators with Evidence
According to an IIoT World article dated November 11, 2026, anomaly detection on the welding equipment at Goiana helped avoid 14 failures in six months. The key point is how the plant earned technicians' trust: by showing them the data.
On November 11, 2026, IIoT World published an article based on an Industrial AI Summit 2026 session titled "From Months to Days: Accelerate Industrial AI with a Modern Data Architecture." It describes a concrete case of AI in production: the anomaly detection system on the welding equipment at the Stellantis plant in Goiana, in Pernambuco, Brazil. The speakers were Aron Semle of HighByte and Denis Leite of Mekatronik.
A note on sources: the article is labeled as sponsored by HighByte, so the results come from the vendor and its partner and have not been independently verified. They should be read with that caution, but the lessons in method remain interesting.
The Reported Results
The system helps maintenance teams spot potential welding equipment failures before they interrupt production. Over six months, according to the speakers, it:
- helped prevent 14 equipment failures;
- recovered 15 hours of production in a single month;
- supports an expected annual capacity increase of 810 vehicles.
The team says results improved further after these data were collected. Extending the approach to other plants and applications is under consideration, but where to start has not yet been decided.
Contextualized Data Before the Model
The welding data came from TXT files, SQL databases and operational systems. Instead of sending raw data to the data lake, the team first organized and contextualized it around individual welding operations. This is an architectural choice: an anomaly detection model only makes sense if every measurement is tied to the process step it belongs to.
Production from Day One
The team skipped the pilot phase and planned for reliability, maintenance and scalability from the start. Stellantis also set aside the time needed for its own engineers to take part in development, bringing their direct knowledge of production into the project.
Operator Trust
The most instructive part concerns adoption. To let engineers compare normal welds with the anomalies flagged by the model, the team used dashboards, time series and distribution charts. The first visualizations, however, did not clearly separate normal from anomalous behavior: engineers saw insufficient reason to intervene and questioned the model at every alert.
After the visualizations were refined, engineers recognized that the alerts warranted investigation.
As technicians confirmed real problems, trust grew, and the model became a tool the plant relies on in 24/7 operations.
Key Takeaways
For engineering and IT leaders, the case suggests three practical lessons:
- Contextualize data by process step before training models, not after.
- Design for production from the start, considering reliability, maintenance and scalability, and involving plant engineers.
- Treat explainability and evidence as part of the deliverable. A prediction creates value only if the people doing maintenance decide to act on it.
In short, model accuracy alone is not enough: if an alert does not come with readable evidence, it gets ignored. Setting aside the numbers, which remain vendor claims, the Goiana experience shows that visualizing normal versus anomalous behavior is a project requirement on par with data architecture.