Stellantis Goiana: Welding AI Earns Engineers' Trust
At Stellantis' Goiana plant, an anomaly detection system prevented 14 equipment failures in six months. For automation professionals, the most interesting part is how the team built engineers' trust: by showing the evidence.
On November 11, 2026, IIoT World published an article on a concrete case of AI on the factory floor: the anomaly detection system for welding equipment at Stellantis' Goiana plant in Pernambuco state, Brazil. The piece draws on an Industrial AI Summit 2026 session titled "From Months to Days: Accelerate Industrial AI with a Modern Data Architecture", featuring Aron Semle of HighByte and Denis Leite of Mekatronik, moderated by Hamish Mackenzie of IIoT World.
A necessary caveat: the article is labeled as sponsored content from HighByte, so the story is backed by a vendor. The numbers should be read with that in mind, but the case still offers useful lessons.
The reported results
The system flags likely welding equipment failures before they interrupt production. In the first six months, the reported results are:
- 14 equipment failures avoided;
- 15 hours of production recovered in a single month;
- an expected annual capacity gain of 810 vehicles.
The last figure is described as "expected": it is a projection, not a measured result. The team says results have improved further since these figures were captured, but it provides no updated numbers.
Contextualized data and a production-first design
Welding data came from multiple sources: TXT files, SQL databases and operational systems. Before sending it to the data lake, the team organized and contextualized it around individual welding operations rather than dumping raw data. This step often makes the difference between a model that can be trained and one that learns noise.
The team also skipped the pilot phase, planning for reliability, maintenance and scalability from the start. Stellantis made its engineers' time available, so the project could draw directly on knowledge of the production process.
Trust comes from evidence
The most instructive part concerns adoption. Engineers used dashboards, time series and distribution charts to compare normal welds with the anomalies flagged by the model. The first visualizations did not clearly separate the two behaviors, and the engineers saw no reason to intervene.
After the visualizations were refined, the engineers recognized that the flagged cases deserved investigation. Each alert confirmed as a real problem increased their confidence: according to the article, the model went from being questioned on every alert to being expected to be right.
Trust did not come from claims of accuracy, but from evidence engineers could verify.
Key takeaways
For those working in automation and industrial software, the case suggests a few practical points:
- design how engineers will validate the system, not just how to build the model;
- treat dashboard and HMI work as a core part of the project, because an unclear visualization can make people ignore a correct alarm;
- contextualize data by operation or event before moving it to a central platform;
- involve the people who know the process from the very beginning;
- think about production from day one, to avoid pilots that never get beyond experimentation.
Limits and next steps
The article does not describe the model type, sensors or architecture in detail, so the case cannot be replicated "by recipe". Stellantis is considering extending the approach to other plants and applications, but does not say where it will start. It remains a useful example: in an industrial environment, a model is adopted when its users can check its alerts with their own tools and their own experience.