In this edition of the MARS Results Series, we speak with Pilar Herrera Plaza, Head of Innovation and European Artificial Intelligence Projects at AIN, the Navarra Industrial Association, about the development of advanced artificial intelligence algorithms for failure detection and quality prediction.
Ensuring consistent product quality across different machines and manufacturing sites can require extensive inspection and often means that deviations are only detected after production. Developed by MARS partner AIN, the Deep Learning models analyse signals collected directly from manufacturing processes to predict quality indicators such as surface roughness and dimensional accuracy. This can support earlier detection of quality deviations, reduce inspection effort and provide manufacturers with a more efficient basis for monitoring production.
Pilar Herrera Plaza works at AIN in Cordovilla, Navarra, where she leads innovation and European projects in the field of artificial intelligence. Within MARS, AIN contributes expertise in artificial intelligence modelling and develops transferable Deep Learning approaches that can adapt to different manufacturing sites, machines, sensors and operating conditions.
We asked Pilar Herrera Plaza how the models work, what makes them transferable across manufacturing environments, what has already been validated and how they could support near real-time quality prediction in the future.
INTERVIEW with Pilar Herrera Plaza
In simple terms, what is this MARS result, and what real-world problem does it address?
Pilar Herrera Plaza: This MARS result consists of Deep Learning models that predict product quality directly from manufacturing process-monitoring signals. The models estimate quality indicators such as surface roughness and dimensional accuracy without requiring extensive manual inspection. By analyzing sensor data collected during production, they enable earlier detection of quality deviations and support more efficient and reliable manufacturing operation.
What makes this result different from what existed before (and why does that difference matter)?
Pilar Herrera Plaza: What makes this result different is its ability to develop and adapt AI quality prediction models across different manufacturing sites, processes, sensors, and operating conditions. Unlike conventional Machine Learning approaches, which rely heavily on manual feature engineering, the Deep Learning models can learn relevant patterns directly from process-monitoring data. Domain adaptation techniques also help address differences between training and real manufacturing conditions. This improves the potential for robust quality assessment across distributed manufacturing environments.
How do you know it works – what has been tested, demonstrated, or validated so far?
Pilar Herrera Plaza: The models were trained and validated using industrial datasets generated through Design of Experiments (DoE) and real manufactured parts across multiple manufacturing sites. Testing showed strong performance for quality classification, particularly for surface roughness prediction, with successful validation under real operating conditions. These results demonstrate the feasibility of AI-driven quality assessment using manufacturing process signals.
How does this result fit into the bigger picture of the MARS Platform – what role does it play?
Pilar Herrera Plaza: The Deep Learning models provide the core intelligence for quality assessment within the MARS platform. They transform process-monitoring data into quality predictions that can support manufacturing decision-making and process monitoring. In addition, the models constitute the basis for future integration into the MARS federated learning framework, enabling distributed AI across manufacturing sites.
What happens next – what are the next steps and what could we realistically see after the project ends?
Pilar Herrera Plaza: The next step is the deployment and integration of the Deep Learning models into manufacturing environments and federated learning infrastructures. After the project, these models could be used to provide near real-time quality prediction, reducing inspection effort and supporting more consistent production quality across different manufacturing sites.