05/06/2025

Pi Center develops predictive maintenance project for Stone.

The IMPA Center for Projects and Innovation (Centro Pi) has completed the predictive maintenance project developed for Stone, a Brazilian fintech company specializing in financial solutions. Initiated in the second half of 2024, the project aimed to predict potential battery failures in the company's card payment terminals. As a result, the team was able to generate models with a good predictive rate for device battery life in various contexts.

PI Center project scientist Lucas Nissenbaum explained that the “project sought to determine with the greatest possible precision when a card machine would have battery problems that would lead to irregular operation.” To solve the problem proposed by the fintech company, several models were built, based on machine learning , time series processing, and data analysis, to generate a solution that had a good correlation with battery health.

The Pi Center used varying time windows for the project's development, ideally with a forecast weeks in advance. The idea was to determine the occurrence of two events: episodes of extremely low battery life in card machines and cases where the devices malfunctioned. Mathematically, the solution was developed using machine learning and data processing.

Henrique Madeira, leader of the Logistics Business Intelligence team at Stone, worked directly on the project as a strategic interface between Stone's technical team and the Pi Center team. According to him, “the collaboration was highly productive: we were able to translate operational data into variables with strong predictive power, allowing the company to adopt a more proactive approach to failure management. This represents an important advance in the maintenance of our hardware infrastructure, contributing to reduced replacement costs and, above all, improving the experience of our main reason for existing: our customers. This project marks another significant step in consolidating our strategy of using Artificial Intelligence applied to the business.”

In addition to Nissenbaum, the team includes project scientist Francisco Ganacim, four doctoral students, one master's student, and one postdoctoral researcher . Vanessa Fernandes is a doctoral student in Computer Science at UERJ (Rio de Janeiro State University) and was one of the collaborators from the Pi Center on this project. She highlights the relevance of working with feature engineering and neural networks.

“It was one of the best experiences I've ever had. I loved being part of a real team, on a real project, for a large company, and all of this within IMPA, which is one of the leading research institutions in Brazil. I felt incredibly privileged. I think I grew on several fronts: personally, interpersonally, computationally, and mathematically. It was an opportunity to deepen my reasoning on various aspects of the model, better understand the evaluation process, and even reinforce concepts that we sometimes think we already know. I also learned a lot from interacting with the team.”

Melvin Poveda, a doctoral student at IMPA, highlights the final part of the project: the extraction of features (machine characteristics) to explain battery health and train machine learning models . “I feel I was able to refine mathematical knowledge I learned in various courses at IMPA. Especially the machine learning class , as well as algorithms, Information Theory, and other research areas. I also enjoyed and learned a lot from interacting with Stone. They were always available for any questions and helped us a lot to create a better solution to the problem together,” he said.