Authors

N. Jia, S. Strelnikov, Acerta Analytics, Kitchener, Canada

Year

2026

Print Info

Production/Publication ÖVK

Summary

Factory Acceptance Testing (FAT) is a critical quality gate in hydrogen fuel cell stack manufacturing, but long test durations constrain throughput, consume costly hydrogen, and delay feedback to production, ultimately making the process inefficient and expensive. To reduce test duration, machine learning models can predict final FAT outcomes from early-stage measurements; however, production deployment is challenging because decision errors carry highly asymmetric costs that are not captured by standard metrics. This paper presents a production-ready machine learning system that can reduce manufacturing costs by enabling early FAT decision-making that preserves strict quality guarantees. The solution combines an end-to-end architecture for data ingestion, model training, edge deployment, and monitoring with a decision framework that explicitly separates prediction from decision-making. Classifiers generate continuous risk scores from shortened FAT data, while operational actions are determined by a cost-based decision layer rather than fixed thresholds. Decision policies are evaluated using realized operational cost and test coverage, making trade-offs between throughput and quality risk explicit. Robust policies are identified via Pareto analysis and stability assessment and distilled into conservative, balanced, and aggressive operating modes. The approach enables clear ownership of risk posture and deterministic execution on the shop floor and is broadly applicable to industrial testing processes with long cycles, partial early information, and asymmetric decision costs.

ISBN

978-3-9504969-5-6

DOI

https://doi.org/10.62626/tjjx-19cx Copy link

Order Members Login