Identifying Customer Returns in a Printed Circuit Board Production Line Using the Mahalanobis Distance

This paper discusses as its primary research question the viability of using the Mahalanobis Distance as a multivariate method for detecting outliers in an industrial setting. An algorithm is used to detect future customer returns in a printed circuit board production line situated in Sibiu, Romania. From the literature, there is a lack of methods, tools and guidelines concerning the paradigm of Zero-Defect Manufacturing. The novelty of the method presented includes separation of highly specialized, future outliers from other outliers, and further automation using Python, a Docker container, a graphical user interface, a search-engine and a reporting tool. This allows the method to be used without external assistance. The data used is extracted industrial datasets from Continentals datalake. The algorithm detects 20% of future outliers and has been implemented by Continental. This can possibly be improved by increasing domain knowledge. The generality of the algorithm in principle allows for use at any of Continental’s production lines. There are strong assumptions regarding the requirements for the method, including benefits of employing domain knowledge critical variable identification and detection rate improvements. Further improvements of detection rate are also discussed. The paper concludes that the algorithm can detect a percentage of highly specialized outliers with simple automation in Python, but also acknowledges limitations in terms of increased demands from data quality and domain knowledge.