Data-Driven Bottleneck Identification for Serial Production Lines

Abstract

Bottleneck identification is an important challenge for manufacturing firms to address in pursuit of increased manufacturing capacity. Prominent efforts used simulation-based approaches but are seldom extended to industry due to lack of data availability. This thesis presents an application of three methods to an industrial data set. A data pipeline was developed to automate data collection from machinery, which was used to realize the first data-driven application of bottleneck identification methods in the electronics industry. It was determined that three methods agreed on the identified bottleneck machine for just 8% observed production runs. Once different perspectives on bottleneck machines were explored, diagnostic insights were sought to identify root causes of capacity loss. Fault events collected by the data pipeline were analyzed using first principles and unsupervised learning techniques. Four data visualization approaches were explored, and a web interface was developed to communicate these insights. The principal causes of unplanned downtime were identified and aligned with domain knowledge at the case company. Improved data quality will improve the quality of bottleneck identification results, and further data collection will improve the understanding of fault behaviour. Ultimately, the work reveals the opportunities and challenges associated with data-driven applications in manufacturing.