Data-driven throughput bottleneck analysis in production systems

Doctoral thesis, 2021

Production systems and production management are getting smart. Manufacturing companies are increasingly adopting digital solutions to monitor and manage production systems. By adopting digital solutions, it has become possible for manufacturing companies to collect huge volumes of widely varying production system data. Alongside this, significant advances in recent years in machine learning and artificial intelligence fields have opened up opportunities to develop data-driven approaches for analyzing the huge emerging volumes of data, deriving insights, and using those insights to improve shop floor productivity.

One way to increase shop floor productivity is to strive to achieve an even flow within production systems. However, even system flows may be disturbed by various factors such as random breakdowns, minor stops, setups, variations in cycle time, waiting for an operator, and so on. Such disturbances may constrain production system throughput. However, previous research has shown that not all disturbances in different machines in the production system
constrain the production system throughput. The disturbances are significant in a set of machines in the production system that constrains the system throughput. This set of machines is called throughput bottlenecks. Quick and correct identification of throughput bottlenecks will help practitioners plan appropriate eliminating actions.

Existing academic research efforts to investigate throughput bottlenecks have largely adopted analytical approaches or discrete-event simulation-model-based ones. However, now that companies are collecting huge volumes of digital data, this data can be analyzed directly by developing data-driven approaches. This allows the derivation of insights into throughput bottlenecks in production systems.

This doctoral thesis constructs a series of data-driven approaches to analyze throughput bottlenecks. Firstly, a data-driven approach is proposed to identify historical throughput bottlenecks in a production system. But merely identifying bottlenecks is not enough if informed actions are to be taken. Secondly, a data-driven approach is proposed for diagnosing historical throughput bottlenecks. Specifically, a diagnosis is made based on a maintenance perspective. Combined with identification and diagnosis, practitioners may then plan and execute different corrective actions. However, when such corrective actions are applied, the dynamics of the production system change. That means that bottlenecks will not act as bottlenecks in the future. Thirdly, to thus predict how the system dynamics will change (and thereby to predict future throughput bottlenecks), a data-driven approach is proposed to predict future throughput bottlenecks in a production system. Fourthly, to help practitioners plan for proactive actions on the predicted throughput bottlenecks, a further data-driven approach is proposed; one which prescribes actions to combat the bottlenecks.

The different data-driven approaches proposed in this thesis have been tested using production system data sets extracted from different real-world production systems. The insights obtained from applying these approaches may help practitioners to better understand the dynamics of throughput bottlenecks and plan for specific actions to eliminate them. Such elimination helps achieve an even flow in production systems, thus increasing productivity.