Modelling and Control of Production Systems based on Observed Inter-event Times: An Analytical and Empirical Investigation (Ph.D. Thesis)

TL;DR

This thesis investigates how observed inter-event times can be used for performance evaluation and control in manufacturing systems, emphasizing the importance of accounting for autocorrelation and applying machine learning for cycle time prediction.

Contribution

It introduces analytical models for control policies considering autocorrelation and demonstrates empirical data analysis and machine learning applications in manufacturing performance prediction.

Findings

Ignoring autocorrelation leads to incorrect threshold estimation.

Optimal control policies are state-dependent threshold policies.

Machine learning methods can accurately predict cycle times.

Abstract

Technological advances allow manufacturers to collect and access data from a production system effectively. The objective of data collection is to deploy the collected data in developing decision support systems for performance evaluation, problem identification, and production control. The goal of this dissertation is to investigate how the collected data can be used to evaluate performance and optimize manufacturing systems, analytically and empirically. In the first part of the thesis, I investigate how can the collected data from the shop-floor be used in the efficient control and design of manufacturing systems? To investigate the impact of possible dependency in the inter-event times on the optimal control and performance measures of the system, a manufacturing system that is controlled by using a single-threshold production control policy is analyzed. It is shown that ignoring autocorrelation in inter-arrival or service times leads to overestimation (underestimation) of the optimal threshold level for negatively (positively) correlated processes. Then, it is shown that the optimal control policy of a production/inventory system with correlated inter-arrival and service times modeled as Markovian Arrival Processes is a state-dependent threshold policy. In the second part of the thesis, an exploratory data analysis is conducted by using a large industrial dataset related to the flow of the products at a semiconductor wafer fabrication. Then, the cycle times of the products are predicted using Machine Learning methods, and the performance of different prediction algorithms is assessed. The analytical and empirical results presented in this dissertation show that effective use of collected data from a manufacturing system enables controlling the system effectively and predicting its main performance measures accurately.