Analytics-Enabled Quality and Safety Management Methods for High-Stakes Manufacturing Applications

Abstract

Quality management is a critical aspect of the management of manufacturing processes, particularly in industries where product reliability and safety are paramount. With increased digitization and automation, there is growing potential for analytical tools combined with ubiquitous data to aid the transition from quality management practices based on expert intuition and qualitative insights to more data-driven decision making. To assist in bridging the gap between this potential and current implementation practices, this thesis develops new methods for analytics-enabled quality and safety management.

Chapter 2 focuses on the problem of detecting clinically-relevant quality variation in pharmaceutical manufacturing of biologic drugs. Currently, both pre-market clinical trials and post-marketing studies focus on variability in safety outcomes due to individual patient-drug factors. However, the inherent complexity of biologic drug manufacturing and distribution raises potential risks that temporal variability in these systems could also impact clinical outcomes. The chapter describes a data-driven signal detection method using Hidden Markov models designed to monitor for manufacturing lot-dependent changes based on reported clinical outcomes. The method is tested on three lot sequences from a major biologic drug. The results suggest correlated lot-to-lot variability in two of the three, possibly related to changing manufacturing and supply chain conditions that may impact the per lot AE rates.

Chapter 3 explores the problem of creating structured access to unstructured quality data captured in free-text documents. Though operator reports and logs are ubiquitous in many manufacturing processes, one of the main barriers to their effective use in decision making is that unstructured data are often unclassified, which makes trend identification and other actionable analyses challenging. This chapter describes a machine learning and optimization-driven methodology to classify unstructured text in process environments into a known taxonomy of categories without access to an existing labeled training set. To accomplish this, the proposed method leverages information from existing reference documentation and formulates a linear program to select a set of key words that distinguish the categories from each other. Results from three test datasets with ground-truth labels indicate that the method delivers strong classification accuracy, both in absolute terms and relative to alternative methods.

Chapter 4 focuses on a quality test for an optical transceiver module, a high-tech hardware product, manufactured by an industrial partner. Currently, human experts review all test logs for quality problems. This chapter proposes a two-stage machine learning classification model that is able to automatically pass the vast majority of tested products and drastically reduces the need for manual review. Assessment on out of sample real test result data suggests that the two-stage model is able to reduce the manual review burden on the operator by 75-99% while on average satisfying the requirement to limit the number of passed defective modules.