Identification and Measurement of Variables Affecting Component Lifetime: A Case Study of Polymer Degradation in the Biopharmaceutical Industry

Abstract

Modern maintenance methods are dependent upon understanding and monitoring component degradation; however, many industrial situations exist where monitoring is impossible, and component degradation mechanisms are not well understood. This research explored the hypothesis that data to support maintenance operations may be extracted within industrial situations where pre-existing data on component degradation is scarce and understanding of the physics of the degradation process is low. In order to investigate this hypothesis, a case-study in the biopharmaceutical industry was followed which attempted to discover the significant variables leading to component degradation. In such industrial scenarios, data fusion techniques, where multiple data sources are pooled together, are necessary to adequately model complex manufacturing systems. In addition, the interaction between humans and the system is an important consideration that should be taken into account. In order to develop models that will predict component lifetime, a beneficial modelling framework is: 1) data inspection, 2) selection of appropriate historical data, and 3) data pre-processing. However a knowledge gap remains to guide the elicitation of the appropriate data required for such models in many real world industrial scenarios. To help bridge this gap a methodology is proposed here to aid in significant data identification and measurement in complex industrial environments where data is both scarce and siloed. The case study followed in this work was that of valve diaphragms used in the biopharmaceutical industry, manufactured from the polymer Ethylene Propylene Diene Monomer (EPDM). The central question in this work was therefore, what data, both machine and human related, are useful to predict the lifetime of EPDM diaphragms? In order to answer this question, data from the process level, the system level, the equipment level, and the component level were required. These disparate data sources, including data from both diaphragm supplier and end-user, each contain valuable information which needs to be extracted. The significance of that data can then be determined, and if significant, this information can be then used within appropriate models. The first step in this work was the modelling of the degradation of the component as a stepwise process, known as a Markov chain. The stepwise nature of the Markov-chain approach allowed the use of bespoke statistical techniques to determine the significant degradation data, a significant step given the scarcity of data. To enable the multi-state approach, a qualitative assessment method was created which categorised the components into different health states for the first time. Once the initial list of significant data had been filtered using the developed method, four key chemical interactions were investigated further using multiple material characterisation techniques. This acted as an additional aid to the selection of the set of significant process variables contributing to component degradation. It was necessary to perform these analyses as the statistical techniques employed could not identify, with a sufficient level of confidence, which data should be taken into account when developing models of component lifetime. The material characterisation afforded, for the first time, an understanding of the chemical and physical mechanisms responsible for EPDM degradation that occur during exposure to the chemicals commonly used in biopharmaceutical production. An understanding of the physics of failure of the components also acts as a first step in potentially developing condition based monitoring solutions for this application. To fully encapsulate all possible root causes of component degradation, it was critical to assess if human interaction during maintenance activities played a role in the sudden failure of the components in-situ. To accomplish this, a methodology was developed enabling the integration of human factors as quantitative data within component degradation models, utilising human knowledge of the maintenance process. The case study investigation suggested that incorrect installation has a significant impact on component lifetime, and that correct procedures were not always being following in the valve installation process. Modelling the factors which influence technician performance in this way allows individual components to have their own local hazard rates, a first for human-system interaction effects due to maintenance intervention. In conclusion, the central hypothesis of this work was substantiated, as useable quantitative data was extracted from scarce and siloed data sources via a multi-disciplinary approach which unified several sources of information. The gathered data would enable the development of predictive lifetime models. It is important to note that the approach of this work is not intended to provide an exact data elicitation solution for all industrial scenarios, as this will change depending on a multitude of factors related to the case study in question, such as data availability, required resources, time constraints etc. Rather, this work should be interpreted as a template to guide the identification and quantification of appropriate data for inclusion in reliability models which can be adapted as required for different systems.