Application of multivariate data analysis to improve and optimise industrial processes

Abstract

ABB, who is the sponsoring company for this research work, is a global leader in power and automation technologies based in St. Neots, Cambridgeshire. The thesis discusses the work carried out on a portfolio of projects as a part of the Engineering Doctorate programme. Application of multivariate statistical process control was central to the successful implementation of the projects. The first project focussed on a Process Analytical Technology (PAT) software solution developed by ABB. The US Food and Drug Administration (FDA) have defined PAT as a process for designing, analysing and controlling manufacturing through timely measurements of Critical Quality Attributes (CQAs) of raw and in-process materials in order to achieve final product quality. The project’s overall objective was to enable seamless roll out and maintenance of chemometric models for at-line testing across multiple worldwide locations. The work presented in the thesis discusses a solution that allows global maintenance of at-line analyser measurement stations whilst providing ‘real time’ quality data at the right business level to enable more efficient business decisions. This required optimising the software during the preliminary stages which included developing hierarchical Partial Least Square (PLS) Models, maintaining a process within control and exporting data using the Model Data Exporter plug-in. Likewise the project involved development of a combination of test sets that could assess and improve the robustness of the product. Following the Factory Acceptance Test (FAT) and Site Acceptance Test the product was successfully commissioned at customer site. The second project investigated a recurring uncharacteristic event in the polymerisation process. This unusual phenomenon led to downgrading of the batch further causing a loss of revenue. Previous investigations indicated that the most likely reason for this unusual behaviour was due to the occurrence of crystallisation in the polymerisation reactor. These batches were identified by monitoring a ‘kink’ in the heat up profile during the polymerisation process. The root cause of this crystallisation was initially examined by monitoring the rate of reaction and analysing the behaviour of one variable at a time. However, these approaches were unsuccessful to identify the underlying issue with the crystallised batches. This body of work illustrates a series of steps developed using multivariate analysis techniques to identify unusual batches in the polymer reactor. Exploratory data analysis using Principal Component Analysis (PCA) and Multi-way Principal Component Analysis (MPCA) was performed on the historic batch data (quality, process and Overall Equipment Effectiveness (OEE)) to identify ii the root cause of the problem and develop a well defined method that can be used by the operators to identify abnormal batches.