Resource Consumption Models for Different Manufacturing System Configurations

This thesis will focus on resource consumption models of productivity, cost of quality, and cycle time, to describe and select preferred configurations as a function of the system and operating parameters. Consider a production system which consists of workstations or machines that can be arranged in different configurations, where each configuration can affect the performance of the system. This thesis will model and analyze four main configurations, namely serial, parallel, and serial-parallel with and without crossover. In a production system, parts are processed at each workstation or machine, where a value is added to the part. Inspection is performed to classify a quality characteristic of a part at each workstation or machine to be accepted, reworked, or scrapped. The short term probability of accept, rework, and scrap are utilized to model the long term probabilities using an absorption Markov Chain methodology. Each configuration will result in unique long term probabilities depending on the number of processes, order and location of each process. The long term probabilities and process flow are used to develop the resource consumption models for each alternative. At first, a two process production system is analyzed using pure serial and parallel configurations, and their performance is evaluated using sensitivity analysis. Then, a four process production system is analyzed using the four configurations, and the performance of each alternative is evaluated using a numerical example. Finally, general models for the resource consumption of an n-process production system are developed for the serial , parallel , and serial-parallel without crossover. Mathematica® is utilized to develop the matrix calculations and equations for the n-process models. A case study of a biopharmaceutical company is used to apply the proposed models. All required data for three production systems are collected then analyzed to be used in identifying the productivity, quality cost, and cycle time for different configurations. It is shown that there is a relationship between system configuration and its performance measured by productivity, quality cost, and cycle time. The proposed methodology in this thesis can be used to select the preferred configuration, where production systems with different parameters can result in different conclusions. The selection of the best configuration can be done by evaluating the resource consumption models for each possible alternative considering the different operating constrains, where the models in this thesis allow the manufacturer to select the best alternative based on specific performance targets. The developed models can also be used in discrete and continuous manufacturing systems, such as the biopharmaceutical industry case study. Acknowledgments I would like to thank my advisor Dr. Valerie Maier-Speredelozzi for her help and support. I also would like to thank all who supported this research from XY Biopharmaceutical Company by providing me with all the tools and access to their production plant and data.


186
x LIST OF TABLES

INTRODUCTION
Many studies have focused on efficiently designing the production system which has fluctuations in demand in an ever-changing global market. These studies also focused on the continuous need for improving the production system to meet c1:1stomer expectations. Many models have been developed to improve the overall performance of the production system . These models have been concerned with production planning , facility layout, line balancing, and material handling, to name a few. The question then becomes, does the system configuration of equipment play a significant role in the overall system performance? If yes, then what are the performance measures?
The ability to cope with continuous changes in demand has become one of the fundamental challenges that modern manufacturers currently face. Recently there are some research projects that focus on diagnosing the relationship between the system configuration or layout and the performance. Many metrics have been used to evaluate the performance of each configuration .
This chapter will include a brief discussion about system configuration, the problem statement, the justification and significance of the study, a discussion of 1 scrap and rework in production systems , a review of the biopharmaceutical industry, and a description of the thesis organization. Koren et al. (1999) defined the system configuration as the machine or workstation arrangement and connections. The machine arrangement means how the processes flow through the line, but not necessarily the physical layout of the machines. The serial layout is the traditional system configuration by which the material flows from one stage to another where each stage adds value to final product, there is only one path for the product flow, and the line is dedicated for one product or product family . This configuration was traditionally used in the automotive industry where the high volume and low product mix were required.

System Configuration and Challenges
Another configuration is the pure parallel by which each workstation is arranged parallel to the other and each one can perform all required tasks to complete the final product. In this configuration machines should be flexible and multi-task which means a high investment cost. Other than these two extreme configurations (pure serial and pure parallel) , there are the parallel-serial and hybrid configurations. The parallel-serial includes two or more identical serial lines arranged in parallel , with or without crossover. Without crossover, the number of production flows equals the number of serial lines. The number of paths is larger when crossover is included, because of the flexibility in transferring products between lines. In contrast to the serial line, one benefit of the parallel-serial 2 configuration is the ability to keep the system producing products even when one or more machines fail. It also facilitates the ability of the system to not be completely idle during the conversion from one product to another. In hybrid configurations, the number of identical machines at each stage is not equal as it is in the parallel-serial configuration, and crossover does not necessarily exist after each stage.
A number of performance metrics have been proposed in the literature to assess the performance of each of the above described configurations. Most of the studies in this area considered the following performance measures: productivity and reliability, quality, convertibility, number of product types, scalability, and cost. How the different types of configuration impact these measures has been the core concern of many studies. Diverse models have been pursued in analyzing the effect of configuration on performance measures separately, where an integrated model that combines all measures might be one option in finding the preferred system configuration that will satisfy all performance measures.

Problem Statement
There should be useful methods and principles for designing the manufacturing system and facilitating the selection among configuration alternatives. The effect of scrap and rework should be considered in modeling production systems, since it will reflect more accurate calculations for resource 3 utilization . This could generate a strong baseline for selecting the preferred configuration.
Biopharmaceutical companies have been challenged to reduce their production costs to be able to deliver their products to the market at a reasonable price, and to remain competitive, maintaining the quality level that is required to provide customers with reliable products . Recognizing the effect of system configuration on performance, in addition to understanding biopharmaceutical processes, regulations, and limitations, will contribute significantly to improving the manufactur)ng system and decreasing high production costs that most biopharmaceutical companies currently face .
Different manufacturing system configurations result in different performance which is measured by many metrics. When the demand and product type or lifetime change, the efficiency of the production line deteriorates, unless the equipment layout is flexible and optimized to compensate for that change.
Equipment configuration should be analyzed to improve the manufacturing line efficiency. Configuration selection requires proposing potential configurations, analyzing the model of each configuration and then identifying the preferred one.
Not all materials which enter the production line make it to the end as finished products, because there is material that leaves the line as scrap, which could be generated from processes or rework stations after parts fail to meet specifications . Rework and scrap rates strongly affect the production cost because 4 they increase labor, material, quality and other costs, and hence reducing the rework and scrap rates improves the line efficiency and profitability.
Consider a production line with many equipment workstations where there is a need to improve the system output. Capturing how the configuration selection contributes to achieving this target is a large challenge . Different configurations such as: serial , parallel , serial-parallel with and without crossover, and hybrid configurations could be proposed and analyzed. The way of analyzing these alternatives and then finding the best one is another challenge that exists. In this thesis, the production flow through all stages is modeled , which includes: raw material, main processes, rework stations, scrap bins, and finished product. The probability of transition between stages is considered in building the model, where these probabilities are combined with the resource utilization at each stage. A Markov chain is a stochastic model that is applicable in this instance. After creating the model and depending on the type of resources that need to be considered, the model is analyzed to find system resource utilization for each configuration alternative and then the configuration that provides the best performance can be identified.  Zhong et al. (2000) built models used to measure productivity and quality of different system configurations. They emphasized the importance of the convertibility metric in measuring the system performance, especially for RMS, and pointed out that finding an integrated model that reflects the impact of configuration on all performance measures is very important in selecting the right configuration . Spicer et al. (2002) illustrated many principles that help in designing the system configuration. They compared the effect of different configurations on throughput, 6 line balancing, and scalability. Freiheit et al. (2003) proposed a model to determine the productivity for mixed serial-parallel configurations, and showed how this non-traditional configuration is equivalent, in the ability of improving productivity, to the buffered serial configuration. Colledani et al. (2005) provided a decomposition method for evaluating the performance of configuration or reconfiguration of the production system. Shabaka et al. (2007) mentioned two levels of reconfiguration : system and machine. They explained how the ability of re-assigning the resources at these two levels leads to increases in the flexibility of the production system which becomes more able to cope with changes in product demands. The authors demonstrated how different configurations of machine axes of motion cause different capabilities for the machine. Bohn and Terwiesch (1999) mentioned that in most production systems not all raw materials entering the line to be processed make it to the end and turn into a good quality final product. They explained that at the checking point components are tested and classified either as good or as defective items. They explained the high impact of scrap and rework on the system capacity utilization especially if defects are detected at bottleneck machines. The authors studied the effect of yield losses, caused by reworked and scrapped items, in yield-driven production processes. They demonstrated the importance of improving the yield in increasing the productivity. They summarized the rework and scrap effects on material, labor, capacity, and variability related costs. Bowling et al. (2004) presented a model to determine the expected profit from n-stage serial production systems by considering the cost of processing, scrap, and rework . The authors used absorbing Markov chains to identify the long term probabilities of the three states of the product during the production processes: rework, scrap, and accepted. Pillai et al. (2008) showed that Markov chain models can be used to represent the production system under uncertainties due to reworking and scrapping. The authors modeled material flow through a serial production system with rework and scrap by using the absorbing Markov chain. By using the model they adopted, system design and also production and inventory control could be efficiently improved. Using an absorbing Markov chain model enables the authors to build equations to determine the following: material requirements, number of machines, production cost, and manufacturing lead time.
As can be seen from the above, many studies focused on building models to study the impact of system configuration on productivity, quality, scalability, and convertibility, but they do not typically consider resource consumption which is strongly related to the production cost or profit. On the other hand, some studies focused on the effect of scrap and rework on the production system performance, but they only focused on the serial configuration and did not consider other types of system configurations. This thesis will use the Markovian approach to model the production flow of four different system configurations: serial, parallel, serialparallel without crossover, and serial-parallel with crossover. Scrap and rework 8 probabilities will be considered , and different system resources such as money and time will be built into the model. After that, the proposed model will be used to evaluate many configuration alternatives and select the best one. Scrapping and reworking make deterministic modeling unreasonable, so instead stochastic modeling is required which could be achieved by using Markov chains, as will be adopted in this thesis. Moreover, although most of the studies of the effect of configuration on system performance have been applied in machinery or similar production systems, in this thesis a case study will be applied at a biopharmaceutical company to validate the model. All of the above make this research unique with significant contributions to the literature.

Scrap and Rework States in Production Systems
The production system consists of stages where items are being processed, inspected, and forwarded to consecutive stages, such as reworked, scrapped, or finished products. Scrap and Rework are two states in the production system that affect the quality, cost, and cycle time, and if their probabilities are known, then the production system characteristics will be better understood which in turn will improve the process of achieving the production target.
In production , parts could conform to quality specifications or be defective.
Parts that conform to the quality specifications continue through subsequent production processes , where the defective parts, which are items that do not confirm to predetermined specification limits and attribute requirements, could be 9 then either reworked or scrapped . Parts that could not be repaired by the rework process have to be scrapped, but this is not the only criteria to classify items as scrapped or reworked; sometimes the items are scrapped because it is more economically feasible than repair. In specific types of production , such as machining , items are scrapped or reworked if the quality characteristic fall outside the specification limits.
Raw materials, unfinished products, rework, scrap or finished goods represent all possible states that material could transfer through when raw material enters the production line. Building a stochastic model that combines all of these states will be one of the challenges in this thesis.

Biopharmaceutical Industry
A Biopharmaceutical can be identified as: "A therapeutic product created through the genetic manipulation of living things, including (but not limited to) proteins and monoclonal antibodies, peptides, and other molecules that are not chemically synthesized, along with gene therapies, cell therapies and engineered tissues" (Odum, 2002) . Newcombe et al. (2008) mentioned that "more than a third of all drugs under development by pharmaceutical and biotechnology companies are biopharmaceuticals". They said that in 2005 there were nearly 300 licensed biopharmaceuticals on the worldwide market and the number is growing annually 10 by 20%. They also said that the revenue for these biopharmaceuticals is approximately $85 billion.
Biopharmaceutical companies have been challenged to reduce their production costs to be able to deliver their products to the market at a reasonable price, and to remain competitive, maintaining the quality level that is required to provide customers with reliable products. There are many regulations that biopharmaceutical production processes have to meet before products exist in the market, which also increase the challenges. In addition, there are limitations that affect the ability of the companies to upgrade their production systems. For example, when a company changes from producing one product to producing multiple products it increases the risk of cross contamination which makes it more difficult to ensure the quality and safety of the products. Improving production and reducing operational costs in both continuous and discrete production environments is essential for the pharmaceutical and biopharmaceutical industries.
Recognizing the effect of machine configuration on system performance, in addition to gaining a better understanding of the biopharmaceutical processes, regulations, and limitations will contribute significantly in improving the manufacturing system and decreasing high production costs that most biopharmaceutical companies currently face .

.7 Thesis Organization
This thesis is organized as follows : the introduction has been presented in Chapter 1. The literature review is addressed in Chapter 2. The research theory and development of models for a multi-stage production system with two, four, and n processes is provided in Chapters 3, 4, and 5 respectively. After that, a case study at a biopharmaceutical company is applied in chapter 6. The discussion and conclusions are included in chapter 7.

LITERATURE REVIEW
As the ever-changing global market and the growth of competitiveness increases the pressure that modern manufacturing systems face, the production layout problem has attracted the interest of many scholars, with new models developed to address the complexity of today's production system design. Coping with the continuous demand changes has been considered as a fundamental challenge in designing the production layout. While much of the research in the field of manufacturing systems has been concerned with physical machine arrangement, others focus on the way that products flow through the workstations.
Finding the relationship between system configuration and system performance is relatively new in literature. Some scholars have developed models for analyzing the system performance that is measured by one or more me tries such as: productivity, quality, scalability, convertibility, or cost. They have used the proposed models to analyze different configuration alternatives and find the best one. In this thesis a new model is developed to assess the performance for different configurations by also considering the scrap and rework rates which are in turn associated with resource utilization. Many scholars have studied and analyzed the effect of scrap and rework and used them in modeling the production 13 system, but they have focused on serial configurations more than on other types of configurations. Moreover, many research projects for improving production layout have included case studies that have been applied in different types of production systems like automotive and machining , but not many studies have been implemented in biotech or pharmaceutical industries. However, some studies using Industrial Engineering principles and tools have been published in these industries.
This chapter presents som e literature that focuses on several relevant topics, which are: 1. Changing demands and needs for non-traditional system configuration .
2. Layout design problems and challenges.
3. The relationship between different configurations and system performance.
4. Scrap and rework effects and considering them in modeling the production system.
5. Challenges in the biopharmaceutical and pharmaceutical industries.
6. Studies that apply industrial engineering concepts or tools in system design at biopharmaceutical and pharmaceutical industries.

Changing Demands and Needs for Non-Traditional System Configuration
Continuous changes in customer demands and short product life cycles mean that manufacturers need to be more flexible when producing a large range of products. The challenge is to construct system configurations that have the ability to rapidly react to these changes at the lowest cost. Recently, research in 14 this area has been increasing and generally focuses on proposing different configurations other than the traditional serial line. Kochhar and Heragu (1999) proposed a methodology for layout design in a dynamic facility where the product mix and volume are not constant. The authors demonstrated that there is a continuous change in the product range and volume, and getting an accurate forecast becomes not possible any more. Based on that, they emphasized the importance of designing a dynamic layout that has the ability to be highly responsive to continuous changes. Suh et al. (1998) demonstrated that there is a strong relationship between manufacturing system design and productivity. They illustrated that the different functional requirements and constraints which need to be satisfied generate different manufacturing system designs. They introduced the principles and methods for the axiomatic design of manufacturing systems which is based on the chosen functional requirements. The authors introduced the challenges that industry faces to produce a mix of products, and asserted the importance of using scientific principles and approaches rather than empirical knowledge when designing the manufacturing systems. The authors explained that constraints related to the available technology, cost, and performance metrics sometimes may affect the functional requirements which need to be considered in the design which, in turn, leads to weakness in the manufacturing system design. Koren et al. (1998) explained that the traditional serial line configuration is not useful in "low volume -high variety" production systems. Freiheit et al. (2003) mentioned that increasing buffer size to improve the system capacity is not beneficial anymore due to the continuous changes in demand which requires efficient reaction to market changes by reducing work in process, detecting quality problems quickly, and avoiding the risk of overproduction. In terms of productivity, Freiheit et al. proposed a model to analyze how the nontraditional configurations, such as mixed serial and parallel configurations, or reserve capacity configurations, could be equal to traditional serial systems with buffers. They described the importance of having a reliable material handling system to achieve the productivity improvement in the proposed nontraditional configurations. Spicer et al. (2002) illustrated many fundamentals that help in the design of the system configuration. They mentioned how the continuous change in customer demand generates the need to obtain new alternatives for system configuration other than the traditional serial layout. These alternatives could be facilitated by the reduction in the cost of Computer Numerically Controlled (CNC) machines, and the introduction of new technologies. They also mentioned that multi-axes CNC machining systems generate multi-task systems that have the ability to align with different possible system configurations. The authors differentiated between two configurations with and without crossover, and between the symmetric and asymmetric configurations. In contrast to the symmetric configurations, in the asymmetric configurations the parts are processed by various plans and on different numbers of machines. They mentioned that the system configuration is defined not only by the machine layout but also by the machine connections. They introduced the effect of setup arrangements on system configuration length, which is determined by the number of machines the part should be processed on . The configuration length can be adjusted either by operation division or setup combination . They introduced the system configuration width as a number of parallel machine sets. Besides determining the upper and lower bounds for the configuration length, they also determined these bounds for the configuration width.

Layout Design Problems and Challenges
There have been many articles developed to study, investigate, and build models for layout design. Many targets and constraints have been considered in the layout design and many challenges have faced the researchers through their studies. In some of the literature, models for block layout (or physical arrangement of departments and machines) were built, where other literature looked at the layout as how the product parts flow through the machines, regardless of the physical location. Kochhar and Heragu (1999) proposed a methodology for block design of a multiple-floor facility to be more flexible in the ability of rearranging the departments, workstations or machines in future periods . The objective of their methodology, the Dynamic Heuristically Operated Placement Evolution (DHOPE) , was minimizing the rearrangement costs and material flow costs by designing a dynamic facility over two consecutive periods. They pointed out that the infrastructure and the supportive tools should be designed to have the ability to facilitate the rearrangement actions. Hassan (1994) studied the problems that relate to the machine arrangement in production floors for modern manufacturing, and he argued that block layout is not applied practically in actual production systems and hence, other methodologies are required . The author demonstrates the importance of machine layout and the factors that play roles in its development. He differentiated among three types of layout: single row, multiple rows , and the loop layout ( fig.1 ).
The author also explained different measures that affect choosing among these layouts. Hassan (1994)  The comparison was in terms of problem formulation , solution methodology, layout type (row, multiple rows, or loop layout), and problem size. Evaluation of the solutions and results were mentioned also for each methodology. He criticized many of these suggested layout procedures. For example, he argued that these procedures were not performed in real manufacturing plants.   (Hassan, 1994). Klampfl et al. (2005) mentioned that although the mixed model assembly line responds quickly to variations in customer demand, many efforts are required to find the preferred production layouts. The authors proposed three formulations for optimizing the layout in automotive assembly lines. Their method was based on the lean system concept of decreasing the non value added time. They defined and used the e-Workcell tool, which is a tool that has been used by Ford to develop a 30 image of the workcell elements according to data from the production plant. They describe their optimization problem, and the objective of reducing the non-value added time which is affected by the arrangement of the workcell elements. The authors proposed several formulations (unconstrained, one-dimensional constrained, and two-dimensional constrained) to solve the objective function and optimize the layout. Their proposed methods for optimizing the layout efficiency enhance the objective of increasing the flexibility in the production system to respond efficiently to different product mix. Kaebernick et al. (1996) presented an integrated model for designing the layout of a Cellular Manufacturing system . The authors pointed out that there are three phases of Cellular Manufacturing design: parts/machine grouping, the layout of machines within cells, and the layout of cells in the shop floor. They emphasized the importance of layout design as an essential element in achieving the desired benefits from Cellular Manufacturing. Their developed model is based on evaluating the integrated effects of the three phases of design on the overall system. It provides multiple alternatives which need to be assessed based on the predetermined criteria. 20

Performance
Some authors have started to recognize the strong relationship between the system configuration and its performance in their studies. The traditional serial configuration may be insufficient for adapting to dynamic customer demands, and hence, the need for new configuration designs is introduced with many associated challenges and problems which need to be considered. This section presents literature that has shown the impact of system configuration on its performance, and then presents some research that have focused on studying different performance measures. Koren et al. (1998) discussed how different configurations of the manufacturing system lead to different performance. They described that choosing the preferred configuration is based on the following measures of performance: productivity, reliability, quality, capacity, and cost. They mentioned that there is a relationship between the number of machines in the system and the number of configurations that could exist. They analyzed the effect of system configuration on performance by proposing a methodology of comparing many configuration alternatives based on the above performance measures. Spicer et al. (2002) demonstrated the relationship between the system configuration and its performance by comparing four symmetric configuration alternatives (pure serial, pure parallel and short serial lines in parallel with and without crossover) based on four metrics: throughput, line balancing, system investment cost, and capacity scalability. Zhong et al. (2000) pointed out that there are different metrics for evaluating the performance of different configurations for the production line, where an integrated method that combines all measures is more efficient in selecting the best arrangement for machines. The authors proposed methodologies to determine three measures of performance: productivity, quality, and convertibility. As a case study, they compared four configurations of the machining ,system (serial, parallel , and two hybrid configurations) in terms of productivity, quality, and convertibility.

Productivity and System Reliability
System productivity is often the most important target that most manufacturers try to improve. There are many researchers who have used productivity as a basic objective in building their models for improving the system performance.
In terms of throughput, Spicer et al. (2002) explained examples of the effect of configuration on the gross throughput, on the expected throughput, and on the probability distribution function . Zhong et al. (2000) used machine availability and complexity and processing rate as parameters for determining system productivity. sun et al. (2008) investigated the relationship between the machine configuration, machine-level reliability and the system-level reliability. They developed a methodology for analyzing the quality reliability at both the machine level and the system level. They analyzed the machine level quality reliability by considering the joining effect of product quality and tool degradation. Then they used the proposed integrated model for machine-level reliability to predict the system level reliability for serial-parallel hybrid manufacturing systems with various selected configurations

Quality and Stream of Variation Theory
Quality has been considered as a fundamental measure of system performance. Many studies have analyzed the effect of good quality on improving the production system. Colledani and Tolio (2006) illustrated that configuration has an impact on both the productivity and quality control performance . They emphasized the fundamental importance of considering the quality and throughput measures together during the design phase of the production system . Colledani and Tolio proposed a method for evaluating the production system performance by considering both of these measures. They mentioned that there are many tools and techniques that have been developed to predict the production throughput and monitor the quality level, but only recently some literature has begun recognizing the importance of using an integrated method to deal with both of them. Zhong et al. (2000) used the mean and standard deviation as measures for the quality of the finished product.

Stream of variation
Recent works have analyzed the effect of the production and process system design on the quality of the finished parts. The stream of variation theory has been introduced to predict how the variability of products propagates through the workstations of a production system. Hu and Koren (1997) suggested a model for the flow of dimensional errors when the product moves from one station to the next. The authors said that the assembly variation is larger than the variation in the component, since the dimensional variation is accumulated while the product is transformed through the assembly system. The developed model helps in solving the quality problem earlier, since it increases the ability to predict the variation evolution through the assembly line. They demonstrated that sheet metal parts in the automotive industry can be deformed through the assembly process, and that causes an increase in the dimensional changes. Hu and Koren (1997) explained the two elements of the Stream-of-Variation theory: 1) analyzing how the dimensional variation of the components is propagated through the assembly line so that the variation of the final product can be anticipated and 2) determining the variation causes and sources in the assembly line. He studied the relationship between the propagation of the assembly variation and the assembly configurations. He showed that variability and the diagnosability features of the serial and parallel 24 assemblies are different. The author identified two causes of variation in the automotive body assembly: (1) by fixture and (2) by part or welding process deviation. He also developed a fault diagnosis strategy to reduce the variation. Djurdjanovic and Ni (2006) demonstrated that the dimensional variations are transformed and accumulated from one operation to the next through a multistation machining system . The authors introduced methods for measurement scheme analysis in multi-station machining systems to predict the root causes for the dimensional faults . In this methodology, the combinations of measurements are evaluated in terms of the goodness of information they provide about the machining process parameters. The best set of measurements is the one that is more informative in reflecting the root causes of the machining dimensional errors.
The stream of variation model was the basis for their methodology.

System Responsiveness
Zhong et al. (2000) pointed out that the system responsiveness is measured by the convertibility and scalability of the system, where the convertibility reflects the system ability to produce different products, and the scalability reflects its ability to adjust the production volume depending on the demand, with less cost and less time. They asserted the importance of the convertibility measure in the Reconfigurable Manufacturing System (RMS). They proposed using the time required to convert the system as a measure for its convertibility. In terms of capacity scalability, Spicer et al. (2002)  Many research projects have been constructed to investigate and analyze the impact of the RMS. Koren et al. (1999) showed that the reconfigurable system is required due to short pro duct lifetimes and continuous change in customer demand . They argued that there were studies and technologies created to develop the product design, but not that many efforts were invested to develop a method for designing the production system . They demonstrated that the reconfigurability of the system makes the system more adaptable to the changes in the capacity functionality, which will reduce the system design time. The authors compared three systems:

Maier-Speredelozzi
Dedicated Manufacturing Lines (DML), Flexible Manufacturing Systems (FMS), and Reconfigurable Manufacturing Systems (RMS). DML produces high throughput but it is not scalable or flexible, and FMS is scalable but expensive and has general flexibility, while the RMS is scalable and has customized flexibility.
They defined the RMS and its main aspects to be CNC machines, Reconfigurable Machine Tools (or adjustable machine structure), and reconfigurable software and hardware. All of these enhance the quick response to the fluctuation in demand and product mix. The authors demonstrated five main characteristics for the reconfigurable system: modularity, integrability, customization, convertibility, and diagnosability. They also explained issues in RMS: system-level design, machine-level design, and system ramp up. They supported their discussion with examples about a reconfigurable machining system.

Azab and EIMaraghy (2007) proposed a mathematical model for
Reconfigurable Process Planning and sequencing for the reason that the customer demand becomes unpredictable and requires an effective changeable or reconfigurable manufacturing system . They formulated the problem and set up the objectives and constraints, taking into account choosing the best location for the new required features . Hon and Xu (2007) discussed the complexity that dynamic demand adds to the system design problem . The authors investigated the effect of the dynamic product life cycle on the timing and extent of reconfiguration. They asserted that companies should know when and how to perform a reconfiguration , where this decision depends on the individual circumstances and characteristics of each manufacturing system. The authors mentioned that there are three common methods for adding capacity to the system configuration , which are: using better tools or machine reconditioning, replacing the current machine, or adding a new machine. Th ey used the simulation method to analyze the multi-stage multiproduct manufacturing system. The three steps in their simulation approach are: modeling the original configuration , optimizing the reconfigured system , and then optimizing the product portfolio. Colledani et al. (2005) illustrated the meaning of changeability which reflects the ability of the company to respond to the unpredictable customer demand with minimum cost. They discussed the system reconfigurability which is one aspect of the system changeability. They proposed an approximate analytical method to assess the Reconfigurable Manufacturing System (RMS). They introduced a model based on two levels of decomposition: machine and buffer.
Their suggested model was applied as a case study to evaluate different configurations as alternatives to increase the system capacity. Shabaka et al. (2007) mentioned two levels of reconfiguration for RMS: system and machine. They explained how the ability of re-assigning the resources at these two levels leads to increases in the flexibility of the production system which becomes more able to cope with changes in product demand. The authors demonstrated how different configurations of machine axes of motion cause different capabilities for the machine. They proposed a methodology to construct the machine tool configuration based on part geometry and features , operations precedence constraints, tool approach directions (TAD), operation clusters, and the integration between the product features and machine capability.

Scrap and Rework in the Production System
Scrap and rework are two factors that have an impact on achieving the goal of improving the performance of a production system . Their effects could significantly increase cycle time and production cost, and decrease system efficiency. There are many scholars who have recognized that and hence, they 29 have started to focus on studying and analyzing scrap and rework in the production system.

.s.1 scrap and Rework Definitions and Effects on Production System
Bowling et al. (2004) classified an item as scrap if the value of the quality characteristic of interest is less than the lower specification limit and as rework if the value is more than the specification limit. Bohn and Terwiesch (1999) mentioned that in most production systems not all raw materials entering the line to be processed make it to the end and turn into a good quality final product. In other words, the number of parts entering the process does not necessarily equal the number of parts coming out as finished products. They mentioned the possibility of having some scrapped items either directly from the main processes, or from the rework processes. They also explained that at the checking point components are tested and classified either as good or as defectives items. If the item is good then it proceeds to the next process, and if not, it could be reworked or scrapped . They explained the high impact of scrap and rework on the system capacity utilization especially if defects are detected at bottleneck machines. They described that more value is added to the items as they move forward through the production system, and hence, having more reworked or scrapped items at later processes means more losses. The authors studied the effect of yield losses, caused by reworked and scrapped items, in yield-driven production processes.
They demonstrated the importance of improving the yield in increasing the productivity. Besides that, they analyzed the economics of multi-stage yield-driven production processes and showed the effect of consuming cost and time due to scrap and rework processes. They summarized the rework and scrap effects on material, labor, capacity, and variability related costs.

.s.scrap and Rework Considerations in Production System
Modeling Bowling et al. (2004) presented a model to determine the expected profit by considering the cost of processing , scrap , and rework. The authors used absorbing Markov chains to identify the long term probabilities of the three states of the product during the production process: rework , scrap , and accepted. The long term probabilities were used to build a profit model for one, two, and n-stage serial production systems to determine the optimal process target at each stage.
They applied sensitivity analysis by varying the model parameters to study their impact on the output. Pillai et al. (2008) showed that Markov chain models can be used to represent the production system under uncertainties due to rework and scrap. The authors modeled material flow through a serial production system with rework and scrap by using the absorbing Markov chain , where both the scrap and finished product states were considered as absorbing states. By using the model they adopted, system design and also production and inventory control could be efficiently improved. Using an absorbing Markov chain model enables the authors to build equations to determine the following: material requirements, number of machines, production cost, and manufacturing lead time. Davis and Kennedy (1987) considered rework , work in process , inspection, and scrap in their modeling of a serial production system. They also took into account the tool wear and the scheduling impact of the tool wear rate utilizing a Markov chain model for the production system . They analyzed the serial production system under different scenarios using the Markov models to have outputs that assist in solving production planning problems .

Challenges in the Biopharmaceutical and Pharmaceutical Industries
Biopharmaceutical and pharmaceutical companies, like other industries, are facing the ever-increasing challenges of improving their production systems to produce high quality products with the lowest cost. Many production regulations in the biopharmaceutical and pharmaceutical industry increase the challenges and complexity in designing the production systems . Researchers have studied these challenges, analyzed data from these industries, and pointed out the issues that biopharmaceutical and pharmaceutical companies should start to consider for improving their outcomes. Abboud and Hensley (2003) talked about how pharmaceutical companies and the FDA have begun recognizing the importance of focusing on improving the production system instead of focusing only on developing new drugs for the market. They mentioned that the increasing number of recalls and the high percentage of defects in pharmaceutical production were strong indications for the FDA (Food and Drug Administration) about the need to start improving and introducing new methods and techniques in the production process. The authors of this article mentioned that many production shortcomings in the pharmaceutical industry are due to the production system's insufficient tools and methodologies.

They included results from the analysis of financial statements by Raymond
Scherzer, 2001 , in which he illustrated that the production cost for the top pharmaceutical companies are 36% of the total costs, which is two times the expenses for Research and Development (R&D) . They also suggested that concentration on improving the manufacturing processes is the rational way to save money and increase the pharmaceutical companies' profits, since there are fewer chances to develop new drugs or to increase sales . The authors gave examples from some companies, like Pfizer, that start investing in their production systems by applying new technologies to improve production efficiency. Basu et al. (2008) discussed a study aimed at analyzing the production cost for some pharmaceutical companies, which was divided into three categories: brand, generic, and biotech companies. They also studied the relationship between the material and production cost or Cost of Goods Sold (COGS) , and the Research and Development (R&D) spending . They mentioned that some studies indicated that the cost for production was more than three times the expenditure for R&D. The authors illustrated that one of many factors causing the high production cost in the pharmaceutical industry are the FDA regulations that require the companies to abide by the "current Good Manufacturing Practices" (cGMP). By comparing the annual reports of the top pharmaceutical companies, the authors demonstrated that in 2005 the revenue of the brand name companies ($10 billion) was two times the revenue for the generic companies ($5 billion). On th er hand they expected that the generic companies would have much the o ' higher revenue in 2008, due to the expectation that more than two thirds of prescriptions would be converted to generic substitutes instead of brand name drugs.
By comparing the data of pharmaceutical companies with the highest market share from the above three groups over a study period of twelve years (from 1994 to 2005) , Basu et al. (2008) found the following : _ In terms of COGS% (COGS as a percentage of total sales), the generic companies have a higher average than that for brand name compan ies which are higher than the average of the biotech companies .
-In terms of R&D% (percentage of R&D spending divided by sales) and operating income, biotech companies have a higher average, and then the brand name has a lower average, while the generic companies have the lowest average.
In terms of general expense, brand name companies have the higher average, then the biotech , followed by the generic companies.
-For the brand name companies there is a "strong negative correlation between the COGS% and R&D%".
-For the generic companies, there is "strong negative correlation between the COGS% and the operating income".
-The manufacturing costs for the generic and brand name companies are around 50% and 27% of the total sales revenue , respectively. Basu et al. (2008) gave a hint that reducing the production cost by improving the efficiency of the production system will reduce the price of the drugs, increase the investment in R&D, and increase the pharmaceutical companies' profits. Odum (2005) pointed out that biopharmaceutical design engineers should reduce the production cost and capital expenditures, and improve the flexibility in the production system . Odum identified three attributes that influence the facility design. These are product, process, and facility attributes. He emphasized that there are fundamental principles that should be fully understood to improve design and operation in the biopharmaceutical industry. One of these principles is that the process and facility designs are connected to each other. The author addressed the implementation of closed process system, controlled processing capabilities, and manufacturing flexibility.
He mentioned that there are companies starting to produce multiple products, and that there is new trend to apply concurrent manufacturing to help achieve multiproduct facilities. He represented the need for building strong strategic planning for the manufacturing capability in order to cope with the dynamic markets of many biological products. He also explained the changes in equipment design Principles due to the increasing trends to move to a multi-product manufacturing system which require the equipment design to be more flexible and cost effective.

7
Studies that Apply Industrial Engineering Concepts or Tools in System 2.
Design at Biopharmaceutical and Pharmaceutical Industries Oh et al. (2004) proposed a method to improve the throughput and the capacity of the production lines of Monoclonal antibodies (Mabs). The authors demonstrated that the demand for Mabs products keeps increasing while the biomanufacturers are unable to increase their production capacity to meet this high demand. They suggested using a simulation tool from Superpro® software to determine the constraints or problems in such design, instead of using the traditional method of conducting pilot plant experiments, which is a more costly and time consuming approach. By demonstrating a case study, the authors gave an example of using Superpro to simulate the processes by entering the inputs from a simple laboratory experiment. Then the material flow steam was constructed, and then the batch process simulation model was obtained. After that, the software was used to produce the machine utilization chart, which was used to determine the bottleneck in the system. The authors suggested many scenarios to increase the system capacity and productivity. They economically analyzed one new design scenario by applying a cost and profit analysis of increasing the number of machines to study its impact on increasing the number of batches produced and profit. objective, usually determined in advance, while their method is more flexible and allows for choosing objectives that are required for each specific layout. The authors asserted difficulties and constraints that affect the layout design due to the nature of the production of the pharmaceutical companies. Some of these constraints are: the contamination which could result from using different materials, the predetermined methods for transporting the materials, the high cost for making changes in the design, and also the FDA regulations. All of these constraints should be considered carefully in building the model, which should be designed in a perfect way from the beginning since it may stay and be used for a long time without the ability to be adjusted . Hamamoto et al. (1999) used a genetic algorithm in their model as the search algorithm, and the growth and band layout methods as the decoding algorithms. Their model's objectives are improving the throughput and reducing the traveling time. They compared the layouts they generated , the actual layouts, and also the ones which resulted from existing layout algorithms such as: CORELAP (construction algorithm), CRAFT (improvement algorithm) , and BLOCPLAN (hybrid algorithm). Newcombe et al. (2008) pointed to many hidden sources of variability in the manufacturing process of biological products. Hence it is important, during the development stage, to propose manufacturing processes that will have the ability 37 et P redetermined specifications. They introduced the "design space" term, to me which indicates a range of process parameters that assure producing products with good quality if they fall within this range. They mentioned the increased trends in applying Design of Experiments (DOE) for modern biopharmaceutical development. They mentioned that DOE could be used to define the design space for the process. The authors also showed how the DOE approach is very useful in finding the critical process parameters and the quality attributes. The authors applied the DOE approach in case study to analyze the impact of some factors on stability of an antibody derived biotherapeutic. authors are focused on studying the effect of system configuration on the following metrics: productivity, quality, scalability, convertibility, changeability. Table 2.2 includes the articles that are mentioned under each of the following subjects: layout design, relationship between system configuration and performance, Reconfigurable Manufacturing System, effects of scrap and rework on the production system, and biopharmaceutical industry.  " Colledani et al. (2005) " "

Summary
.... .  Davis and Kennedy (1987) used Markov models for a serial production system to assist in solving problems that related to production planning. Bowling et al. (2004) used the absorbing Markov chain to build a profit model for one, two, and n-stage serial production systems to determine the optimal process target at each stage. Pillai et al. (2008) also used the absorbing Markov chain to represent a serial production system, in order compute material requirements, number of machines, production cost, and manufacturing lead time. In this thesis, absorbing Markov chains are adopted to model four system configurations: serial, parallel, serial-parallel without crossover, and serial parallel with crossover, in order to develop models for cost of quality, productivity, and cycle time, that can be used in comparing the performance of the different configurations to find the preferred configuration. Chapters three, four, and five include develop the models for production systems with two, four, and n processes, respectively. This chapter focuses on a two process production system, where two different configurations are applied: the serial (where products go through two stages) and parallel (where products go through one stage) configurations. A Product is processed at any given stage and then moves to the next step if it 42 meets quality specifications, while it gets reworked or scrapped if it is defective.
Three models for cost of quality, productivity, and cycle time are proposed to define the configurations that obtain the best output of each model. The results show that the different configurations have different values of quality cost, system productivity, and total cycle time.
A brief description of the Markov Chain tool is discussed first, since it is used as a starting point to develop the three models for each configuration. After building the models, simple numerical examples are used to apply these models, and then sensitivity analysis is provided.

Markov Chains
When a product goes through multiple states in a production line, the probability that describes the product at any specific state is a function of that process or state. As the product proceeds in the line, the probability at each state becomes related to the prior process, i.e. the probability of state i depends on that of state i-1, while it is not related to the probability at any states prior to state i-1.
This process is described to follow a Markov Chain process. The independence of probability at state i from all previous states (except i-1) makes the Markov chain method widely applicable in industry to predict the long term and steady state Probability distribution of any given state, since the probability at any state can be defined by only the state that is immediately before, and it is independent of all prior states and any history of prior states.
In a Markov Chain process, the transition probabilities are considered. A transition occurs when the system moves from one state to the next state during one period of time. A transition represents when the system changes its state from ; to j at a probability of Pij· For example, in a production line, when a product moves from process 1 to process 2, then the transition probability is defined to be the probability of moving a product from process 1 to process 2, which can be referred to· as the acceptance probability between 1 and 2. There is a primary assumption related to Markov chains, which is the stationary assumption. It implies that the transition probability of any state i to another state j does not change over time and is always the same. A Markov Chain is usually represented as a matrix of transition probabilities between any two states.
One of the widely used applications of a Markov Chain is the Absorbing chain which includes absorbing states in addition to the transient states. An absorbing state is a state when the system or product enters permanently without any chance to leave to any other state, i.e. the probability of leaving an absorbing state is zero, and the probability that a product or the system stays in the same absorbing state is 1. An example of an absorbing state could be the finished Product state, or scrap state. Absorbing Markov Chains are used to find the long term probabilities of states that the product of the system would eventually enter, like a finished product or scrap state.
44 Two stage Serial System with Two Processes

3.
consider a two stage production line, where one process is in each stage.
A product is processed at the first stage, and then needs to go through the second stage for another required processing step before it reaches the finished product status. This type of production can be referred to as a two stage production system with two processes. These two production stages do not necessarily have to be physically in series, as long as the product needs to go through process X1 and then process X2 before it can be called a finished product. In a serial system raw material is released to enter the first production stage, stage 1, where process X1 will occur. Stage 1 processes the raw material and adds value by performing certain set of operations, which will add new features to the raw material. At the end of stage 1, a quality characteristic needs to meet specific requirements defined by the customer (internal or external customer); hence, each product needs to be inspected against the pre-defined requirements, or the customer specifications. When a product is inspected at stage 1, there will be three outcomes that are assigned to that particular product.
-A product can proceed to the next stage (stage 2) if the quality characteristic at stage 1 falls within the acceptable range of the customer specifications.
-A product can be reworked by re-entering stage 1 if the quality characteristic at stage 1 falls within the rework range in the customer specifications. Each reworked product will go through the same exact 45 operations at stage 1 and will be then subjected to the inspection station at stage 1.
_ A product is scrapped and will never be processed at any other stage if the quality characteristic falls within the scrap range in the customer specification.
Note that the rework and scrap ranges are not defined by the customer, but rather defined by the process capability, engineering assessment, nature of the product, ability to rework a product, etc... For example, consider a turning machine at stage 1, where the manufacturer cuts metallic shafts and produces to customerdefined specifications of 1 O+/-0.5 mm in diameter, i.e. the acceptable range is 9.5 mm to 10.5 mm . If the shaft is produced at 9.3 mm, then it cannot be reworked, since the material has already been cut from the shaft, and it will be scrapped (material cannot be added to the shaft to bring the diameter up to the acceptable range). While if the shaft is produced at 10.7 mm, then it can be reworked by reprocessing at stage 1 to reduce the diameter to within the acceptable range (9.5 mm-10.5 mm).
A product that can be reworked will be sent back to the same stage, stage 1, to be processed again, while a scrapped product will be sent to scrap bin and never leave that bin. However, a good part at stage 1, which is a product that has a quality characteristic that falls within the acceptable range of the customer specifications, will be sent to the next stage, stage 2. Hence, only good parts from stage 1 can be transferred to stage 2.
At stage 2, other set of operations are performed on the incoming products from stage 1, where operations will attach a new quality characteristic (different from the quality characteristic at stage 1 ). At stage 2, each product will be inspected against the customer specifications defined for the new quality characteristic at stage 2, and each product can go through one of the three options defined above (rework, scrap, or accept). When the item is accepted at stage 2 (and it is also accepted at stage 1 ), then the part is considered as finished product, and it can be sold to a customer as finished goods.
Overall , we have two production stages, where each product is processed, and a unique quality characteristic is attached to the final product at each respective stage. There are two rework loops, the first rework loop takes items at stage 1 back to the same stage to re-process them until they are accepted or reworked. The second rework loop is defined at stage 2, where a rework-able item will be sent back to the same stage (stage 2) to be processed again. Similarly, there are two scrap states, where an item at stage 1 is sent to the scrap bin and will never leave if it is considered scrap-able, and also scrapped items at stage 2 will leave the production system and will be sent to the scrap bin is the quality characteristic at stage 2 is within the scrap range. Finally, we have two acceptance states, where good items at stage 1 are sent to stage 2, and all acceptable items at stage 2 will be sent to the finished product state.
The quality characteristic at each stage, stages 1 and 2 are unique and independent of each other; the acceptance of a product at stage 1 (implying the 47 quality characteristic is within the accepted range) does not affect the acceptance of a product at stage 2, nor does it influence the quality characteristic at stage 2.
3 .3.1 Model for Two Processes-Two Stages in Series Figure (3.1) shows a flow diagram for the two stage production system in series.

RM
State (1) State (2) State (3) Finished Product State (4) Fig. 3.1: Two processes production system in series In this model , process X 1 (at stage 1) is considered as a transient state, state 1, whereas process X 2 (at stage 2) is considered a transient state 2. A transient state is any state that the system or the product enters and eventually leaves to another state. States 3 and 4 are considered to be absorbing states, since any product that enters states 3 or 4 (scrap or finished product respectively) stays and never leaves. For example, consider a production line where two chemicals are mixed at 48 twO different stages. One chemical is added at each stage and the two production stages tor adding and mixing are considered to be transient states. When the product is finally mixed and meets all the specifications, it is sent to the finished product storage area and never returns to any other states. If the product fails to meet any of the specifications at either production stages, it will be scrapped (if it cannot be reworked), and sent to the scrap bin, where it stays and never leaves.
In this example, the finished good storage, as well as the scrap bin are considered to be absorbing states, whereas all of the production processes are considered as transient states, since the part will eventually leave to another transient state (production process), or absorbing state (finished good or scrap).

Building Absorbing Markov Chain Matrices
In order to build an absorption chain matrix for any production system, we need to understand the material flow and the short term probabilities of moving from each state to the next. By collecting data about the considered quality characteristic at each production stage, we can understand the quality characteristic distributions of acceptance, rework, or scrap at each stage, and then the short term probabilities could be used to develop an absorbing Markov is the probability that an item at state i will go to state j, where i = 1, 2, 3, and 4 and j = 1, 2, 3, and 4. For example P12 is the short term probability that a part which starts at state 1 (process X1) will go to state 2 (process X2).
Matrix Pis divided to four sub-matrices: Q: is the matrix that contains the probabilities of moving from one transient state to another transient state.
R: is the matrix that contains the probabilities of moving from a transient state to an absorbing state. 0: is the zero matrix.
I: is the identity matrix.
The absorbing Markov chain that includes the long terms probabilities of moving from transient to absorbing states equals: (I -ar 1 R, such that: Where: a 11 : is the average number of times that an item from state i stay at state j.
Hence, the long term absorbing Markov chain matrix becomes: .. · . the long term probability that an item from state i ends up in state j.

b11· IS
For example, b13 is the long term probability that a part from state 1 (process X 1 ) ends up at state 3 (scrap pin), and bz4 is the long term probability that a part from state 2 (process X2) ends up at state 4 (finished good storage).

Resource Consumption Models
In any production system it is important to improve the cost of quality, productivity, and cycle time to stay competitive. There are four categories of cost of quality as determined by Juran et. al ( 197 4 ): ( 1) internal failure costs, such as scrap and rework, (2) external failure costs, such as warranty charges, (3) appraisal costs, such as inspection and test, ( 4) prevention costs, such as quality planning. In this thesis, however, the cost of quality is considered to be a total of scrap and rework costs only. Scrapping a part means throwing away the unfinished product and losing all values that are added to that part such as raw material, labor hours, processing time, and overhead. Reworking a part means reprocessing that part again which consumes resources such as labor hours, Processing time, and overhead. Reworking a part means reprocess the part again and that consumes resources such as labor hours, processing time, and overhead. Productivity is considered as a ratio between the system output and input. It is preferred to have higher productivity, but there are many factors affect t nd reduce the productivity such as the existence of the scrap and rework, tha a and hence not all the raw material (input) ends up as finished products (output).
Cycle time is the total of the processing and reworking time from the point the part enters the system until it leaves as finished product. In this section three models are developed to identify the resource consumption per part. These resources could be quality cost defined by rework and scrap waste which will be translated into cost, material consumption measured by the productivity, and time consumption to produce one item measured by the cycle time. Equations (3.2) and (3.3) are used to develop the three models.

Cost of Quality Model
The cost of quality model is developed by recognizing that the expense of quality problems existing in the production system will be the sum of the total cost of reworking or scrapping an item. The quality cost is therefore:  (1) Assume that we have two processing stages with two processes, A and B, and their quality characteristics follow the normal distribution.   That means that on average each item needs 8.10 minutes to be produced.

One Stage Parallel System with Two Processes
Consider a one stage production line consisting of two production processes in parallel, namely process X'1 and process X' 2, both manufacturing the same product. Process X' 1 manufactures a product by adding value to meet pre-defined specifications required by the customer (internal or external). An item is pulled from the raw material store and process X' 1 performs certain operations on the item, where an inspection is being done to measure and evaluate a specific quality characteristic, and then a decision is made on either accepting the part and sending it to finished goods store, scrapping the item, or reworking it. If the quality characteristic falls within the acceptable range of customer specifications, then the part is considered as a good item, and it is sent to the finished goods storage ready to be sold for the customer. On the other hand, if the part falls within the rework range, then the part will be sent back to the same process (process X'1) to be reworked, by performing more operations to bring the quality characteristic into 58 Cep table range, and the item will be inspected again. A part is scrapped if the ac the quality characteristic falls in the scrap range, at which time an item is rejected, and sent to the scrap bin to never leave. Process X' 2 manufactures the same product, and each item that is pulled from the raw material store goes through the same set of operations, and similar decisions will be made: accepting the part, reworking, or scrapping the part. Processes X' 1 and X'2 run independently in parallel. Any item that is manufactured by process X' 1 will be reworked only on process X' 1 , while any scrapped or accepted part will end up in the same scrap bin or finished goods storage. Rework loops are uniquely defined by each process; each process X' 1 and process X'2 has a separate rework loop, and there is no crossover between processes at the rework loops (a reworked item at process X' 1 can not be reworked at process X' 2 , and vice versa).   (1) State (2) Finished Product

Model for Two Processes-One Stage in Parallel
State (4) State (3) Fig. 3.2: Two processes production system in parallel Process X' 1 and process X'2 perform the same tasks, where the assumption is that both processes X' 1 and x· 2 could perform the same processes that both processes X1 and X 2 can perform in the serial configuration model. Returning back to the example of adding two chemicals (described in section 3.3.1) in a parallel configuration each process has the ability of adding and mixing the two chemicals. This could require the two machines at process x· 1 and X' 2 to be more flexible and probably more expensive than X 1 and X 2 in the serial configuration. In this example, there are no transitions which occur between X' 1 and X' 2 . In other WOrds, P12 and P21 are equal to zero. The transition only occurs from X'1 or X'2 to either scrap or finished good storage which are the absorbing states. .. · . average number of times that an item from state i stays at state j.

Pi4
(1-Pi I) P24 1-P 22 bii: is the long term probability item from state i ends up in state j.

Resource Consumption Models
(3.14) Equations (3.13) and (3.14) are used to build the models for cost of quality, productivity, and cycle time for the parallel configuration system.

Cost of Quality Model
Using the same procedure that is used in developing equation (3.6), quality cost for the parallel configuration system is: l ·ty cost == Rework Cost + Scrap Cost aua1 Where; QC: cost of quality per item.
RCi: the cost of reworking one item from process (X' i).
sc 1 : the cost of scrapping one item from process (X' i).
Ni: total number of raw material parts that enter process (X'i). Productivity is the percentage of the production system total output to the Production system total input. The output is the total products accepted as finished d from both process X' 1 and process X' 2. The input is the total raw material goo s which enters to both processes. Hence: productivity = [(N1 x probability an item from process x· 1 is accepted) + (N 2 x probability an item from process X'2 is accepted)] I (N1 + N2) Therefore:

Cycle Time Model
The expected cycle time for the two process production system that is shown in . ycle time at state i, which includes processing time and reworking time at C I· C process X' i · PTi: processing time at process X' i RTi: reworking time at process X' 1 (2) Considering the same example of adding and mixing two chemicals, the assumptions are:

Numerical Example
1-A' and B' have the same mean, standard deviation, specification limits, processing and rework times, and rework and scrap costs.
2-The considered quality characteristic at A' and B' is the same as in process B in the serial configuration system (the specification limits equal to that for process B). 1 . The scrap cost at each of processes A' and 8' equals the summation of scrap costs at processes A and B.

3-The considered quality characteristic at
Based on the above assumptions, the parameters required for calculating quality cost, productivity, and cycle time for the parallel production system are summarized in table 3.2. In this section assumptions and values from example (1) and (2)    and sc 2 = 2.2 stay constant. It indicates that the quality cost for both o. , configurations increases as SC1 increases. It also shows that when SC1 is less than $0.264 the serial configuration has lower quality cost than the parallel. On the other hand when the SC1 is more than $0.264 the parallel configuration has the lower quality cost.  Figure (3. 7) shows the effect of mean 1 on system productivity while mean 2 = 22, StDev 1 = 1.1, and StDev 2 = 2 stay constant. It indicates that the system productivity of the parallel configuration almost stays constant as mean 1 increases. For the serial configuration, the system productivity increases and then become constant. This figure shows that the parallel configuration has higher productivity than the serial.

70
Effect of Mean 1 ::: 1.1, and StDev2= 2 stay constant. It indicates that for both serial and parallel configurations, system productivity increases as mean 2 increases. It also shows that the parallel configuration has higher productivity than the serial for all values of mean 2.

Effect of PT
. .   In terms of quality cost, figures 3.3 -3.6 show that within the specified range of each factor, parallel configuration performs better than the serial at different rework cost, while at different scrap cost the parallel or serial configuration could perform better. In terms of productivity, all of the figures 3. 7 -3.10 show that system productivity for parallel configuration is higher than that for the serial (within the specified range of each factor). In terms of total cycle time, figures 3.11 -3.14 show that at different processing time both configurations perform almost the same since there is slight difference, while at lower rework time the serial configuration performs better, however at higher rework time the parallel configuration performs better (within the specified range of each factor).

Abstract
In this chapter, four processes are considered in serial, parallel, and mixed serial-parallel configurations. Three resource consumption models are evaluated for each configuration to select the one that results in a better objective. Similar to previous analysis, a Markov chain is used to build the resource consumption models (cost of quality, productivity, and total cycle time), which represent the long term probabilities of good, rework, and scrap parts at each process.
Four machines can be arranged in different patterns giving different configurations. As shown in figure 4.1, four machines c9uld be arranged one after each other obtaining a pure serial configuration; all of them could be arranged in pure parallel, or two serial machines could be parallel to another two serial machines. The mixed parallel and serial configuration may or may not allow crossover between production system stages. Each configuration is studied and modeled, and resource consumption models are developed, in order to analyze the performance of each configuration in terms of the output of each model.
Simple numerical examples are used to apply these models.

Notation
In this chapter, notations similar to that in chapter three are used. All notations that are used in chapter four's sections are listed below.
• Pii: the probability that an item at state i will go to state j.
• 0: matrix that contains the probabilities of moving from one transient state to another transient state.
• R: matrix that contains the probabilities of moving from a transient state to an absorbing state.

79
. •Xi: process i in the production system.
•QC: quality cost per item.
• RC 1 : rework cost per item from process (X1) . • N1: total number of raw material parts that enter process (X 1 ).
• CT1: cycle time at process (Xi), which includes processing time and reworking time, where: 80 CTi ::: (processing time at process (Xi)) + [(probability an item is reworked at process (Xi)} x (rework time per item at process (Xi))] • PTi: processing time per item at process (Xi).
• RTi: reworking time per item at process (Xi).
W . . the percentage of accepted parts from process x "i that go to process x"i· • lj·

four Stage Serial System with Four Processes
Consider four processes in series, where raw material enters the first process and a product is processed and it moves to the second process and so on. At each process, value is added to the product by performing certain operations. Typically, a product quality characteristic is inspected at each respective process, and compared with pre-defined specifications. When the quality characteristic falls within the acceptable range , the product is accepted at the respective process and then transferred to the next process, and when the part finishes process four, it is then moved to the finished goods storage. On the other hand, when the product falls outside the acceptable range, it could be either reworked or scrapped , depending on the value of the quality characteristic. If a Product falls in the rework range at any given process, the product is re-entered into the process to be reprocessed and re-inspected against the same specification limits. A scrapped part is a part the falls in the scrap range of the 8pecifications, and it could not be reworked, and hence will be discarded from the 81 d t · ion line.

Resource Consumption Models (4.3)
Equations (4.2) and (4.3) are used to build the models for cost of quality, productivity, and cycle time for the four stage serial configuration system.

Cost of Quality Model
Since there are four processes in the considered production system , the expected scrap and rework could exist at any of these processes, and hence the cost of quality model is: (4.10) (4.12)

Productivity Model
The productivity of a system of four processes in series is simply the multiplication of the probabilities of acceptance at each state.

Productivity =
[(probability of accepting item at X 1 ) x (probability of accepting item at X2) x (Probability of accepting item at X 3 ) x (probability of accepting item at Xi)] Assume there is a production system with four serial machining processes (A, B, C, and D). Each process adds a new feature to the product. The mean, standard deviation, and the specification limits of the key quality characteristic at each process are summarized in table 4.1, where they are used to calculate the Probability of accepting, scrapping, and reworking parts at each process. The table also includes rework costs, scrap costs, processing times, and rework times at each process.

86
The assumptions are: _ The quality characteristics of all processes follow the normal distribution.
_ A part is considered to be acceptable if it falls within the specification limits, scrap if it falls below the lower specification limit and rework if it falls above the upper specification limit.
_ The sequence of the processes is: A, B, C, and then D.  Then: The total cycle time = 16.1 O minutes per part (by applying equation 4.14 ).

One Stage Parallel System with Four Processes
Consider four processes in parallel, where raw material enters each process which is capable of performing all required operations to produce a finished product. Parts from each process could be finished product if they meet the predetermined specifications, or can be reworked or scrapped if they do not meet specifications. The (/ _ Q) matrix becomes: 1-P., And hence: (l-P 44 ) (l-P44)

Resource Consumption Models
Equations (4.20) and (4.21) are used to build the resource consumption models for the parallel configuration system with four processes. The productivity of the production system that consists of four processes in parallel is the expected number of finished products divided by the total number of raw material parts that enter the system. Based on that: productivity = [(N1 x probability an item from process X' 1 is accepted) + (N 2 x probability an item from process X' 2 is accepted) + (N 3 x probability an item from process X'3 is accepted)+ (N 4 x probability an item from process X ' 4 is accepted)] I Therefore: (4.31)

Cycle Time Model
The cycle time is the summation of processing and reworking times at each of the four processes weighted to account for the proportion of raw material in each process path, therefore:  (2) Assume there is a production system with four parallel machining processes (A', B' , C', and D'). Each process can perform all production operations that are required to produce a finished product. Table 4.2 includes all required data, where the assumptions are: -The quality characteristics of all processes follow the normal distribution.
-The considered quality characteristics at A', B' , c·, and D' are the same as process D in the serial configuration system in example 1 and the specification limits are equal to those for process D.
-A part is considered to be acceptable if it falls within the specification limits, scrap if it falls below the lower specification limit and rework if it falls above the upper specification limit.
-A·, s·, c·, and o· each has the same mean, standard deviation, specification limits, processing and rework times, and rework and scrap costs.

Resource Consumption Models
Equations (4.35) and (4.36) are used to build the resource consumption models for a serial-parallel production system with four processes when there is no crossover.

Cost of Quality Model
Each of the four processes at stages one and two contributes to the total quality cost per part: Where: ( 4.41) (N, +N 2 ) ) N,b, 5 SC, (4.45) (N, +N 2 )

Productivity Model
The productivity is the ratio between the finished products delivered from processes X\ and X" 4 and the total raw material parts that enter the two parallel lines, therefore: 100 productivity= [(N1 x probability an item is accepted at x"1 and X"3) + (N 2 x probability an item is accepted at X"2 and X"4)] I (N1 + Nz) Then: (4.46)

Cycle Time Model
For the parallel configuration without crossover, the expected cycle time becomes: (4.47) Where: (4.48)

Numerical Example (3)
Assume that there is a production system with four machining processes (A", B", C", and D\ Processes A" and c" are parallel to processes s" and o".
Each pair of processes can perform all production operations that are required to produce a finished product. Table 4.3 includes all required data, where the assumptions are: _ The quality characteristics of all processes follow the normal distribution.
_ The considered quality characteristics at A" and 8 11 are the same as in process B in the serial configuration system in example 1 (the specification limits are equal to those for process B). On the other hand, the considered quality characteristics at c" and o", are the same as in process D in example 1.
-A part is considered to be acceptable if it falls within the specification limits, scrap if it falls below the lower specification limit and rework if it falls above the upper specification limit.
-Processes A 11 and 8 11 have the same mean, standard deviation, specification limits, processing and rework times, and rework and scrap costs. The process time, rework time, rework cost, and scrap cost for each process equals, respectively, the summation of process time, rework time, rework cost, and scrap cost of that for processes A and B in example 1.
-Processes c" and o" have the same mean, standard deviation, specification limits, processing and rework times, and rework and scrap costs. The process time, rework time, rework cost, and scrap cost for each process equals, respectively, the summation of process time, rework time, rework cost, and scrap cost of that for processes C and D in example 1 . ,>: RT (min/item) 5 5 6.5 6.5

5000
The P, A, and B matrices are: (4.51)

Resource Consumption Models
Equations (4.50) and (4.51) are used to build the models of quality cost, productivity, and cycle time for the serial-parallel production system with four processes when there is crossover.

Cost of Quality Model
Since the processes at stage two, X" 3 and X\, could receive parts from either process x" 1 or x" 2 , weights are given to identify the proportion of parts which enter X"3 and X" 4 . "wi/' term is used to represents the weights, where: 107 .· the percentage of accepted parts from process x"i that go to process x"j· W1J· For this type of system configuration: ( System productivity is the ratio between the finished products delivered from processes X"3 and X\ and the total raw material parts that enter the two parallel lines. When considering the weights given to identify proportions of parts that enter x"3 and x"4 from x "1 and x"2, the system productivity becomes: The assumptions in this example are similar to that in example 3, except that Parts from either process in stage one can go to any process at stage two. 109 75 % of parts that are accepted at process A" proceed to c" and 25% proceed to DA, whereas 50% of parts that are accepted at process 8 11 proceed to c" (assuming c" has higher capacity than 0 11 ) and 50% proceed to o". Table 4.3 includes all required data for this example. 2.5 2.5 3.5 3.5 r·SC 1$/item_l 5.2 5.2 7.7 7.7 t·PTJmin/iteml 7 7 9 9 RT Imin/itemI 5 5 6.5 6.5 Number of entered RM 5000 5000 The P, A, and B matrices are:   Based on the data and assumptions, serial-parallel with crossover and serial-parallel without crossover configurations have same quality cost, productivity, and total cycle time. Their total quality costs are equal, although their f(RC3) and f(RC 4 ) values are not the same. The quality cost, productivity, and cycle time equations for the system with crossover are not the same as that for the system without crossover (as shown in section 4.5 and 4.6 respectively), Which in turn indicates that the result in this example is just a special case where the two configurations perform the same due to the assumptions that the Processes at each stage are similar in terms of the parameters used in the three models 112 In terms of the quality cost, the parallel configuration has the lowest cost, followed by the serial-parallel configurations, and then the serial configuration. In terms of productivity, the parallel configuration has the highest value, followed by the serial-parallel configurations, and then the serial configuration. In terms of total cycle time, the serial configuration has the lowest time, followed by the serialparallel configurations, and then the parallel configuration.
In summary, the parallel configuration of the assumed production system has the best performance in terms of quality cost and productivity, followed by the serial-parallel configurations and then the serial configuration. On the other hand, the serial configuration has the best performance in terms of total cycle time followed by the serial-parallel configurations and then the parallel configuration.

Abstract
In chapters three and four, the resource consumption models are developed for two and four process production systems respectively, while in this chapter the work is extended to build a general model for n process production systems for each of the serial, parallel, and serial-parallel without crossover configurations.

Introduction about Building the n Model
It can be concluded from the models for two and four process production systems (in chapters three and four) that there are only two terms needed to develop the resource consumption models: 1-aii from matrix A, where i= 1,2, .. , n, and n =number of the processes.
2-bi(n+1) from matrix B, which is the long term probability that an item from state i will end up as scrap, where state (n+1) is the scrap state and state (n+2) is the finished goods state. For example, when the production system consists of two processes (n=2), then b 13 is the long term probability that an item from process 1 (state 1) ends as scrap.

114
Mathematica® software is used to generate the model for different values of n, and then the trend for both of aii and bi(n+1) is analyzed to generate a general expression. For the definitions of the notations in next sections, refer to section 4 .2 in chapter four. Figure 5.1 represents a production system with n processes arranged in series. The general expression for aii is:

117
Therefore, by applying equations 5.3, 5.4, and 5.5 respectively the quality cost, productivity, and cycle time can be found:  The general expression for aii, bi(n+1), and the resource consumption models are: For example, assume there is a production system with 3 processes, and the transition probability matrix is:  i=t 1 2 h=~1 z + t (5.14)

Parallel Production System with n Processes
As an example, suppose that there are five machines arranged in two parallel lines with three processes in the first line and two processes in the second line. Table 5.1 shows the probabilities of scrap, rework, and acceptance for each machine. It also includes rework and scrap costs, and processing and reworking times.  The n process serial-parallel configuration with crossover involves a complicated model for the long term scrap probabilities bi(n+1) as well as more complicated resource consumption models. However, there is an example of four process serial-parallel configuration with crossover in section 4.6, and an example of six process serial-parallel configuration with crossover in section 6.5.2, which can be used for similar production systems in each case. The level of complexity and difficulty to follow a pattern in the probability matrices and the resource consumption models is too high and it is advised that the reader uses mathematical software packages to help develop the model for configurations with a large number of processes.
CHAPTER SIX CASE STUDY AT BIOPHARMACEUTICAL COMPANY

Abstract
The methodologies and models in this thesis are applied in a biopharmaceutical industry case study. Two production systems from a biopharmaceutical company have been utilized to apply the serial-parallel model (with and without crossover), and the serial model. This chapter includes a brief background about the company and the description and analysis of the two production systems.

Company Background
XY Company is a fast growing biopharmaceutical company that develops and produces drug products. It has many locations outside and inside the United State. It has short and long term objectives that aim to satisfy their customer with a high quality product that meets the Good Manufacturing Practices (GMPs), established by the Food and Drug Administration (FDA), in all aspects, and hence it maintains the safety, identity, strength, quality, and purity of its products.
As in many biopharmaceutical companies, the main processes for

Media Preparation Production System
Media is defined as: "a (usually sterile) preparation made for the growth, storage, maintenance, or transport of microorganisms or other cells" (Odum, 2002). Different types of biopharmaceutical products need different types of media, and for each product there are two types of media, growth and production media. Figures 6.2 and 6.3 show the value stream map (VSM) for producing 750 liters of media starting from the point the raw material is received to the point where finished media is ready to be transferred to the internal customer (cell culture department). The media production system consists of two preparation tanks, two hold tanks, and transfer lines as show in figure 6.4. Media is prepared at tanks A and B, where the components are added to WFI (Water for Injection) followed by titrating the solution to adjust the pH value. The holding tanks (A' and B') are used to keep the media under pressure after going through the filters at the transferring lines. Filters should pass two tests, pre-test and post-test, to ensure 129 that they were in good condition and did not cause any contamination of the media. To meet the GMP requirements and avoid cross contamination, GIP (Clean in Place) and SIP (Steam in Place) processes are applied to clean the tanks and the transfer lines before producing the media.
As can be seen from figures 6.2 and 6.3, the VSM tracks all activities occurring at preparation tanks A and B, and hold tanks A' and B'. Both figures show the percentage of value added time at each tank. For example, hold tank A' has 65% value added time whereas hold tank B' has 5% value added time, which is the lowest percent. The two most common non-value adding activities are different types of waiting and paperwork.

Media Production System Configuration
Two types of media, Land M, are used in this case study, where both are used for the same drug product. As shown in figure 6.4, the production system of media L or M consists of three main production stages. The first stage is preparing media at tank A or B by adding chemicals and serum to WFI. The second stage is titrating the media at tank A or B, and the third stage is filtering and transferring the media to holding tank A' or B' . The system configuration is considered to be serial-parallel, where there are two parallel lines, and each line consists of three processes. Crossover may also happen between the processes at stages two and three. That means the media after titration could be transferred to either of A' or B' as shown in figure 6.4.   Two configuration alternatives are analyzed to find the performance of each alternative in terms of quality cost, productivity, and cycle time. These configuration alternatives are, serial-parallel without crossover and serial-parallel with crossover. In the case of this specific production system it is not appropriate to arrange the six processes into a pure parallel configuration, because this would violate the requirement of filtering the media before it enters the hold tanks and is ready for use by the cell culture department. In other words, filtering the media should not be performed in the same processing stage where it is prepared. Arranging them in a pure serial configuration is not applicable as well.
For example, if the two processes in the preparing media stage were arranged in series where adding water is performed at tank A and adding chemicals is Performed at tank B, then that will be not feasible since water has to be transferred again from tank A to tank B in order to be able to add chemicals, and hence filling water into tank A becomes meaningless. That means that preparing and transferring the media can only be performed by three stages, not one stage (as in a pure parallel configuration) or six stages (as in a pure serial configuration). The models developed in chapter five for the "three stage serialparallel system with six processes" can be applied in this case study.

Resource Consumption Models at the Media Production System
In this section the resource consumption models are applied to the two alternative configurations, and then the results are used to analyze their performance.

Serial-Parallel Configuration (without Crossover)
The required data to build the resource consumption models have been collected, with the exception of data that are related to costs because this is not available to be included in the thesis. As a starting point, the types of media that are used in this case study are identified . After that, the key quality characteristic for each process is defined, and then the short term probabilities of acceptance, rework, and scrap are determined.  (1) and (4)), the operator 134 should make sure that the total amount of WFI is within the acceptance range after adding the chemicals. OIT (Operation Interface Terminal) technology is used to control the amount of added WFI before and after adding the chemicals, so at this process all of the prepared media is accepted, since it always falls within the pre-specified range, and moved to the titration stage.

Quality Characteristics and Short Term Probabilities
At titration, the considered quality characteristic is the pH value. The prepared media can not proceed to the next stage unless the pH value at titration falls within a given acceptance range. The solution is re-titrated until pH falls within this range; hence, there is no scrap from this stage. Once the media is accepted at titration it is then filtered and transferred to the hold tanks.
There is no rework at stage three (filtering and transferring media). The amount of scrapped media at stage three could range from a few liters up to all of the prepared media. If the filters do not pass the post-test, the media is considered to be non-qualified media and is all scrapped. Although losing the entire amount of media rarely happens, there are always small amounts lost during the filtering and transferring process.
Historical data for the pH values at the titration stage are collected in order to find the rework rate. Minitab software is used to find the probability distribution that the data follows. Data for media L shows that the pH value of the titrated media from tank A (state 2) fits the normal and lognormal distributions, whereas the pH value of the titrated media from tank 8 (state 5) fits the lognormal distribution (figures 6.5 and 6.6 show part of the Minitab outputs for finding the 135 best distribution. The figures of other distributions are attached in the appendix).   For media M, the titrated media from tank A (figure 6.9) fits the largest extreme value distributions, while the titrated media from tank B does not fit any distribution, so Johnson transformation is applied to the data (figure 6.10) and the resulting P-Value is accepted (the figures for other distributions are attached in the appendix). Figure 6.11 shows the pH value histogram for the titrated media from tank A, while figure 6.12 shows the histogram of the transformed data of titrated media from tank B.   After finding the distributions for the pH data, the probability of rework at each of the titration processes for both types of media is calculated, and summarized in Table 6.1. The probability of rework is the probability that the pH value will be above the upper specification limit (pH = 6.8) or below the lower specification limit (pH = 6.6). Each campaign of the selected product includes, on average, seven lots of type L and fifteen lots of type M. The rework time and processing time for the titration process are almost equal. Since cost data is not available, the rework cost at any process is measured by the labor time in minutes at that process. On the other hand, the scrap cost is measured by the labor time in minutes consumed for producing the media starting from receiving the raw material to the point where the scrap happens. Tables 6.2 and 6.3 summarize all of the data that is required to apply the resource consumption models for the production system of media L and M.  The procedure used in previous chapters is used to find the cost of quality, Productivity, and cycle time models.

Probability Plot for pH_ Titration -A
(Refer to the definitions of aii, and bii from section 4.2).

• Productivity Model:
3. Cycle Time Model: Where: CT; =PT; +(a;; -1 )RT; Table 6.4 summarizes the calculated quality cost, productivity, and cycle time after applying the equations. Prepared at tank B -titrated -filtered and transferred to Tank B'

Quality Characteristics and Short Term Probabilities
The considered quality characteristics are the same as in section 6.5.1 when crossover is not allowed between stage three and stage four. The probability of re-titrating the prepared media does not change (table 6.1 ). The only change is the scrap rates during the filtering and transferring process. Table   6.5 shows the percentages of titrated media from Tanks A and B that is filtered and transferred to the hold tanks A' and B'.

Cell Culture Production System
Odum, 2002 defined cell culture as: "Cells taken from a living organism and grown under controlled conditions ("in vitro"). Method used to maintain cell lines or strains". There are four main stages which happen during the cell culture production system: initiation, cell expansion , refeed, and harvest. A brief description of the production processes for two types of drug products is shown in section 6.6.1.

Process Flow for Cell Culture Production System
The production for product 1 starts with the initiation process followed by four cell expansions (CE), one refeed, and finally fourteen harvest processes (HA), with an average processing time of 76 days, as shown in figure 6.16. On the other hand, the production for product 2 includes one initiation process, three cell expansions (CE), one refeed, and three harvest processes (HA), with an average processing time of 56 days, as shown in figure 6.17. During the initiation and cell expansion processes the growth media is used, where during the refeed and harvests processes the production media is used instead. Each campaign of product 1 includes six lots; three as main lots and three as back-up lots that proceed only to cell expansion three. Each campaign of product 2 includes seven lots; five as main lots and two as back-up lots that proceed only to cell expansion two. The back-up lots are only used if any of the main lots are lost because of a bioburden or contamination problem.

Initiation and Cell Expansion Processes
The  Reviewing the historical data shows that using the back-up lots rarely happens, and most of the time the back-up lots end up discarded. The total number of seeded bottles for each back-up lot is 476 RB for product 1 (from initiation, CE1, CE2, and CE 3 ) and 226 RB for product 2 (from initiation, CE1, and CE2). Since each campaign of product 1 has three back-up lots, the total quantity of seeded bottles (which will be mostly discarded) becomes 1,428 RB, and since each campaign of product 2 has 2 back-up lots, the total quantity of seeded bottles becomes 452 RB.

154
Through an online search for the market price of the empty RB, it can be assumed that its price from the supplier equals $5 (http://www.coleparmer.com).
Then, the total cost of empty RB that are used in the back-up lots becomes $7,140 per single campaign of product 1, and $2,260 per campaign of product 2.
By adding the costs of labor hours, overhead, and other raw material, the total cost becomes even larger. This is an example that shows the type of challenges and difficulties that the biopharmaceutical industry faces to be competitive and meet the market needs. Back-up lots that cost a lot of money are produced just to make sure that the company will not risk a shortage of critical product available in the market for their patients which in turns means loosing the customer satisfaction, not helping the patients, and loosing millions of dollars. Tables 6.7 and 6.8 include the total number of RB per campaign for product 1 and product 2.
The percentages of discarded RB per campaign at each initiation and cell I expansion processes are now found. The discarded RB are the unused seeded RB from main lots and all RB from the back-up lots. Table 6.9 includes the total non-discarded and discarded RB for each process. Each of the cell expansion processes includes two steps, trypsinization and inoculation. The final expansion before the refeed is performed by robots, whereas all of the previous cell expansions are performed manually. When the robots perform the trypsinization, some RB might be discarded if any problem happens during this semi-automated process. There is no available historical data regarding how many RB each robot might loose during the process, but based on two days of personal observations the percentage of the discarded bottles from all robots during the last cell expansion averages 1.57% for product 1 and 6.14% for product 2. Assuming that these observations are representative of typical production, and based on figures 6.15 and 6.16, the percentages of the discarded RB for each of the main lots at each of the initiation and cell expansion processes can be found . Table 6.10 summarizes these percentages for product 1 and product 2.

Refeed and Harvest Processes
The refeed and harvest processes follow the cell expansion processes.
The refeed process involves replacing the media in the RB and returning the RB back to the storage room to continue growing the cell. The harvest processes include collecting product from the RB, adding new media, and then returning them back to the storage room. The process flow is shown in detail in figure 6.20.
During the refeed and harvest processes the roller bottles (RB) are visually checked to determine if any is contaminated. An RB is discarded if its contents are contaminated or damaged by an operator or robot, or if the bottle itself is damaged. Two years of historical data for the number of discarded bottles at refeed and harvest were used to calculate the average percentage of discarded RB for each process. Since the discarded RB does not add value to the finished product, they will be considered to be waste or scrap. Table 6.11 shows the percentage of discarded RB (scrap rate) for refeed and all harvest processes of product 1 and product 2.
The data describing the number of discarded RB at all processes of cell culture (initiation , cell expansions, refeed, and harvests) are used to develop a transition probability matrix for the serial configuration of the entire cell culture production system. Then, the quality cost, productivity, and the cycle time can be found.

Resource Consumption Models for the Cell Culture System
The initiation and cell expansion processes (except the last cell expansion) are not considered in the P matrix in developing the Markova chain, but they will be considered in determining the quality cost, productivity, and cycle time. That is because at each of these processes the scrap rate is based on a constant quantity that is needed to be achieved each time and it could not be variable with time, so it will not be changed in long run. On the other hand, scrap at the last cell expansion, refeed , and harvest processes is caused by robot, operator, or contamination, so it may vary with time. The transition probability matrix is developed based on data shown in tables 6.10 and 6.11. Since any RB with any quality problem will be discarded to maintain the highest possible production quality, there is no rework in the cell culture production system. to complete last 1 ' : rac:k.' 3 : 1 _ : :-WaTt-olher io-bOftD -:  (3J There 3re ::ooo RB (20 racks) and 3 robot:: one robot 1 .s userJ for 6 r:;tcl::: of RB, -.,.yt1ere the ott"1er t<M) robot-s are t .J::.:ed tor 7 tocks e~cf1 (4J Before tr:lns.fenirig. welt 15 min bi:.> tote t~k1ng CJ:il mple 1 ~nd -. , .. ; art 5 min to confirm from UF Tn:;nsdenlng takes· '30 min on ~verage .0007 .0010 .0021 .000 8 .0014 .0006 .0009 .0011 .0010 .0008 .0008 A modified model for the n stage serial configuration is used by considering the short term probabilities for the initiation and cell expansion processes (except the final cell expansion) and the long term probabilities for the 161 rest of the processes obtained from the Markov chain model. Table 6.12 summarizes the results of applying the resource consumption models for the serial configuration cell culture system.

7 Cell Expansion 3 for Product 2
In this section another case study is developed for the final cell expansion process in producing product 2. The final cell expansion process is performed by three robots, and it includes trypsinization and then an inoculation process. Three robots are used to perform the cell expansion, robots X, Y and Z. Based on observations, the percentage of wasted RB during the trypsinization step is assumed to be 10. 71 % at robot X, and 3.8% at robots Y and Z. On the other hand, the scrap rate during the inoculation process is assumed to be 0.0025% for robot X, while it is 0% for robots Y and Z. Robot X has some operation problems caused by different factors such as the sensor, robot arm, or keyboard.
During the trypsinization process the cell suspension is prepared in a spinner, and then at the inoculation process the spinner content is distributed to empty bottles in order to expand the cells. Table 6.13 includes the average time required to perform the trypsinization and inoculation processes when they are completed at the same robot. Different alternative configurations are considered, and then the resource consumption models are applied. Since Robots Y and Z are similar in terms of the scrap rate, the number of alternatives is reduced to eight, as shown in figure   6.21 . The current configuration is a pure parallel, by which both trypsinization and inoculation are performed at the same robot. On the other hand, the other alternative configurations have a combination where the two processes could be performed at the same robot or at two different robots. If the inoculation is performed at different robot from the one that performs the trypsinization, then the average time for inoculation is assumed to be increased by 10 minutes for material handling and setup.
( 1) (3) To develop the results for the eight configurations the model from chapter three for the pure serial and serial-parallel configurations is used assuming that the appropriate process has zero parameters. For example, in configuration #5, a hypothetical process of zero parameters is assumed to be in series with process z. Table 6.14, and figures 6.22, 6.23, 6.24 summarize the results for all configurations.

Case Study One: Media Production System
The production system for two different types of media is analyzed, where there are two configuration alternatives: serial-parallel with crossover and serialparallel without crossover. All data required to model each configuration is collected, and then the productivity, quality cost, and cycle time are determined for each alternative.
The results for media L show that the quality cost is 19.81 (labor minute per unit of media) for the serial-parallel configuration without crossover, and 23.11 (labor minute per unit of media) for the serial-parallel configuration with crossover. The total expected time to produce one unit of media for both configurations is 259.87 minutes, so this indicates that 19.81 minutes is 7.62% and 23.11 minutes is 8.89% of the total time that is required to produce one unit of media. In other words, 7.62% and 8.89% of the total time required to produce 1 unit of media is a waste because of the scrap and rework. Converting labor minutes to labor costs of two operators, and then adding the material and overhead costs, will indicate how much money is lost due to quality costs. In terms of productivity, the configuration without crossover has 77.42% productivity, whereas the configuration with crossover has 73.33%.
The results for media M show that the quality cost per unit of media is 7.26 (labor minutes) when the configuration is without crossover, and 7.04 (labor minutes) when there is crossover, which are, respectively, 2.80% and 2.72% of cycle time (259.22 minutes). When the configuration is without crossover the productivity is 93.93%, while it is 94.20% when the configuration is with crossover.
As a conclusion, the cost of quality and productivity for the media M production system is better than that for media L, since the rework and scrap rates are different as shown in Tables 6.2 and 6.3. The without-crossover configuration of media L has a lower cost of quality and higher productivity than a configuration with crossover, while the expected cycle time is the same for both configurations. On the other hand, for the media M production system, a configuration with crossover performs slightly better than a configuration without crossover by generating less cost of quality and higher productivity. The expected cycle time, however, is the same for with and without crossover configurations of media M.

Case Study Two: Cell Culture Production System
The cell culture production system has a serial configuration and this is the only applicable configuration, because of the nature of the product, where cells should stay for a certain amount of time after each process to keep growing before proceeding to next process. The developed resource consumption models for serial configuration still can be used to find the quality cost, productivity and cycle time for product 1 and product 2 production systems.
For product 1 the scrap rates at cell expansions 2 and 3 are the highest and they affect the quality cost and productivity more than the remaining processes. The cycle time equals 1824 hours (76 days), which is just the summation of the processing time at each process since there is no rework time.
The quality cost is 105.72 hours per lot (approximately 4 day and 10 hours per lot), which is 5.8 % of the total time. The productivity is only 24.84% and that is obviously because of the high scrap rate at many processes.
For product 2 the Initiation, cell expansions 1, and cell expansion 2 processes have the highest scrap rate, so they have greatest effect on quality cost and productivity more than the remaining processes. The quality cost is 99 .95 hours per lot (approximately 4 days and 4 hours per lot), which is 7.44 % of the total time. The productivity is only 40.51 % which is better than that for product 1 but still not high. There is no rework in producing product 2, so the cycle time equals the summation of the processing time at each process which is 1344 hours (56 days). The company should consider improving the processes to reduce the scrap rate especially for processes that have the highest rates, in order to reduce the quality cost and increase productivity.

Case Study Three: Cell Expansion
Different alternative configurations are analyzed to find which one performs better. Figures 6.22 shows that configuration 4 is the best in term of quality cost since it has the lowest quality cost, while configuration 7 is the worst since it has the highest quality cost. Figure 6.23 shows that configurations 4 and 6 have the highest productivity and hence they perform best, while configuration 3 has the lowest productivity so it is the worst one. Figure 6.24 shows that configuration 8 has the lowest cycle time so it is the best, while configurations 169 1,2, 3, and 4 have the highest cycle time so they are the worst. Table 6.15 includes the ranks of the eight configurations from best to worst. It can be concluded from table 6.15 that although performance metrics may involve trade off, it is possible to achieve multiple goals at the same time. For example, configuration 4 performs the best in term of two metrics: quality cost and productivity. Configuration 8, which is the current configuration, has the best performance only in term of cycle time. Selecting the best configuration depends on the metrics that the management wants to consider in improving the production system performance. The managers also could give weights to the performance measures depending on the importance of each measure. For example, if the weights for the quality cost, productivity, and cycle time are 0.4, 0.4, and 0.2 respectively, then table 6.16 shows that in terms of all metrics, configuration 4 performs the best while configuration 7 performs the worst. The 170 applicability of each alternative should also be considered before selecting the best one.

DISCUSSION AND CONCLUSION
This chapter provides a summary of the results obtained , with a comprehensive discussion . The conclusions are then proposed .

Summary and Discussion
A summary of the numerical example results for the two process production system (using a numerical example as shown in sections 3.3.4 and 3.4.4) is shown in table 7.1. The results show that the selection of the better alternative depends on the performance measure of interest. For example, if this manufacturer is interested in reducing the cost of quality, then the parallel configuration performs better than the serial, or in other words, changing the production configuration from serial to parallel can save 24.2% of the quality cost. Such an improvement is also combined with an increase in the productivity of the production system from 79.03% (serial) to 93.16% (parallel). However, the cycle time slightly increases using the parallel configuration . The manufacturer can make a decision based on the performance target of interest.
Next, a full summary of the sensitivity analysis for the two process numerical example is shown in tables 7.2, 7.3, and 7.4. Note that the results apply for the ranges of parameters selected for the sensitivity analysis. Table 7.2 indicates that the parallel configuration performs better than the serial configuration when rework costs change, while the preferred configuration depends on the value of the scrap costs. Table 7.3 indicates that the parallel configuration performs better than the serial configuration when means and standard deviations change. Table 7.4 shows that the serial configuration performs slightly better than the parallel configuration when the processing times change, while the preferred configuration depends on the value of the rework times of processes 1 and 2.  Parallel has higher productivity than the serial

Mean2
Parallel has higher productivity than the serial

StDev1
Parallel has higher productivity than the serial

StDev2
Parallel has higher productivity than the serial Table 7.4: Sensitivity analysis summary for cycle time of the two process d f t I Q_rO uc ion s_ys em exam_Q_!e Parameters Results summary PT1 Serial has slightly lower cycle time than the parallel PT2 Serial has slightly lower cycle time than the parallel RT1 Serial has lower cycle time than the parallel at low RT1 RT2 Serial has lower cycle time than the parallel at low RT 2 With a two process production system, there are only two possible configurations to be applied, pure serial and parallel. In the following discussion, a four process production system is analyzed, and the four possible configurations are implemented, which are pure serial and pure parallel, and serial-parallel with and without crossover. Table 7.5 summarizes the results of the resource consumption models using a numerical example as in sections 4.3.3, 4.4.3, 4.5.3 , and 4.6.3. Similarly, the selection of any given configuration depends on the performance target assigned by the manufacturer. The parallel configuration shows the highest productivity and the lowest cost of quality compared to the other alternatives, while the serial process flow has the lowest cycle time.
In table 7.5, the results show that the serial-parallel configurations (with and without crossover) performed the same, based on the assumptions and parameters used for that example. However, when applying the serial-parallel models (with and without crossover) on the case study analysis at a biopharmaceutical company in Chapter 6, results were different between the two alternatives. The parameters used to develop the case study were based on actual production data, where results were dependent on the operating parameters, assumptions, and production constraints. A production line for two media products containing six processes was analyzed , utilizing serial-parallel configurations (with and without crossover). Referring to section 6.5 , table 7.6 summarizes these results. Note that media L and M are two different products that are manufactured on the same production system .  Table 7.6 shows that the results for configurations with and without crossover are different. In terms of cost of quality and productivity, the configuration without crossover for media L performs better, while the configuration with crossover for media M performs better. However, the cycle time remains the same between the two configurations for each media type, because the processing time and rework parameters -including the rework time and probabilities -stays the same when changing configurations.
In another case study implemented at the same biopharmaceutical company, three robots within cell expansion 3 are arranged into all possible configurations of pure parallel and serial-parallel configurations. Note that the model was adjusted so that it can be used for configurations of any number of processes, odd or even, by making the appropriate assumptions and adding virtual processes. Based on the performance target, the eight different configurations were ranked to achieve the best production output.

.2 Conclusions
This thesis provides an analytical approach to model a production system and select the preferred configuration . There is a relationship between the configuration of any production line and the performance. Based on the production and operating parameters, the preferred configuration could vary accordingly. Cost of quality, productivity, and cycle time are three important metrics in measuring the system performance. They are affected by the system configuration and the operating parameters . Besides the transition probabilities between the states of the production system , the main parameters that affect the quality cost are rework and scrap costs, wh ile the main parameters for cycle time are the processing time and rework time. An n process general model can be applied for odd or even numbers of processes in any production line, with direct calculations for the resource consumption models which provides great benefits for analyzing large production systems. This thesis provides an excellent reference for the biopharmaceutical industry, with models that are applicable for the performance analysis of both continuous and discrete production systems.

.3 Future Work
Based on the results obtained, the following are suggested as possible extensions to the research in this thesis: 1-Use a simulation tool to model the production flows of the four different system configurations and then to verify the proposed resource consumption models.