Workshop on
Performance analysis of manufacturing
systems:
Bridging the gap between industry and academia?
June 19 & 20, 2006
EURANDOM, Eindhoven, The Netherlands
Adan, Ivo (TU/e, EURANDOM)
Mean value analysis for polling systems
Polling systems typically consist of a number of queues attended by a single server in a fixed cyclic order. They have a wide range of applications in manufacturing, communication, transportation and maintenance systems. This talk is concerned with two commonly used service disciplines, the so-called exhaustive and gated discipline. Exhaustive means that a queue must be empty before the server moves on to the next one, and gated means that only those customers are served present at the polling start. In many applications the mean delay of a customer is an important performance characteristic. Unfortunately, only in very special cases, explicit closed-form expressions are known for the mean delay. In this talk we present a novel mean value approach to compute the mean delay figures, the key components of which are Little's law and the PASTA property. The merits of the mean value approach are in its intrinsic simplicity and its intuitive appealing derivation.
Armbruster, Dieter (Arizona State University & Eindhoven University of Technology)
Continuum models and their validation for semiconductor production
A review of continuum models for production flows involving a large number of items and a large number of production stages is presented. The basic heuristic model is based on mass conservation and state equations for the relationship between the cycle time and the amount of work in progress in a factory. Heuristic extensions lead to advection diffusion equations and to capacity limited fluxes. Comparisons between discrete event simulations, factory data, numerical solutions of the heuristic PDEs and queuing network models are made. A first principle model based on the Boltzman equation for a probability density of a production lot, evolving in time and production stages is shown to lead to an advection diffusion equation.
De Kok, Ton (TU/e)
Modelling manufacturing systems
In this presentation we discuss a quantitative model that has been derived on the basis of discussions with experienced industrial engineers. The quantitative model translates process information into a queueing model, which is a submodel of a queueing network model. The quantitative model enables to model a variety of processes, including flowlines, ovens and machines processing similar products simultaneously. We discuss various concepts that enable modelling, including set-up behaviour and break-down behaviour. We briefly discuss some case studies and the analysis techniques.
Etman, Pascal (TU/e)
STW-project Effective Process Time: an overview
For a successful improvement of throughput and flow time performance, insight in the factors that are responsible for capacity losses in the manufacturing system is essential. Even small reductions in capacity loss may yield significant financial benefits or savings. For this reason, industry puts great efforts in the reduction of capacity losses due to disturbances such as machine downs, setup, rework, etcetera, for instance by using metrics such as the OEE. In addition to capacity loss, the various disturbances in the manufacturing system cause variability in processing. A high level of variability also adversely affects the throughput/flow time performance. A metric quantifying the effective level of variability at workstations is therefore desirable.
Hopp and Spearman (1996, 2000) introduced the term 'effective process time' and defined it as the process time seen by a lot from a logistical point of view. The effective process time (EPT) aggregates the raw process time and all time losses due to events such as machine downs, operator delays, and other disturbances into a single distribution. The advantage of this approach is that both the mean effective process time (which relates to capacity loss) as well as its variance (variability) can be determined, and that these two quantities can be directly used in a queueing theoretic framework. Hopp and Spearman compute estimates of the EPT mean and the EPT variance from the contribution of the individual sources of variability. In practice this may turn out to be difficult, in particular when only part of the disturbances are known or being measured.
In the STW research project, we start from the idea to measure the aggregate process time distributions from basic factory floor data such as arrival and departure events. For simple flow lines consisting of single-lot machines this boils down to an application of the sample-path equations. We are generalizing this idea towards aggregate modeling and parameter estimation of more complicated machines, workstations and networks. Our aim is to arrive at an aggregate modeling methodology that enables us to build simple yet accurate models of manufacturing networks using operational factory data without the need to characterize all contributing disturbances and shopfloor realities. Key elements in our approach are (i) the effective process time paradigm in aggregate modeling and parameter identification, and (ii) efficient queueing network approximations through iterative decomposition-based algorithms.
In the presentation we give an overview of the project, present some of the main findings, and discuss possible future directions.
Gershwin, Stan (MIT)
Manufacturing Systems Design and Analysis --- Past Successes and Future Research
Every time a new product is introduced, a manufacturing system must be built or modified. When product lifetimes were long, there was plenty of time to build the system and then gradually improve it. Now, however, many product lifetimes are short, so systems must be built quickly and operate well immediately.
Therefore, society needs (1) accurate, reliable, and fast tools to predict the performance of proposed factory designs; (2) rules or tools to suggest good designs; and (3) tools for real-time management (control) of factories. These tools should not be used as black boxes; they should be used by Manufacturing Systems Engineering professionals who have an accurate, quantitative, intuitive understanding of production systems.
This talk discusses large stochastic flow systems that arise in manufacturing and other contexts. One of the crucial characteristics of these systems is temporary storage: material may rest in in-process buffers while its immediate next destination is unavailable. Such buffers increase throughput, but at the expense of increasing inventory.
A natural approach to the analysis of these systems is to formulate them as Markov processes, and obtain performance measures (production rate, average inventory, etc.) from their steady-state probability distributions. However, because the storage capacities are finite, there is no exact analytical solution for these systems. Because the systems are large, their state spaces are enormous.
A practical, effective approximation is decomposition. The system is broken into small subsystems, and the influence of the outside world on each subsystem, and of each subsystem on the outside world, is modeled in a simplified way. All the subsystems are modified iteratively so that their different representations of each other are reconciled.
The development of this approach and its application to widening classes of systems is described. We emphasize our current work on the interaction of quality- and quantity-focused strategies on the quality and quantity performance of manufacturing systems. We show how inventory capacity can influence system yield and productivity, sometimes in counter-intuitive ways.
Kock, Ad (TU/e)
Lumped parameter modelling of the litho cell
Litho cells, are the most expensive equipment in a wafer fab. To support decision-making on this equipment, accurate simulationmodels for throughput and flow time are helpful. The simulation models that are typically developed incorporate various shop floor details. To properly model these details, they should be quantified, which is difficult and time-consuming. In this paper, a lumped parameter model is proposed for the litho cell. The model consists of two parts: a detailed representation of the processing inside the track and scanner, and an aggregate representation of the factory floor feeding the loadport. The track-scanner is modelled as as a tandem flow line with blocking. The shop floor is represented by a delay distribution that incorporates all contributions outside the machine. Preliminary simulation results show that the suggested method provides a simple, yet accurate approximation of the litho cell.
Lefeber, Erjen (TU/e)
Modeling and Control of Manufacturing Systems
This contribution deals with the modelling of manufacturing systems for control. First the concept of effective process times is introduced as a means to arrive at relatively simple discrete event models of manufacturing systems based on measured data. Secondly, a control framework is presented. Thirdly, this framework is implemented on a case, which shows that suitable EPT-realisations can only be obtained when job authorisation events are recorded.
Nijsse, Frank (VDL Steelweld BV)
EPT throughput control in the automotive industry
The core competence of VDL Steelweld is to develop, produce and build production lines for the automotive industry. In the year 2001 VDL Steelweld went to the Technical University of Eindhoven with throughput simulation questions. Our main issue was the collection and simulation of the MTBF and MTTR. This is the current input for discrete event simulation of flow lines with blocking in the automotive industry. In cooperation with the TUE we developed a total new way to control the throughput in the automotive industry based on the Effective Process Time. This presentation gives an explanation on one of our practical applications of the EPT at the car factory of Volvo Cars in Gent.
Resing, Simone (CQM B.V)
Performance analysis of manufacturing systems: a practitioner's point of view
CQM is a consultancy firm that supports its customers in decision making and in process improvement. This is done fact based, with analytical techniques and mathematical models. One of our consultancy areas is the logistical design and performance analysis of manufacturing systems. In this talk we present our view on EPT, and show some recent projects.
Rose, Oliver (Dresden University of Technology)
Benefits and drawbacks of simple models for complex production systems
Semiconductor wafer fabrication facilities are among the most complex production facilities. A large product variety, hundreds of processing steps per product, hundreds of machines of different types, and automated transport lead to a system complexity which is hard to understand and hard to handle. For teaching planners and developing adequate material flow control mechanisms, simple models for this complex environment are required. We outline two simplifying approaches, one for fostering the understanding of the factory and one for the development of a control rule for a complex type of machines. In the first example we show how a simple model can be used to predict the factory behavior after a bottleneck work center breakdown. In the second example, we discuss an approach how a few simple characteristics of a cluster tool can be used to run a fast sequencing algorithm for this tool type. In both cases, we present the practical benefits and the drawbacks of very simple modeling approaches.
Scholz-Reiter, Bernd (University of Bremen)
The influence of production networks’ complexity on the performance of autonomous control methods
Autonomous control of production networks means the decentralised decision-making and routing of autonomously acting logistic items themselves. The items render their decisions based on local information following predefined methods. In previous work, two different methods were developed. The first one is based on the calculation of the future throughput times; the second one bases on information left by other items that have already passed the network. In presence of high fluctuations in demand and/or internal disturbances, these autonomous control methods are able to control highly complex systems. The influence of the complexity of manufacturing systems on the performance of autonomous control methods is analysed using discrete event simulations. This gives a hint to the limits of the applicability of autonomy in manufacturing i.e. in which structures which autonomous control method is advisable. As an exemplary scenario, a dynamic model of a job shop is used consisting of multiple parallel lines and stages in a matrix form. In this scenario, the items are free in their decision on which machine at their current stage they will be processed. They render these decisions based on local information like the buffer content or the processing time at the respective machine combined with the information stored on the item itself. The scenarios’ complexity is indicated by the number of elements within the network and the number of different order classes. The ability to cope with rising complexity of two different autonomous control strategies is analysed by observing the system’s logistic performance in terms of mean and variance of the throughput time.
Van Campen, Edgar (Philips Semiconductors)
Cycle time in semiconductor manufacturing: challenges for the EPT
In this presentation the relevance of cycle time performance in the newest generation 300 mm wafer fab is explained. Various elements at equipment and factory level affecting the cycle time will be addressed. The current practice regarding cycle time monitoring and modeling at Crolles 2 wafer fab is used to indicate the challenges for the EPT.
Van der Eerden, Joris (ASML)
Litho area cycle time improvements in an advanced 300mm semiconductor manufacturing line
Semiconductor wafer fabrication is one of the most complex and costly manufacturing processes. Building a wafer fab costs over 2 billion dollars and operating costs run into hundreds of millions of dollars a year. A good fab design, fast ramp-up and efficient usage of the equipment is essential to stay ahead of competition.
The cycle time of products through a manufacturing line is an essential business objective for semiconductor manufacturers. An optimized manufacturing line has short and predictable cycle times. Short cycle time will result in a lower inventory and faster time to market.
Lithography is by design the bottleneck process of the fab. Being one of the processes that is repeated most during manufacturing, any reduction in the lithography cycle time will reduce the overall manufacturing cycle time. Given the complex processing steps involved in the litho area and the high utilization of the litho clusters the complexity of cycle time reduction projects is high.
Texas Instruments and ASML together started a study to investigate the current litho cycle time performance of an advanced 300mm semiconductor manufacturing line. Target of the project was to reduce the litho cycle time by 35%.
First set of data was based on 6 months of litho area production data with more than 200,000 number of lot processing steps involved. In total 26 improvements have been identified. A second set of data has been taken consisting of 100,000 number of lot processing steps over a 3 months period to study the effect of the improvements. At the end of the project the litho cycle time was reduced by almost 50% with still a number of improvements to be implemented.
New methodologies have been used to identify root causes of cycle time losses. Methodologies based on the Effective Process Time (EPT). One advantage of the EPT methodology is that it also includes variability, one of the key parameters influencing cycle time. This is the first time EPT methodology is applied to litho clusters (resist track + exposure scanner) which are considered cascading systems. The methodology has been applied to single processing and batching systems but never to cascading systems.
Once the systems are identified that have the highest contribution to litho cycle time, a more thorough analysis is used to find the root causes. In the thorough study, several new methodologies are used to identify the true (effective) utilization of litho clusters, and cluster uptime (uptime of a combined cluster of clean track and exposure scanner).
Van Vuuren, Marcel (TU/e)
Performance Analysis of a Production Line with an Assembly Node
Inspired by a case in the automotive industry we designed a method to predict performance measures such as throughput and sojourn time of a production line with an assembly node. To tackle this case we developed two separate methods. First we present an improved algorithm for approximating tandem queue with small buffers. Also, we present an algorithm to approximate an assembly queue. Finally, we combine the two algorithm in order to tackle the case.
Weiss, Gideon (University of Haifa)
Joint work with Anastasia Kopzon
Virtual infinite buffers as a tool for heavy traffic modeling
We consider a multiclass queueing system with two machines (service nodes) and four classes (queues of parts waiting for processing), where there are two streams of products, one stream is processed first in station 1 and then in station 2, the other stream moves in the opposite direction. This push pull system is distinguished from the network of Rybko and Stolyar in that we assume that there is an unlimited number of customers available in each of the two streams.
We shall discuss this model as an alternative to the balanced heavy traffic paradigm, in which systems become congested as the traffic intensity rho approaches 1. In our push pull system both machines can be busy all the time, i.e. work with utilization rho=1, and yet the system is not congested: it is possible to control the queue sizes, so that the number of parts in process is positive recurrent. This is in contrast to the Rybko-Stolyar network, which becomes congested as rho approaches 1.
To get some concrete results we assume processing times are memory-less. We then construct a family of generalized threshold policies for which the push pull system is stable, i.e. positive recurrent, with full utilization. We derive the steady state distribution for some of these policies.
This workshop is sponsored by
This page is maintained by Lucienne Coolen