Wednesday, June 4 Speaker: George Tiao, University of Chicago Abstract: TBD
Abstract: Shortly after his paradigm shattering paper introducing response surface experimental designs (JRSS,B (1951)) George Box was invited by Gertrude Cox to spend a year as a research professor at the Institute of Statistics at NC State. In his letter of acceptance he wrote that he was interested not only in experimental design but in testing the “robust” nature of the F and t tests, an interest that naturally flowed from some earlier work he published in Biometrika (1941). In January 1953 George Box arrived in Raleigh and I and Sigurd Andersen were assigned to him as graduate students. This brief talk hopes to capture some of the spirit and joy of that wonderful year. Wednesday, June 4 Title: TBD Title: TBD Title: TBD Wednesday, June 4 Title: Simulation on Demand Abstract: For years the operations research community has worked on making manufacturing/distribution/service system simulations faster or more statistically efficient. However, in many contexts, simulation time is plentiful, but decision-maker time is scarce. The philosophy of “Simulation on Demand” is to use intensive simulation up front to build a high quality response surface model that will allow for very fast, but very precise, queries throughout the factor space without the need for additional simulation runs. Since the emphasis is on the efficiency of obtaining useful simulation results, rather than on the efficiency of the simulation run itself, this is a design of experiments problem for building the response surface model to a specified precision through simulation experiments. An example of a semiconductor fabrication plant will be used to show how, “Simulation on Demand” represents a bridge between the flexibility and accuracy of a simulation model and the insight provided by an analytical queueing model. (Joint work with Barry Nelson, Feng Yang, Jerry Mackulak and John Fowler.)
Abstract: An integrated approach for estimation and reduction of measurement variation (and its components) through a single parameter design experiment is developed. Systems with a linear signal-response relationship are considered. The noise factors are classified into a few distinct categories based on their impact on the measurement system. A random coefficients model that accounts for the effect of control factors and each category of noise factors on the signal-response relationship is proposed. Two different data analysis strategies - response function modeling (RFM) an performance measure modeling (PMM) are proposed and compared. The effectiveness of the proposed method is demonstrated with a simulation study and Taguchi's drive shaft experiment. (joint with Arden Miller and Jeff Wu) Title: Bayesian Optimal Single Arrays for Robust Parameter
Design Abstract: It is critical to estimate control-by-noise
interactions in robust parameter design. This can be achieved by using
a cross array, which is a cross product of a design for control factors
and another design for noise factors. However, the total run size of such
arrays can be prohibitively large. To reduce the run size, single arrays
are proposed in the literature, where a modified effect hierarchy principle
is used for the optimal selection of the arrays. In this article, we argue
that effect hierarchy is a property of the system and cannot be altered
depending on the objective of an experiment. We propose a Bayesian approach
to develop single arrays which incorporate the importance of control-by-noise
interactions without altering the effect hierarchy. The approach places
no restrictions on the number of runs or levels or type of factors and
therefore, is very general. The advantages of the proposed approach are
illustrated using several examples. (Joint work with Lulu Kang.) Abstract: Six Sigma has been applied with great publicity and success at a large variety and number of companies, beginning in the United States and spreading to some extent to Europe and Asia. We investigate the extent, success and nature of penetration of Six Sigma methods in China. The focus of the research reported on in this presentation is on the extent (how much activity), the success (how well have typical Six Sigma financial and process goals been accomplished) and the nature of the penetration (how the methodology has been adapted to deal with the unique nature of business in China, ranging from MNCs with activity in China to SOEs). The results are based on literature surveys and in-depth interviews with executives and other key personnel (including Master Black Belts and Black Belts) involved in Six Sigma initiatives in a set of companies in China, and with key representatives of consulting firms who offer Six Sigma consulting services in China. We conclude with observations of unique strengths and weaknesses in Six Sigma as it is being introduced in Chinese firms, and with suggestions for strengthening the applications. Wednesday, June 4 Invited Plenary Session 3– Presentation of the 2008 Conference Honoree Award, C.F. Jeff Wu, Georgia Tech Honoree’s Address: Quality Technology in the High-Tech
Age Thursday, June 5
Abstract: In an effort to speed the development of new products and processes, many companies are turning to computer simulations to avoid the expense and lost time of building prototypes. These computer simulations are often very complex and may take hours to complete one run. If there are many variables affecting the results of the simulation, then it makes sense to design an experiment to gain the most information possible from a limited number of computer simulation runs and build a response surface type of model, usually called a surrogate model. The absence of noise is the key difference between computer simulation experiments and experiments in the real world. Since there is no variability in the results of computer experiments, optimal designs based on reducing variance have questionable utility. Replication, usually a "good thing", is clearly undesirable in computer experiments. Thus, a new approach to experimentation is necessary. This talk takes a case study approach using computer models to demonstrate new software that supports both computer simulation design and analysis. Title: Comparing Designs for Computer Simulation Models Abstract: The use of simulation as a modeling and analysis tool is wide spread. Simulation is a powerful tool, which enables the users the ability to experiment virtually on a validated computer environment. Often the output from computer experiments results in a complex response surface. Determining which experimental design should be used is a non trivial task. An evaluation of several computer experiments is made assuming the modeler is interested in fitting a high order polynomial to the response data. Optimal and space-filling designs are explored. Additionally, a new class of hybrid designs are introduced, which can be used for deterministic or stochastic simulation models. Discussant: Geoff Vining, Virginia Tech Thursday, June 5
Abstract: Repeated measures degradation studies can be used to assess product or component reliability when there are few or even no failures during the study. This talk will describe an algorithm that can be used to compute approximate large-sample variances and covariances from a mixed effect linear regression model for repeated measures degradation data. The algorithm has several important applications: assessing the effect of sample size (number of units and number of measurements within the units), evaluation of proposed test plan properties to allow the design of statistically efficient test plans, and the computation of approximate confidence intervals for functions of the parameters such as quantiles of the degradation distribution, failure probabilities, and quantiles of the failure time distribution. These applications will be illustrated with examples. The algorithm has been implemented in R and S-PLUS. This talk is based on joint work with William Q. Meeker and Luis A. Escobar.
Abstract: Materials degradation experiments are used to investigate the impact of various experimental factors on the degradation of specified response variables as a function of time. Results obtained from studying degradation in a controlled setting provide information that can be used to develop predictions of the performance of materials as they age. Often, degradation experiments involve acceleration factors, such as temperature, that allow the experimenter to investigate degradation processes over a much shorter time span than simply observing the materials over ordinary time. A number of statistical issues arise in the design of materials degradation experiments, including selection of runs, sampling frequency, and choice of acceleration levels. Analysis may involve a variety of types of responses, such as images or functional data. Examples of materials experiments will be used to illustrate the statistical challenges associated with degradation studies.
Abstract: In accelerated reliability testing, the occurrence time of failures are used to estimate the life distribution for extrapolation to field conditions via an acceleration model. Often, because of time limitations on testing, many if not most of the units undergoing stress do not fail, resulting in censored observations. In certain situations, however, it is possible to do an analysis even without any actual failures, provide that there is a measurable product parameter that is degrading over time towards a value that is defined to be a failure level. In this talk, we will review key concepts in degradation analysis including physical models, assumptions, analysis methods, and acceleration considerations. We will illustrate, using an example in microelectronics, how field reliability projections can be obtained from accelerated life tests. We also will provide some simulation results to reveal factors that can significantly influence the estimation error. Thursday, June 5 Title: Pitting Corrosion: Analysis of Designed Experiments
with Extreme Value Distributed Responses Abstract: This talk discusses how Extreme Value Statistics can be used to validate and improve designed experiments with extremal responses, and how to extrapolate and compare results. A main motivation is corrosion tests: Localized, or "pitting", corrosion can limit the usefulness of magnesium and other new lightweight materials aimed at reducing weight, and thus CO2 emissions from cars. It makes judicious choice of alloys and surface treatments necessary. Standard methods to evaluate corrosion test are based on analyzing weight loss by ANOVA. These methods may be misleading in two ways. Usually it is not weight loss but the risk of perforation, i.e. the depth of the deepest pit which is of interest. Further the standard ANOVA assumptions of normality and homogeneity of variances typically are not satisfied by pit depth measurements, and don't give credible extrapolation into extreme tails. The talk presents a streamlined approach to analysis of such experiments. It consists of 1.) a preliminary analysis to check that the experiment has worked out as intended, 2.) a separate analysis for each individual experimental condition, and 3.) a strategy for pairwise comparisons. Of particular interest is to try to separate situations with many but shallow pits from potentially more dangerous situations with rare but deep pits. Further, a class of hierarchical random effects models for extreme value data is discussed. The models are obtained by mixing extreme value distributions over positive stable distributions. They open up exciting new possibilities to develop extreme value analogues of normal random effects models. The big challenge is to develop a full theory of design and analysis of experiments with extreme value distributed responses – the present results are just an early beginning. Short biography: Hologer Rootzn is professor of Mathematical Statistics at Chalmers University of Technology, Sweden, and has earlier worked in Lund, Chapel Hill and Copenhagen. He is Editor of Bernoulli, was founding editor of Extremes, is leader of GMMC, the Gothenburg Mathematical Modelling Centre, and is chairman of the Gothenburg Stochastic Centre. He has led joint research projects with a number of companies, including Volvo Car. His research interests include stochastic theory, extreme values, financial risk, and applications in engineering, medicine, and industry.
Title: Estimating Extreme Quantile Regions for Two Dependent
Risks Abstract: When simultaneously monitoring two dependent, positive risks one is often interested in quantile regions with very small probability p. These extreme regions contain hardly or no data and therefore statistical inference is difficult. In particular when we want to protect ourselves against a calamity that has not yet occurred, we consider the case where p is smaller than 1/n, with n the sample size. In practice, p typically ranges from .01 to .0000001. We consider quantile regions of the type {(x,y): f(x,y)<= c}, where f is the joint density. Such a region has the property that its complement consists of the most likely points, that is, is as small as possible. Using extreme value theory, we construct a natural, nonparametric estimator of such a quantile region and study its behavior. This is joint work with Laurens de Haan (Rotterdam and Lisbon). Short biography: John H.J. Einmahl is professor of Statistics at the Department of Econometrics and research fellow at CentER, both at Tilburg University. He has published in leading journals in Statistics and Probability Theory. His main research area is nonparametric statistics, including statistics of extremes, empirical likelihood, generalized quantiles, and empirical processes. He obtained his Ph.D. from Nijmegen University. He is is an Associate Editor of Bernoulli and Extremes. He is an elected fellow of the Institute of Mathematical Statistics, and since 2005 a member of its council. He was a co-director of the Stochastics of Extremes and Risk Analysis program at the European research institute Eurandom (Eindhoven) from 2001-2004. He was an Associate Editor of The Annals of Probability (1999-2003), and of The Annals of Statistics (2001-2004). In 1998 he visited Florida State University as a Senior Fulbright Scholar. Thursday, June 5 Title: Screening Designs with Reasonably-Balanced Projections Abstract: Screening designs are useful in product development to quickly identify a few factors having relatively large impact on a response from a much larger set of factors. They are also useful in robustness studies to demonstrate that product performance is insensitive to a host of variables within specified ranges. Regular fractions of 2- and 3-level factorial designs are often used for screening designs and tables and properties of these designs are well-understood. This paper introduces a balance metric which is obtained by examining all possible projections of the design into one, two and three dimensions. This loss function is shown to be related to a projectivity measure and entropy. Standard fractional factorial screening designs are shown to be optimally-balanced relative to this measure. A general strategy for constructing screening designs with reasonably-balanced
projections is presented. Some example designs are given, including a
design that was employed by the author in the pharmaceutical industry.
Analysis approaches are described. SAS code is available to construct
designs of arbitrary levels and runs.
Title: What Commercial DOE Software Should Do
Abstract: A general algorithmic approach to the construction of optimal designs is introduced. The algorithms use a columnwise exchange procedure, which can be particularly useful when certain designs structures need to be retained during the search for optimal designs. The efficacy and efficiency of the algorithms are demonstrated by their applications in several areas that include optimal balanced designs, mixed-level designs, Latin-hypercubes, and design repair problems. Some comparisons between the columnwise algorithms and other competing algorithms are obtained. Thursday, June 5
Abstract: The environmentally mandated industry migration from Pb-based solder to lead-free solder requires reliability evaluation of new solder joint configurations. Other than being lead-free, properties of the preferred substitute solder (Sn) are generally inferior for the solder bump application, so that candidate alloys and configurations intended for this application require extensive testing. Two of the more troublesome properties are solder bump stiffness and electromigration resistance, both of which are detrimental to solder joint reliability. Solder bump stiffness becomes a problem when packaged parts experience temperature changes. Electromigration (EM) is the movement of metal atoms in response to an applied current. This movement produces voids in the solder joint that increase the joint resistance and eventually lead to chip failure. More attention has been paid to EM recently because the industry is moving to denser chips with higher power. At the same time that the joint size is decreasing, the industry is also moving to Sn based solders which exhibit different behavior than does Pb, primarily because of the asymmetry of the tetragonal crystal structure. Because of limitations on the acceleration of such testing, considerable time and expense are necessary to define the reliability of the lead-free joints. Due to this expense, there are practical limitations on the degree to which solder bump reliability can be determined in the near future, and this requires a conservative limit to current and temperature to be applied to operating conditions.
Abstract: There are many tools and techniques used by reliability engineering practitioners in the execution of reliability programs. Some of these tools include: Quality Function Deployment (QFD), Failure Mode and Effects Analysis (FMEA), Design of Experiments (DOE), Test Planning, Accelerated Testing, and Life modeling. Each of these has a specific purpose, and is often accomplished with different teams within a project. As a result, they are often executed in an autonomous fashion, with little or no leveraging amongst the various tools. While each of these tools accomplishes a specific goal with different approaches, there is a significant amount of synergy that exists if they are treated as a “system” of tools, and are leveraged to the maximum extent. This presentation will discuss the manner in which these tools can be linked, resulting in a more effective and efficient reliability program.
Abstract: Accelerated test methods are essential in developing product maturity. The market place is full of highly competitive companies prepared to offer the latest technologies by the time you wake up tomorrow. Consumers fuel this cadence with demands of; “I want it all… and I want it now!” Engineers and statisticians work diligently to fulfill these demands with reliable products, but find it difficult to agree upon test methods that are both effective in proving reliability and can do so quickly using small sample sizes. The “physics of failure” approach to reliability was the first major step toward finding solutions to this problem and has formed a strong synergistic relationship with the world of accelerated testing. The Calibrated Accelerated Life Test method (CALT) is a product of today’s synergy between the physical sciences and statistics while building upon the legendary work of Wayne Nelson and William Meeker. The CALT method is based upon the concept of a life-stress relationship, which covers a wide range of phenomenon in the world of physics and chemistry. Unique to this method is the ability to select the stress levels that will ideally allow the test process to occur within the time window required by the engineer. This discussion will provide the rational behind this methodology through examples from the automotive industry. Additionally, the concept of Degradation Analysis applied within the CALT framework will be explained with examples shown. The CALT methodology decouples time and sample size from the reliability
requirement. This simple attribute is exceptionally powerful and useful
in today’s world of product development. Abstract: The intermediate and long-term vision for medical imaging systems is for continuing increases in system-level performance and functionality combined with decreasing cost and size. Improvements in image quality, in particular, increased spatial and temporal resolution to allow clinicians to make better diagnoses, leading to better patient outcomes. These requirements, in turn, flow down to increasing channel densities for the sensitive signal electronics hardware inside these imaging systems. This densification requires novel electronics packaging solutions. Present-day imaging systems may have several hundred thousand channels combined with a system-level 10-year life requirement. One technique used to increase channel densities is through the widespread use of Chip-Scale Packages (CSPs) for sensitive mixed-signal integrated circuits. In these packages, the die is mounted on an interposer which makes electrical connections to a Printed Circuit Board (PCB) through an array of solder balls. The most common failure mechanism for properly designed CSP assemblies fail in low-cycle fatigue at the solder joint. We present the results from a CSP reliability study. The primary factors influencing reliability are the solder ball pitch, contact metallurgy and solder ball location. To determine reliability, a series of accelerated life tests under cyclic thermal stress were performed. Two different pitches were tested (0.8 and 1.0 mm), along with two different contact materials (Cu/OSP and Ni/Au). Solder ball location was a key noise parameter. Test assemblies were specially designed to emulate the structure of the final assembly and allow for testing of the critical variables. The assembly consisted of an array of CSPs with a “daisy-chain” interconnect pattern solder attached to a PCB substrate with controlled contact metallurgy and structures to allow easy electrical testing. The accelerated test protocol was air-to-air thermal cycling from 0C to 100C over several thousand cycles. Results from the accelerated life testing indicated that all the configurations investigated should exceed the 10-year system life requirement. Thursday, June 5 Title: Optimal Designs for Network Traffic Measurement
Schemes Abstract: Computer network traffic can be sampled at routers to determine the origin and destination of different packets. This is in turn used to estimate the volume of traffic between all origin-destination pairs. Use of higher sampling rate leads to better estimates but uses more resources. In this work we formulate the problem of determining the network-wide sampling rates as an optimal design problem. This problem is solved using semi-definite programming. The proposed approach is evaluated on a number of real and synthetic data sets. Title: Analysis of Window-Observation Recurrence Data Abstract: Many systems experience recurrent events. Recurrence data are collected to analyze quantities of interest, such as the mean cumulative number of events. Methods of analysis are available for recurrence data with left and/or right censoring. Due to practical constraints, however, recurrence data are sometimes recorded only in windows. Between the windows, there are gaps over which the process cannot be observed. We extend existing statistical methods, both nonparametric and parametric, to window-observation recurrence data. The nonparametric estimator requires minimum assumptions, but will be inconsistent if the size of the risk set is not positive over the entire period of interest. There is no such difficulty when using a parametric model for the recurrence data. For cases in which the size of the risk set is zero for some periods of time, we propose and compare two alternative hybrid estimators. The methods are illustrated with two example applications. (This is joint work with Jianying Zuo and William Q. Meeker). Title: Maximum Likelihood Estimators of Clock Offset
and Skew under Exponential Delays Abstract: Accurate clock synchronization is essential for many data network applications. Various algorithms for synchronizing clocks rely on estimators of the offset and skew parameters that describe the relation between times measured by two different clocks. Maximum likelihood estimation (MLE) of these parameters has previously been considered under the assumption of symmetric and exponentially distributed network delays. We derive the MLEs under the more common case of asymmetric exponentially distributed network delays and compare their mean squared error properties to a recently proposed alternative estimator. We investigate the robustness of the derived MLE to the assumption of non-exponential network delays, and demonstrate the effectiveness of a bootstrap bias-correction technique. Thursday, June 5 Title: Translating Understanding of a Complex System
into a Statistical Model Abstract: Developing appropriate statistical models for system reliability of complex systems involve many diverse aspects and expertise from a variety of different experts. Understanding the system structure typically requires the expertise of several engineers, while building a model that can incorporate multiple types of data involves advanced statistical thinking. Through a case study, the talk describes the process and some of the potential pitfalls for eliciting information from collaborators with a variety of backgrounds, synthesizing different types of engineering and statistical knowledge, and verifying that the final statistical model reflects actual system characteristics.
Abstract: Classroom discussion is an effective way for students to interact and learn. A class on design of experiments usually covers many related design strategies. It can be useful to have one running example (similar to a case study) that can be used throughout the class to extend earlier topics to later more complex topics. This talk presents one such running example that studies plant growth, mostly in a greenhouse. The greenhouse experiment begins with a simple t-test and ends with response surface methodology. By making small changes to the scenario, most of the standard topics covered in a one-semester design of experiments class are discussed.
Abstract: This non technical talk describes an innovative and successful use of technology, through a virtual process, to aid in the teaching of statistical concepts and methodology. The virtual process simulates a manufacturing process for automobile camshafts that has a number of processing steps and many inputs. At the start of an upper year undergraduate course Stat 435/835: Statistical Methods for Process Improvement, each team of students is given a budget and assigned the task of reducing variation in a critical output characteristic of a different version of the virtual process. Throughout the term, the teams plan and analyze a series of process investigations (~1/week) to first learn about how their process works and, by the end of term, how to improve it. The teams interact with the virtual process through a web interface. Each team submits a weekly written report describing their recent progress and twice per term presents to the class at a “management review meeting.” The virtual process is also used as the context for all midterms and exams. Based on anecdotal evidence and survey results, students find interacting with the virtual process fun, stimulating and challenging. The goals of this talk are to show how the virtual process aids in the teaching of material and concepts in Stat 435/835 and to describe its main pedagogical benefits. With thought and some adaptation something similar should be possible for other applied statistics courses. Stefan Steiner is an Associate Professor in the Statistics and Actuarial Science department, and director of the Business and Industrial Statistics Research Group (BISRG) at the University of Waterloo. Thursday, June 5 Title: Analysis of Optimization Experiments Abstract: The typical practice for analyzing industrial
experiments is to identify statistically significant effects with a 5%
level of significance and then to optimize the model containing only those
effects. In this article, we illustrate the danger in utilizing this approach.
We propose methodology using the practical significance level, which is
a quantity that a practitioner can easily specify. We also propose utilizing
empirical Bayes estimation which accounts for the randomness in the observations.
Interestingly, the mechanics of statistical testing can be viewed as an
approximation to empirical Bayes estimation, but with a significance level
in the range of 15--40%. We also establish the connections of our approach
with a less known but intriguing technique proposed by Taguchi. A real
example and simulations are used to demonstrate the advantages of the
proposed methodology. (Joint work with V. Roshan Joseph) Abstract: This talk will review a number of recent advancements related to adaptive-One-Factor-at-a-Time (aOFAT) experimentation. Previously, our research group showed that aOFAT provides high expected values of improvements and high probabilities of exploiting two-factor interactions, especially when these interactions are large. More recently, we have investigated the properties of adaptive experimentation from a Bayesian perspective. This enables one to determine the degree to which aOFAT represents an efficient response to step-by-step accrual of sample information and helps one assess aOFAT in comparison to "greedy" algorithms. We also extend the aOFAT approach by incorporating concepts from "ensemble methods." In the approach, the experimenter plans and executes multiple adaptive experiments on the same system with differences in starting points and orders in which factors are varied. Decisions are made by aggregating the multiple outcomes of aOFAT experiments. This approach is compared to alternative methods using both a case study to illustrate the approach and a hierarchical probability model to seek more general conclusions.
Abstract: Given a group of tasks and two non-identical processors with the ability to complete each task, which tasks should be assigned to which processor in order to complete the given group of tasks in as short amount of time as possible? Processors may be airport runways, shipping trucks, manufacturing lines, computer processors, or surgeons. The tasks may be airplane take-offs and landings, packages to be shipped, products to be built, computer codes to be run, or patients to be operated on. This problem has been formalized in the scheduling literature as the minimization of the makespan (time required to complete all tasks) for two unrelated parallel processors. One possible approach to solving this problem is to create a computer
simulation model of the process and apply this model to all possible task
assignments (complete enumeration), selecting the assignment that produces
the minimum makespan. A typical implementation of the complete enumeration
approach is to calculate the makespan for each assignment schedule one
by one, noting the minimum makespan and associated assignment schedule
observed thus far. This talk discusses the benefit realized by implementing
the complete enumeration approach using a 2k full factorial experimental
design framework, as illustrated by a printed circuit board assembly example.
Also explored is the possibility of employing 2k - p fractional factorial
experimental designs in the solution of the two unrelated parallel processors
problem. Thursday, June 5 Madison, Wisconsin, the host city for the 2008 Quality and Productivity Research Conference, is closely linked with several seminal texts on productivity and quality. This session brings three authors together to discuss the enabling factors that this city provided to develop what can be viewed as the fundamental tools for improving quality and productivity, as well as their individual works. Brain Joiner’s “Fourth Generation Management” provides the organizational framework to realize quality gains. Peter Scholtes’ books, “The Team Handbook” and “The Leader’s Handbook”, provide the tools and philosophy necessary to run and manage effective improvement teams. J. Stuart Hunter’s “Statistics for Experimenters” provides the statistical engine to extract the maximum amount of information from the data collected by the teams. The work of all three authors fit together to drive overall improvement. Speaker: J. Stuart Hunter, coauthor of “Statistics
for Experimenters”. Speaker: Brian L. Joiner, author of “Fourth Generation
Management”. Speaker: Peter R. Scholtes, author of “The Team
Handbook” and “The Leader’s Handbook”. Thursday, June 5
Abstract: We show how to use Bayesian methods to make inferences from an accelerated life test where test units come from different groups (such as batches) and the group effect is random and significant both statistically and practically. Our approach can handle multiple random effects and several accelerating factors. We illustrate the method with an application concerning pressure vessels wrapped in Kevlar fibers, where the fiber of each vessel comes from a single spool and the spool effect is random. Bayesian analysis using Markov chain Monte Carlo (MCMC) methods is used to answer questions interest in accelerated life tests with random effects that are not as easily answered with more traditional frequentist methods. For example, we can predict the lifetime of a pressure vessel wound with a Kevlar fiber either from a spool used in the accelerated life test or from another random spool from the population of spools. We comment on the implications that this analysis has on the estimates of reliability (and for the Space Shuttle, which has a system of 22 such pressure vessels. Our approach is implemented freely available WinBUGS software so that readers can apply the method to their own data.
Abstract: Reliability analytics are important not only for determining when a part or complex system will fail but also in developing an overall strategy for determining on-going operational costs, quantifying financial risk, and developing a comprehensive maintenance and repair strategy. In this talk we will discuss various reliability and repair models and will provide examples from industry of how these techniques are used to quantify the risk associated with customer service agreements.
Abstract: We consider the problem of monitoring parameters of a lifetime distribution based on the sequence of life tests with censoring. This problem arises in a number of settings, for example in administration and interpretation of Ongoing Reliability Tests (ORT) data or warranty data. The main technical difficulty of the monitoring problem is related to the fact that information obtained at each point in time affects data for every life test under observation. This feature necessitates use of special methods and special metrics for their evaluation. In this paper we describe a new approach to monitoring such data streams and give examples related to detection of unfavorable changes in warranty data in the context of a large scale system for warranty data analysis. Thursday, June 5 Title: Calibrating a Computer Code in the Presence of
Systematic Discrepancy Abstract: Computer models to simulate physical phenomena are now widely available in engineering and science. Before relying on a computer model, a natural first step is often to compare its output with physical or field data, to assess whether the computer model reliably represents the real world. Field data, when available, can also be used to calibrate unknown parameters in the computer model. Calibration can be particularly problematic in the presence of systematic discrepancies between the computer model and field observations. In this talk we present results on a simulation study that is designed to assess how well the calibration parameter has been estimated, and the conditions under which calibration is possible. By simulating both computer model data, and physical observations from a Gaussian process the uncertainty due to using the incorrect model does not arise. This allows us a more accurate picture of the problems that can arise when attempting to calibrate the model in the presence of systematic discrepancy.
Abstract: A level-set-based multistage metamodeling approach is proposed for design optimization. Different from the existing objective-oriented sequential sampling methods, where design objective and constraints have to be combined into a single response of interest, our method offers the flexibility of building metamodels for multiple responses (design objective and constraints) simultaneously. Uncertainty quantification (UQ) is introduced for each metamodel to represent the confidence interval due to the lack of sufficient samples in the early stages of metamodeling. Based on the extreme values of the optimal solution identified, the level set method (LSM), together with a series of Boolean operations, are used to identify the region(s) of interests with arbitrary topology. Introducing LSM facilitates region manipulations, detection of disconnected regions of interest, and visualization for design exploration. The proposed method possesses a superior efficiency in design exploration to conventional sequential sampling strategy and allows the use of multiple samples in each sampling stage. The next-stage metamodel is adaptively fit based on the pre-existing and newly introduced sample points. As the metamodeling process moves on, the region of interests is progressively reduced, and the optimal design is asymptotically approached. The proposed approach is demonstrated with mathematical benchmark examples and an engineering design problem. The results are compared with those from one-stage metamodeling using the Optimal Latin Hypercubes (OLH) experiments. Our results show that the proposed method can effectively capture the region of interests with any shape and topology and locate the optimum design with improved efficiency.
Abstract: Despite their obvious advantages, the incorporation of computer simulations in the study of complex science and engineering systems introduces many challenges. To justify the inclusion of information obtained from numerical models, some level of predictive capability of the corresponding simulation codes must be guaranteed. In addition, the degree to which a computer model accurately represents the real world phenomena it seeks to represent must be quantified. Moreover, calibration techniques should be applied to both improve the predictive capability of the model and to curtail the loss of information caused by using a numerical model instead of the actual system. Evaluating and calibrating a computer model is dependent on three main components: experimental data, model parameters, and algorithmic techniques. Data is critical as it defines reality, and inadequate data can render model evaluation procedures useless. Model parameters should be studied to determine which affect the predictive capabilities of model and which do not. Those that do are subject to calibration. Techniques for model evaluation and calibration should both sufficiently sample the design space and limit the computational burden. In this talk, we will discuss the problems inherent in model calibration and validation processes in terms of data, parameters and algorithms. In particular, we will focus on suitable techniques of optimization and statistics for some specific numerical models from the areas of electrical engineering, medicine, and groundwater control. We will demonstrate some successful and some not so successful approaches. Thursday, June 5 Title: Diagnosis of Process and Sensor Faults in Manufacturing
Processes Abstract: This paper presents an approach to identify process mean shift faults and sensor mean shift faults simultaneously in manufacturing processes based on Bayesian variable selection techniques. The potential process faults and sensor faults always outnumbers the number of sensor measurements. The proposed approach utilizes the fact that only a few process and sensor faults can occur simultaneously. A guideline on how to select the values of hyperparameters in the Bayesian variable selection is also given. Several simulated examples and a real assembly example are studied to demonstrate the performance of the developed diagnosis method. A sensitivity analysis is conducted for the assembly example, which shows that the proposed approach is a robust procedure. Title: Multiple Fault Signature Integration and Enhancing
for Variation Source Identification in Manufacturing Processes Abstract: Recently, signature matching as a popular method for variation source identification in manufacturing processes has drawn significant attention. In this method, the variation source is identified through matching the variation patterns of specific process faults, also called fault signatures, with the variation patterns in the newly collected quality data. There are situations that a fault occurred multiple times before and consequently multiple signatures exist for the same fault. A technique is proposed in this article which can integrate multiple signatures together to enhance the accuracy of variation source identification. A linearly combined fault signature is constructed to increase the detection power of the fault identification. A numerical study is also presented to validate the effectiveness and robustness of the proposed method.
Abstract: Warranty data can play a critical role in quality control. A quantitative relationship between this product feedback information and manufacturing data can be used to predict the product failures, identify the process parameters that cause them, and ultimately make improvements to the process. The main challenge in establishing such a relationship is the small number of faulty units in a lot. This leads to a highly skewed distribution of the response variable, field return, which prevents the well-known off-the-shelf classification methods from performing well. We propose a statistical procedure in such a class-imbalance problem. The proposed procedure is applied to a real dataset from a cell phone manufacturer and it performs significantly better than well known off-the-shelf methods. Title: Multiscale Wavelet Analysis for Multiple Embedded
Operations Monitoring Using Aggregated Signals Abstract: Aggregated signals are referred to the system responses contributed from the multiple operations embedded in a system. While abundant literature exists in analyzing signal profiles of system responses, little has been found on how to identify features of the aggregated signals in order to map these features to the individual operations. For this purpose, this research develops a multiscale feature mapping algorithm using wavelets. It shows that the proposed mutliscale mapping can enhance the separation of the individual operations that contribute to the aggregated signals not only in the localized time segments corresponding to the working ranges of individual operations, but also at the different scales corresponding to the frequency ranges of the operations. A case study is provided to validate the developed method. Friday, June 6
Abstract: Storage systems have an essential role in enterprise and mission critical applications - to provide uninterrupted access to data, prevent its loss and safeguard its integrity, and to do all of these while maximizing application performance. In this talk, we will present some of the challenges continued to be faced by storage systems designers that cause data loss or data integrity issues. We will also present some of the trends and modeling techniques employed in storage systems in achieving high reliability.
Abstract: At Boston Scientific, our mission is “to improve the quality of patient care and the productivity of health care delivery through the development and advocacy of less-invasive medical devices and procedures.” Consequently, our customers - our patients - expect our products to perform their intended function every time, time-after-time. As you can imagine, it may be either impossible or impractical to physically test every component of every product. However, the statistical methods of acceptance sampling and distribution fitting can be very effective at determining if a process is capable of meeting its specifications with a high degree of confidence. In this presentation, we explain the relation between risk and reliability, and we show how to use reliability methods to address some situations when the usual validation approach may no longer be appropriate.
Abstract: In a Perfect World all components and subsystems would have stringent reliability requirements that are proven via testing of multiple units over long durations at conditions identical to their use environment. However, we must live in the Real World. In the Real World, we must prove our products meet their reliability requirements within the limitations of budget and schedule. By looking at three examples of real components/subsystems with varying degrees of complexity, significance, and resource constraints, we will discuss the process followed within GE Healthcare-Life Support Solutions for developing reliability requirements and demonstrating compliance with those requirements. Friday, June 6
Abstract: George Box is one of the most respected living
statisticians. He has made fundamental contributions to many areas of
statistics. In this talk we will look at some of his contributions to
industrial statistics and quality. We will touch on several events such
as the birth of response surfaces, creation of Technometrics, and the
founding of the centre for quality and productivity improvement at the
University of Wisconsin. Prof. Box's papers form a collection of contributions
that have had a significant impact on the research and practice of statistics
and quality around the world during the last half a century. Abstract: TBD
Abstract: George Box is truly one of the towering figures in the history of industrial experimentation, going back to his seminal work on response surface methodology, Box and Wilson (1951). This talk outlines George pivotal works in industrial experimentation through out his career. Of especial interest are his contributions to sequential experimentation, designs to support second-order response surfaces, the impact of bias due to an underspecified model, evolutionary operations, and the importance of balancing many often conflicting criteria when selecting appropriate experimental designs.
Abstract: In a recent article Kulahci et al. (2006) showed that minimum aberration is not designed to distinguish among the different types of two factor interactions when designing fractional factorial split-plots. Here we expand this idea to fractional factorial designs in situations where more than one split is involved; e.g., multi-step processes. We give design and model constraints to generate fractional factorial designs for multi-step situations, and examples of a 3-step process (split-split-plot), a 4-step (split-split-split-plot), and a 5-step (split-split- split-split-plot).
Abstract: In many industrial experiments, some of the factors are not independently reset for each run. This is due to time and/or cost constraints and to the hard-to-change nature of these factors. A lot of research has been done on the design of experiments with hard-to-change factors. Most of this work restricts the attention to split-plot designs in which all the hard-to-change factors are reset at the same points in time. This constraint is to some extent relaxed in split-split-plot designs because these require the least hard-to-change factors to be reset more often than the most hard-to-change factors. A key feature of the split-split-plot designs, however, is that the least hard-to-change factors are reset whenever the most hard-to-change factors are reset. In this presentation, we relax this constraint and present a new type of design which allows the hard-to-change factor levels to be reset at entirely different points in time. We show that the new designs are cost-efficient and that they outperform split-plot and split-split-plot designs in terms of statistical efficiency. Friday, June 6 Title: Inference for the Step-Stress Model with Lagged
Effects Abstract: We consider models for experiments in which
the stress levels are altered at intermediate stages during the observational
period. These experiments, referred to as step-stress tests, belong to
the class of accelerated models that are extensively used in reliability
and life-testing applications. Models for step-stress tests have largely
relied on the cumulative exposure model (CEM) discussed by Nelson. Unfortunately,
the assumptions of the model are fairly restrictive and quite unreasonable
for applications. We introduce a new step-stress model based on the Weibull
distribution where the hazard function is continuous. We consider a simple
experiment with only two stress level and develop a model that allows
for a lag period before the effects of the change in stress are observed.
Using this formulation in terms of the hazard function, we obtain the
maximum likelihood estimators of the unknown parameters. A Monte Carlo
simulation study is performed to study the behavior of the estimators
obtained for different choices of sample sizes and parameter values. We
analyze a real dataset obtained from an airforce study on altitude decompression
and show that the model provides an excellent fit. This is joint work
with D. Kundu, Department of Mathematics and Statistics, Indian Institute
of Technology Kanpur, India. Abstract: Reliability methods are often used to assess the consistency of different psychiatrists in rating the functional and mental status of their patients. To this end, the intra-class correlation coefficient (ICC) is used as a sensible measure of the inter-rater reliability. The focus here is on an experiment with a binary trait such as diseased/healthy. To conduct such a study, one needs to decide on the number of subjects, n, and the number of raters, k. Although sample size calculation for ICC has been widely studied in the frequentist literature (see Shoukri, Asyali, and Donner, 2004 for a review), there is virtually no literature in the Bayesian paradigm. This work is meant to fill this void and show how sample size calculations for ICC can be addressed from a Bayesian standpoint. The goal is primarily to know how many subjects for a fixed number of raters and how many raters for a fixed number of subjects are required to guarantee an expected width of the credible interval for ICC no larger than a pre-specified value. Optimal allocation of (n,k) for N=nk fixed to conduct a reliability study is also investigated, keeping cost considerations in mind. Title: Methodologies for Recurrent Event Data under
Competing Risks Abstract: The focus of the talk is recurrent events for which the relevant data comprise successive event times for a recurrent phenomenon along with a event-count indicator. Situations in which individuals or systems in some population experience recurrent events are quite common in areas such as manufacturing, software debugging, risk analysis, and clinical trials. While the different application areas dealing with recurrent data enjoy certain commonalities, often the issues specific to an area warrant a non-standard modification of the basic methodologies. In this talk, I shall discuss competing risks analysis in the context of recurrent failures that are subjected to multiple failure modes. Models and methodologies for analyzing a single as well as multiple clusters of recurrent events will be discussed in detail. The situation dealing with missing or masked cause of failure will further be addressed. Friday, June 6 Title: Using Statistical Process Control for Change
Detection in Network Systems Abstract: The occurrence of change in data centers hosting servers and applications is a common phenomenon. The detection of when change has occurred often eludes datacenter operators due to a lack of process or agility requirement of enterprises. The current work proposes an on-line mechanism where real-time network data is compared with historical behavior to determine if a change has occurred in the environment. The detection scheme is performed in terms of quality control parameters namely number of outliers and distance measure between the data and it's dynamically calculated threshold. Dynamic threshold (DT) defines the historical upper and lower bounds of the data which the current method uses to determine data-to-DT interdependence.
Abstract: Classical statistical process control (SPC) is focused on applications where a certain target value for the quality characteristic is available, or where the process is stationary. However, those assumptions are often not met and -often customized- adaptations to the classical concept of SPC are needed. In this contribution, several practical examples from industry will be given where such adaptations are required. Starting from a univariate case of monitoring livestock, the talk will focus on monitoring very high dimensional data (e.g. power spectral densities, near infrared spectra, .) that possibly exhibit a non-stationary behavior. Special attention will be paid to the use of data reduction techniques (supervised and non-supervised) such as principal components –including their recursive versions- for multivariate SPC.
Abstract: The Repeated Two-Sample Rank (RTR) procedure
is a nonparametric statistical process control methodology that applies
to both univariate and multivariate data. The method transforms a sample
of data into univariate statistics; changes in the distribution of the
data are then detected using nonparametric rank tests. In this discussion
we explore its use as a spatio-temporal event detection and monitoring
methodology for use in applications such as biosurveillance, crime mapping,
or IED incidence change detection. The methodology is designed to sequentially
incorporate information from individual observations as they arrive into
an automated systems and thus can operate on data in real-time. Upon a
signal of a possible distributional change, the methodology suggests a
way to graphically indicate the likely location of the distributional
change. Abstract: We adapt the classic cusum change-point detection algorithm for Applications to data network monitoring where various and numerous performance and reliability metrics are available to aid with early identification of realized or impending failures. Three significant challenges that must be overcome are: 1) the need for a nonparametric technique so that a wide variety of metrics (including discrete metrics) can be included in the monitoring process, 2) the need to handle time varying distributions for the metrics that reflect natural cycles in work load and traffic patterns, and 3) the need to be computationally efficient with data processing of the massive number of metrics that are available from IT environments. Friday, June 6 Title: A Discussion of Various Ways to Analyze Data
From Response Surface Split-Plot Experiments Abstract: Split-plot and other multi-stratum structures are widely used in factorial and response surface experiments and residual maximum likelihood (REML) and generalized least squares (GLS) estimation is seen as the state-of-the-art method of data analysis for nonorthogonal designs. We analyze data from an experiment run to study the effects of five process factors on the drying rate for freeze dried coffee and find that the main-plot variance component is estimated to be zero. We show that this is a typical property of REML-GLS estimation which is highly undesirable and can give misleading conclusions. We also investigate approaches that require replication for estimation the variance components. However, we recommend a Bayesian analysis, using an informative prior distribution for the main-plot variance component and implemented using Markov chain Monte Carlo sampling. Paradoxically, the Bayesian analysis is less dependent on prior assumptions than the REML-GLS analysis. Bayesian analyses of the coffee freeze drying data give more realistic conclusions than REML-GLS analysis, providing support for our recommendation.
Abstract: This talk will address a problem from semiconductor manufacturing where the response is non-normal and there is a random effect representing an extraneous source of variation. Levels of the control variables are to be chosen that minimize the variation arising from the random effect when the mean of the response is constrained to a target value. A generalized linear mixed model is used to develop a response surface approach to address this problem.
Abstract: One traditional criterion for comparing response
surface designs is the size of the prediction variance. Another criterion
is based on utilizing the power of an F-test concerning the fixed unknown
parameters in the associated model. The purpose of this talk is to compare
response surface designs on the basis of the prediction variance and power
criteria when the model contains a random block effect, in addition to
the fixed polynomial portion of the model. The proposed approach uses
quantiles of these design criteria on concentric surfaces, either inside
the experimental region or within the so-called alternative space associated
with the unknown parameters. The dependence of these quantiles on the
unknown value of the ratio of two variance components, namely, the ones
for the block effect and the experimental error, is depicted by plotting
the so-called quantile dispersion graphs. (This is a joint work with André
I. Khuri) Title: User Preference and Search Engine Latency Abstract: Industry research advocates a 4-second rule for web pages to load. Usability engineers note that a response time over 1 second may interrupt a user's flow of thought. There is a general belief that all other factors equal, users will abandon a slow search engine in favor of a faster alternative. This study compares two mock search engines that differ only in branding (generic color scheme) and latency (fast vs. slow). The fast latency was fixed at 250 ms, while 4 different slow latencies were evaluated: 2s, 3s, 4s, and 5s. When the slower search engine latency is 5 seconds, users state they perceive the faster engine as faster. When the slower search engine latency is 4 or 5 seconds, users choose to use the faster engine more often. Based on pooling data for 2 s and 3 s, once slow latency exceeds 3 seconds, users are 1.5 times as likely to choose the faster engine.
Abstract: In on-demand computing services, a customer pays for what they actually use and the service provider is free to reallocate unused capacity to other customers. An important issue in such a shared environment is optimal workload consolidation and capacity planning. In this talk, a quantile-based statistical approach is proposed for analyzing the benefits and risks of workload consolidation. An optimization problem is formulated for workload consolidation with the aim of maximizing the benefit while keeping the risk at an acceptable level.
Abstract: Forecasting and simulation of data on internet activity, and communications networks in general, are complicated by nonlinear variability. At high frequencies, time series from the world-wide web and communications networks exhibit heavy tails and irregular bursts of activity characterized by intermittent outliers. At lower resolutions, they also exhibit underlying signals. Empirical tests were run on call series at Sprint-Nextel and information
downloads from company websites. At high frequencies, evidence of heavy
tails is confirmed, while the fractal dimension is non-integer, indicating
a high probability of extreme events. At lower frequencies, the nonlinear
variability dissipates, leaving underlying signals, specifically a stochastic
trend and 24-hour and 7-day cycles.
|