Overview of all Abstracts

The following PDF contains the Abstractbook as it will be handed out at the conference. It is only here for browsing and maybe later reference. All abstracts as PDF

My abstracts

 

The following abstracts have been accepted for this event:

  • A note on smart alarming methods

    Authors: Laurent Bordes, Simplice Dossou-Gbété, Jean-Paul Valois (Laboratoire de Mathématiques Appliquées, France)
    Primary area of focus / application:
    Submitted at 22-Jun-2007 16:33 by
    Accepted
    Methods of Smart Alarming aim at timely novelty or anomaly detection in Data Streams. A review is proposed to highlight the key points of using them. In case of univariate data, the more suitable method is not the same of stationary variable or non-stationary variable. Multivariate data set are often dealt with unsupervised learning based methods, using either factor analysis (mostly PCA) or clustering algorithms. Each of these methods must be applied in a specific situation: the possible anomalies can be prior perfectly known or not, learning data set can be large sized or not, and so on. Some examples are outlined. Discussion underlines the importance to have a prior knowledge of variable behaviour, and to consider the global flow chart, including eventually a data preprocessing.
  • Prediction of Spiralling in BTA Deep-hole-drilling

    Authors: Amor Messaoud, Nils Raabe, Oliver Webber, Dirk Enk, Claus Weihs (University of Dortmund, Dortmund, Germany)
    Primary area of focus / application:
    Submitted at 24-Jun-2007 08:55 by
    Accepted
    Deep-hole drilling methods are used for producing holes with high length-to-diameter ratio, good surface finish and straightness.
    The process is subject to the occurrenec of a dynamic disturbances called spiralling. It leads to multi lobe-shaped deviation of
    the cross section of the hole from absolute roundness which constitutes a significant impairment of the workpiece. A common
    explanation for the occurrence of spiralling is the coincidence of time varying bending eigenfrequencies of the tool with multiples
    of the spindle rotation frequency. In practice, it is necessary that a process monitoring system is devised to predict the occurrence
    of spiralling during drilling. This allows the engineers to know when and how to adjust the process. In this work, the application
    and use of different monitoring strategies are discussed. These strategies are based on control charts in combination with a
    statistical and physical models describing the course of the eigenfrequencies.
  • Taut String as an Alternative to Empirical Distribution Estimators for System Loads in Logistics

    Authors: Sonja Kuhnt (1) and Christoph Schürmann(2)
    Primary area of focus / application:
    Submitted at 24-Jun-2007 09:02 by
    Accepted
    Large systems or networks are frequently analysed in logistics with the aid of simulation models. Meaningful
    conclusions can only be derived if the simulation model is a good image of reality. System loads such as arrival
    times of customers or order quantities are often generated from stochastic distributions. The empirical distribution
    function is the straightforward choice to derive such a distribution from an observed real-life data set if no information
    concerning the underlying model structure is given. As an alternative we consider a distribution estimator based on taut
    string methods. The resulting distribution function generally has fewer knots than the empirical distribution function,
    leading to a reduced simulation effort. We conduct a simulation study to compare the behaviour of empirical distributions
    and taut string estimates with estimates of normal distributions and of distributions from the true family. Distribution
    families relevant for system loads as exponential, normal, uniform and t-distribution are treated. The distance between
    the estimated and true distribution is measured in terms of Kolmogorov-Smirnov and quantile distances.

    Affiliations:

    (1) Department of Mathematics and Computer Science
    Technical University Eindhoven
    Eindhoven, The Netherlands

    (2) Department of Statistics
    University of Dortmund
    Dortmund, Germany

    Specifics: poster presentation
  • Statistical Issues in Search Engine Marketing

    Authors: Steinberg, D.M. (1), Matas, A. (2), Schamroth, Y. (2)
    Primary area of focus / application:
    Submitted at 24-Jun-2007 09:03 by
    Accepted
    Search Engine Marketing - the paid ads you get when you search on the world wide web - is a rapidly expanding economic
    sector. Effective use of SEM requires analytical methodologies that will support decision making. We present here methods
    we have applied and created specifically for this zone of eCommerce, that allow for effective planning of advertising
    campaigns, based on the results from previous adverts. These methods include modeling the effect of a particular bid
    on Cost-per-Click, Position, Click-through-rate, Impressions, Conversions and ultimately Profit. Once modeled, these
    relationships are used to optimize profit at the keyword level. We describe challenges that arise in these efforts,
    particularly those stemming from the dynamically changing environment of competitor behavior, and the corresponding
    solutions we have developed. We apply multiple inference techniques to quantify estimated error allowing for accurate
    decision making per keyword. Finally, we present operations research solutions allowing a particular campaign to be
    globally optimized.

    Affiliations:

    (1) Department of Statistics & OR, Tel Aviv University

    (2) Media Boost, LTD, Israel
  • Phase I statistical control of key indicators in health care

    Authors: Stefano Barone (1), Arturo Caranna (2), Valeria Fonti (3)
    Primary area of focus / application:
    Submitted at 24-Jun-2007 13:57 by
    Accepted
    A current challenge health care providers are facing all over the world is the definition of adequate metrics to
    constantly measure, monitor and control the quality of their services and related costs. The identification of
    quantitative indicators for the key processes is a first necessary step. In a preceding work, the authors presented
    an approach to define a set of indicators and its successful implementation at ''IRCCS Oasi Maria SS.'', a no-profit
    organization for care and research on mental retardation and brain ageing.

    In this article the first results of the statistical analysis of indicators data is presented. Collected data concern
    two-years of activity. They were analyzed by the perspective of statistical process control. Particularly, since they
    were the first analyzed data, they were used to set up the statistical control framework (phase I).

    The main output is a restricted set of control charts, giving top management the possibility to monitor on a
    continuous basis the vital processes of the organization, having effect on the quality of provided care, related
    services and customer (patient/family) satisfaction.

    Keywords:
    Performance indicators in heath care, phase I statistical process control, statistical monitoring, key indicators.

    Affiliations:

    (1) Assistant Professor, University of Palermo, stbarone@dtpm.unipa.it,
    Dipartimento di Tecnologia Meccanica, Produzione e Ingegneria Gestionale,
    Viale delle Scienze, Palermo - Italy

    (2) Dr., IRCCS Oasi Maria SS., acaranna@oasi.en.it, Via Conte Ruggero, Troina - Italy

    (3) PhD. Student, University of Palermo, vfonti@dtpm.unipa.it,
    Dipartimento di Tecnologia Meccanica, Produzione e Ingegneria Gestionale,
    Viale delle Scienze, Palermo - Italy
  • Spiralling in BTA Deep-Hole-Drilling: Combining Statistical and Physical Models

    Authors: Nils Raabe, Dirk Enk, Oliver Webber, Claus Weihs, and Dirk Biermann (University of Dortmund, Dortmund, Germany)
    Primary area of focus / application:
    Submitted at 24-Jun-2007 13:59 by
    Accepted
    One serious problem in deep-hole drilling is the formation of a dynamic
    disturbance called spiralling. Spiralling can be explained by the
    convergence of bending eigenfrequencies with multiples of the rotational
    frequency. In former work we therefore proposed a combination of
    statistical and physical models describing the course of the
    eigenfrequencies.

    In our current work we discuss how this model can be used to predict
    the occurrence of spiralling. As recent experiments showed the
    crossing of multiples of the rotational frequency and bending
    eigenfrequencies yields a resonance effect on the specific multiple.
    By connecting the amplitude of this resonance to quality
    measurements of the workpiece this amplitude can be used as a
    quantification of spiralling. This transforms the binary problem of
    the decision between 'spiralling' and 'no spiralling' into a
    continuous problem and opens up the opportunity of a more explicit
    investigation of spiralling and its impact on the workpiece.
  • Statistical Modelling of Springback Simulation

    Authors: Hilke Kracker (1), Marco Gösling (2)
    Primary area of focus / application:
    Submitted at 24-Jun-2007 14:02 by
    Accepted
    In
    many
    engineering applications physical processes can be simulated by complex
    computer models also called computer experiments. Deep drawing is such
    an
    example, where forming and springback can be simulated using finite
    element
    analysis.

    Such simulations are usually very complex: they are based on many
    input parameters and calculation of a response can be
    time-consuming. Hence, even with a simulation at hand we do not
    entirely understand the input – output relationship at once.

    An approach to gain more insight into the input – output
    relationship is to build a statistical model based on a restricted
    number of evaluations of the computer code. In this talk we will
    compare the adequacy of different statistical methods for
    approximating the springback simulation for a class of workpiece
    geometries. We discuss the adequacy of the statistical models in
    terms of their approximation accuracy and interpretability.

    Affiliations:

    (1) University of Dortmund,
    Department of Statistics, 44221 Dortmund,
    Germany

    (2) University of Dortmund,
    Institute of Forming Technology and Lightweight Construction, 44221
    Dortmund,
    Germany
  • Residual Analysis in Experimental Design

    Authors: Janet Godolphin (University of Surrey, Guildford, Great Britain)
    Primary area of focus / application:
    Submitted at 24-Jun-2007 14:15 by
    Accepted
    Engineers are well aware that standard experimental analyses depend upon various model assumptions about the data to which they relate concerning such key issues as experimental units, factors, blocking and response variable. Violations of model assumptions are usually investigated by plotting and examining least squares and standardised residuals.
    Despite the success of these diagnostic procedures, however, it is sometimes difficult to obtain exact tests of critical model assumptions because, unlike the true errors, least squares residuals are correlated and have unequal variances. In this presentation it is shown how to define and use a set of uncorrelated residuals with a common variance to examine these assumptions.
    Such techniques are demonstrated to be of particular value in industrial experimentation when the nature of the experimental runs means that observations are indexed by time. It is suggested that the resulting test procedures should complement the usual methods. The construction of this set of uncorrelated residuals is established and shown to be straightforward for practical use when checking for model adequacy.

    Specifics: Author prefers to give a talk.
  • Use of Experimental Design to analyze a Neck Forming Process

    Authors: J. Kunert , E. Tekkaya , L. Kwiatkowski , O. Melsheimer and S. Straatmann (University of Dortmund, Dortmund, Germany)
    Primary area of focus / application:
    Submitted at 24-Jun-2007 20:20 by
    Accepted
    Neck Forming is a production process for achieving a diameter reduction in cylindrical bodies. The process depends on many different factors and is - at least at the state of knowledge - too complex for analytical description of the basic forming mechanisms. We have to do physical experiments because simulation of incremental forming processes like Neck Forming is very time-consuming. With these experiments we investigate the main structure of the process.

    The approach we chose starts with the identification of signifant influencing variables using a fractional factorial experiment. For process optimization and robustification we apply methods which have already proved their usefulness for another incremental forming process, namely sheet metal spinning.

    In this talk, we present the methodology with the help of an example. In this example we try to neck-form straight bead welded steel pipes in a inner range to achieve one given geometry.
  • Improving alarm systems by classification

    Authors: Wiebke Sieben (University of Dortmund, Dortmund, Germany)
    Primary area of focus / application:
    Submitted at 25-Jun-2007 10:08 by
    Accepted
    False alarms are a problem of many monitoring systems, especially in
    intensive care. In situations were classical process control methods
    cannot be applied, existing alarm systems can be improved by
    classification procedures. This is the case when no "in control state"
    exists or when the process to be monitored is high dimensional, complex
    and possibly autocorrelated. Annotations to the existing alarm system
    that contain an expert's opinion whether an alarm is considered as
    "true" or "false" can be used as input for data driven alarm rule
    generation. We study the use of ensembles of decision trees as
    classifiers for this problem and at the same time take the unequally
    severe consequences of misclassifying true as false alarms and false as
    true alarms into account. A procedure based on the analogy of this
    classification problem to statistical testing is presented and applied
    to real data.

    The data comes from a standard monitoring system at an intensive care
    unit. So far, the alarms, mostly based on univariate signals, are
    triggered when a physiological variable crosses a preset threshold.
    These standard monitoring systems are known to produce a high number of
    false alarms that distract and annoy the care givers. With our new
    procedure, the expected sensitivity of the resulting alarm system can be
    adjusted to the monitoring environment. This is demonstrated for
    sensitivities of 95 percent and 98 percent for which a false alarm
    reduction by 46% and 30% is achieved on average.