Workshop on Statistics in the Food and Chemical Industry.
DEINDE (DEsign of INDustrial Experiments) 2001

Location: Villa Gualino , Viale Settimo Severo, No. 63 10133 Turin, Italy. Tel: +390116603555, Fax: +390116600082

Program

February 26th 20001

11: 00

OPENING ADDRESS

 

Giovanni Pistone, Politecnico di Torino

11: 30

Isotonic regression: another look at the change point problem.

 

Michael Woodroofe, The University of Michigan, USA

 

Based on isotonic regression, a test for monotonic trends in short range dependent sequences is developed. This test provides another perspective for change point problems which appear, for example, in statistical process control. The isotonic test is shown to be more powerful than some existing tests for trend.

12: 30

LUNCH

 

SESSION I: FOOD

Chair: Mauro Gasparini, Politecnico di Torino

14: 30

Response Surface Modelling of a Pastry Dough Mixing Process.

 

Steven Gilmour, Queen Mary, University of London, United Kingdom

 

An experimental study was conducted on a food extrusion cooker which was used for mixing pastry dough before baking. The ultimate objective was to be able to control the process in order to produce dough with the desired quality characteristics. Three experimental factors were varied: the feed flow rate, the initial moisture content and the screw speed. Day to day variation in some of the responses was expected and, since only four experimental runs could be performed each day, the runs within a day were used to define a block. A modified three-level central composite design, arranged in seven blocks, each of four runs, was used for the experiment. Nine response variables, three relating to the shape of the baked pastry, three relating to its strength and three relating to its colour, were measured on each experimental run. These response variables were analysed by fitting second order polynomial models, allowing for the blocks as a random effect, and in some cases simplifying the models. The fitted models were expressed in canonical form in order to aid understanding of how to control the process. The statistical issues which arose from the practical problems, and how they were dealt with, are discussed. The importance of developing an experimental design appropriate for the objectives of the investigation is emphasised and the utility of the canonical forum of the fitted surfaces is illustrated.

15: 30

Using Designed Experiments in a European-Wide Project Investigating the Sensory Evaluation of Butter Quality.

 

Dave Stewardson, University of Newcastle, United Kingdom

 

This paper describes the use of orthogonal and part-orthogonal arrays within the context of a sensory analysis based research project investigating the homogeneity of a European butter quality standard. The project involved panels of sensory assessors in nine laboratories throughout Europe following the standard used to assess the quality of butter, when submitted for purchase under the common agricultural policy price support mechanism. The main objective was to establish whether results are equivalent Lab to Lab and how results from panels of assessors vary under a variety of conditions. This was the first sensory analysis based cross-nation inter-laboratory study ever conducted in the dairy industry. The paper describes how a factorial design was used to assess the effect of different parts of a butter block, and if butter type and origin were also important, prior to conducting the main trials. Later designs concentrated on the significance of defined butter defects on results and the benefits of a pre-trial calibration sample. It is shown that the factorial approach is able to pinpoint important factors in the sensory analysis context. It was also found to be useful to present statistical information in the graphical form more commonly found in manufacturing. The significance of the five-point scoring method used in the current standard, and also widely used elsewhere, was also investigated. One problem established by the work that has not apparently been addressed by others is the greater uncertainty and consequent increased scatter between results seen under conditions of marginal quality. This may cast doubt on any scoring method involving more than two categories unless the extra variance prevalent in the critical part of the range of available scores is handled within the analysis. A simple pass-fail criterion may be better when assessing food quality in this way. This work described in this abstract was performed during a project that was part-financed by the European Commission under The Standards Measurements and Testing Programme, contract number SMT4-CT98-2221

16: 30

COFFEE BREAK

17: 00

The use of Statistics and experimental design in food research at Nestlé.

 

Marcel Baumgartner, Nestlé, Switzerland
 

Nestlé has developed over several decades a world wide Research & Development network, employing several thousand people. Data produced in food research are very often analyzed using statistical methods. Furthermore, the usefulness of planned design of experiments has been recognized very quickly. The Nestlé Research Center in Lausanne employs a group of 6 members with a strong formation in statistics and mathematics. Its mission is

  • to assist the fellow researcher in the statistical analysis of their result;
  • to plan, together with the subject experts, beforehand experiments efficiently;
  • to develop and adapt statistical methods to the particularities of food research;
  • to develop and give courses to provide basic knowledge about statistics and experimental design to the fellow researchers.
In this talk we will focus on how we use and communicate the methodology of designed experiments in our daily work. We will summarize the content of the courses we prepare for our colleagues. We will also illustrate some type of designs and problems we regularly encounter. Finally, we will discuss some issues about software for experimental designs.

18: 00

A new multivariate approach for the optimisation of the simultaneous distillation-extraction technique for free fatty acids using a face centred cube experimental design: application to Parmigiano-Reggiano cheese.

 

Giovanni Mori, Università di Parma, Italy

 

A face centred cube (FCC) experimental design was used to investigate effects of extraction time, solvent temperature and sample temperature on the simultaneous distillation-extraction (SDE) extraction yields of 13 fatty acids from Parmigiano-Reggiano cheese, and to find the best extraction conditions for these compounds. Each factor was tested at three levels: 20 experiments were carried out, each producing 13 responses. In order to create a mathematical model representing the relationships between factors and responses and to determine the best conditions for the SDE procedure, GC peak areas were processed by using the multiple regression analysis. The backward stepwise selection of the regression coefficients led to the identification of seven groups of compounds having similar behaviour with respect to the three variables investigated. The same set of models was found by performing principal component analysis (PCA) on the t-values of the regression coefficients of the unrefined polynomials and cluster analysis on the scores of the 13 acids obtained by PCA. The best experimental conditions were evaluated for each compoud. The use of the FCC experimental design allowed to evaluate not only the main effects of the variables analysed, but also curvature effects and interaction effects among them.

19: 30

SOCIAL DINNER IN A RESTAURANT DOWNTOWN.

 

February 27th 20001

SESSION II: CHEMICAL

Chair: Jeffrey Eisele, Schering Berlin, Germany

8: 30

On six sigma methods in the production of bromide compounds.

 

Ron Kennet, KPA Ltd, Israel

 

Modern industrial organizations are facing increased competitive pressures and rising customer expectations. Management teams, world wide, are striving to "delight" their customers while simultaneously cutting costs. These challenges require solutions that resolve the apparent conflict between high productivity and high quality. Modern industrial statistics is uniquely positioned to help develop such solutions. Unfortunately many implementation barriers do not permit to uncover these opportunities. The Dead Sea Bromine Group (DSBG) is the world's largest producer of elemental bromine and a recognized leader in the development and supply of a growing family of bromine compounds. Over 3000 customers on five continents rely on DSBG products.Two highly sophisticated installations, one in Israel and the other in the Netherlands, together produce more than 200,000 metric tons of bromine compounds per year. DSBG's revenues in 2000 totalled over $600 million. In this talk we will present the implementation, at DSBG, of a Six Sigma initiative. The approach produced breakthrough results which were specifically recognised as derived from the implementation of statitical methods including Designed Experiments and Statistical Process Control. We will focus on the role of Designed Experiments, the training methods of Black Belts, the Statistical Methods used and the Management Systems that eventually facilitated the process. General insights and reflections on Six Sigma will conclude the talk.

9: 30

Design of experiments in process development from molecule to marketplace.

 

Robin Nicolson, GlaxoSmithKline, United Kingdom

 

The pharmaceutical industry is always seeking new methods to shorten the time required to take candidate compounds from research to full-scale manufacturing, in order to gain a return on investment. Statistically designed experiments (DoE) for laboratory, small scale process development and early manufacturing batches reduce process development and validation times. Such studies are used to develop optimal operating conditions for full scale manufacturing with a high degree of confidence in the robustness of these conditions for routine operation. Complementary programmes of experimentation in development and manufacturing can further reduce the time required to provide optimal processing conditions. General principles and lessons are illustrated by case studies where DoE has been successfully implemented in chemical process development to reduce development times and to improve process performance and costs. Some of the pitfalls and problems are also illustrated, as are some of the barriers that have to be overcome. Experience has shown that in chemical process development; chemists, analysts, process engineers and statisticians working together can help realise benefits to the business by applying DoE. The general principles can be applied and benefits realized more widely.

10: 30

COFFEE BREAK

11: 00

Repeated screening for defective items in a situation with inspection error and no false positive inspections

 

Harald Nusser, Schering Berlin, Germany

 

Repeated screening is a 100% sampling inspection of a batch of items followed by removal of the defective items and further iterations of inspection and removal. The reason for repeating the inspection is that the detection of a defective item may happen with probability p<1, called the sensitivity of the screening. A missed defective is a 'false negative': no false positive are contemplated in this paper, as motivated by a problem coming from the production of pharmaceutical pills. Bayesian posterior distributions for the quality of the lot are obtained for the case of both p known and p unknown. Lot rejection and lot acceptance control limits for the number of defective items at subsequent iterations can then be calculated.

12: 00

Many factors, few runs, awkward constraints and a botched experiment. The real value of good design and analysis.

 

Mark Whitlock, Pfizer, United Kingdom

 

It is unusual for a statistically designed experimental programme, however small, to run as described in most publications on the subject. However, thanks largely due to the robustness of the orthogonal array designs it is nearly always possible to make reliable scientific inferences from such studies. This presentation will show how a mixture of techniques not routinely covered in the standard texts can add great value to their design and analysis. These include: trend robust designs, outlier detection and imputation, Bayesian methods for complex aliased and botched designs, as well as the much neglected area of presenting results for ease of interpretation.

13: 00

LUNCH

14: 30

LEAVE by bus to visit the Martini & Rossi plants and museum.

Back to DEINDE 2001 homepage.