Copyright is owned by the Author of the thesis. Permission is given for a copy to be downloaded by an individual for the purpose of research and private study only. The thesis may not be reproduced elsewhere without the permission of the Author. METHODS TO IDENTIFY, QUANTIFY AND MINIMISE VARIATION OF NET WEIGHTS IN CANNED FOODS DAY ANANDA SIRILAL VITHANAGE A thesis presented in fulfilment of the requirements for the Degree of Master of Technology FACULTY OF ENGINEERING AND TECHNOLOGY MASSEY UNIVERSITY NEW ZEALAND 2006 Methods to identify, quantify and minimise varia tion of net weights in canned foods II Abstract Using a 24 fac torial design model, the methods to identify and quanti fy the major sources of vari ati on of net we ights in canned foods were inves tigated. A piston fill er was selected using modifi ed starch so lution as the fi lling medium. The stroke length and speed of the fill er and the concentration and temperature of the filling medium compri sed the four factors. The data were transfom1ed into means and vari ances of fill weights from both across filling heads and across consecutive fillin g cyc les, and were used as responses. The responses, which were deri ved across filling cyc les fo r each of the filling heads, were used as blocks, to eva luate the head effects. The projecti on des igns were used to optimise variation and fill levels at set piston-stroke leve ls. The fac tor leve l combinations required to minimise vari ation and max imise fill leve l which was computed through a model matri x using all important effects were fo und to be P+,S+,T-,C+ and P+ ,S+,T+C+ respectively. The contributions of fac tors and their interactions to the short-term variance of fill weights were estimated using vari ance across heads w ithin consecuti ve filling cyc les (88 .5%). The analys is across fi lling cyc les within individua l heads estimated the deviation of fi x fac tor leve ls w ithin the trials and contributed to 44%, which appeared as factor effects. Most of thi s variation (52.3%) w hich was caused by the unstable filling mechanisms appeared as the res idual enor. The ana lysis of blocks using heads was successful in partitioning the vari ance due to head differences (3.6%).The high volume operati ons generated a higher contribution fro m unstable fillin g mechani sms to the total variance, and a le ser contribution fro m head diffe rences to the total range of fill weights. The recommendations inc lude methods and materia ls to reduce the enor in the des ign. Future research is recommended in the areas of vacuum and single shot fillers, multi­ filling processes, and particle size vari ation. Methods to identi fy , quanti fy and minimise vari ation of net weights in canned foods 111 Dedication l dedi cate thi s thes is with love and appreciati on to Pi yadasa and Daisy Mary Yithange, my father and mother, and to Dis inahami , my grandmother, fo r their inspiration, guidance and support in shaping my li fe . Methods lo identi fy , quanti fy and min imise va ri ation of nel weights in canned foods IV Acki1ow leclge111e 11ts My supervisors, Dr John Bronlund and Dr Nige l Grigg, not onl y did an exce llent j ob of superv ising the research which was carri ed out at the Heinz Watti e site but a lso stepped out of their line of duti es to rescue and put me back on the right track when the go ing was tough. I deeply apprec iate their inva luable help in accompli shing this task. Working towards a mas ter's degree amidst the pressures of a demanding career has been a di fficult task. I deeply appreciate the se lfl ess and unconditional support of my dear wife, Ajantha Vithanage, in moti va ting me to complete thi s thesis and fo r shouldering most of my share of the household responsibili ties during thi s d ifficul t time. g ive smcere thanks to Anthony Bennet, the manager of project engmeermg, fo r recogni sing the importance of the ro le of DoE in the food-process ing environment of Heinz Wattie Ltd , Hastings and for prov id ing the much-needed funds fo r thi s research. Further, I express my apprec iation and thanks to T ri xie Ackerly for her expert advice in helping to shape up the desktop document numerous times, and to Graham Danes fo r his va luable comments during the rev ising of thi s document. Methods lo ident ify, quantify and minimise variat ion of net weights in canned foods Co11te11ts Contents Tables Figures Glossary of technical terms 1-0 Introduction 2-0 1-1 Scope of research 1-2 Project proposal Literature survey 2-1 Review of statistical methods 2-2 Applications and case studies 2-2.1 Statistical methods in food process improvement 2-2.2 Use of statistical methods 2-2.2. 1 Role of statistical thinking in food industry 2-2.2.2 Benefits of statistical methods 2-2.3 Design of Experiments in variability reduction 2-2.4 DoE as a tool of building process knowledge 2-2.5 Useful statistical techniques in DoE 2-2.6 Applications of DoE in food process variability 2-3 Experimental model of DoE 2-3.1 Tool of experimentation 2-3.1.1 Experimental pattern 2-3. 1.2 Randomi sation 2-3.1.3 Planned grouping (b locking) 2-3.1.4 Replication 2-3.2 Tools of analysis of DoE 2-4 Conclusion of literature survey Methods to identify , quantify and minimise variation of net weights in canned foods Page X XI Xll 3 4 4 4 4 6 6 7 9 11 12 15 19 20 21 21 22 22 23 25 V Contents Contents 3-0 Sources of variability 3-1 Sources and state of control 3-1.1 3-1.2 3-1.3 3- 1.4 3-1.5 3-1.6 Variations related to process dynamics Variation rel ated to product composition Variation related to packaging Variations re lated to filling operations 3- 1.4.1 Vacuum filling operations 3-l.4.2 Piston filling operations 3-1.4.3 Pocketfi llers 3-1 .4.4 Fillers a a contributor to variability 3-1.4.5 Control of fil ler variability Effect of quality control on variation Quantity control 3- 1.6. 1 As an essential function of food proces es 3-1.6.2 Methods of quantity control 3-2 Screening of sources for major factors 3-2.1 3-2.2 3-2.3 improvement focus through partitioning of variations Identifying the factors responsible for net weight variation Time based characteristics of variations 3-2.3.1 Short term variability VI Page 27 27 27 29 29 31 31 36 39 42 44 46 47 47 48 50 50 52 55 55 3-2.3 .2 Long term variability 56 3-2.4 Selection of important factors for the planned experiments 57 3-2.4. 1 Effect of variation of tare-weight of Cans 57 3-2.4.2 Process of factor se lection 58 Methods to identify , quantify and minimise va riation of net weights in canned foods Co11te11ts Contents 4-0 Experimental methods 5.0 4- 1 Experimenta l design approach 4- 1.1 Weaknesses of one factor at a time designs 4-1.2 Avai lable designs 4-1.2.1 Nested designs 4-1 .2.2 Factorial designs 4-1.2 .3 Fractional factorial designs 4-1 .3 Se lection of appropriate design 4-2 Planning for factoria l design and tria ls 4-2.1 Planning fom1 for factorial design 4-3 Methods and material s 4-3 . 1 Production gear 4-3 .2 Product preparation 4-3.3 Trial Runs 4-3.4 Data collection 4-3. 1.1 Collecting response data 4-3 .4.2 Measurement of factor levels Results and analysis 5-1 Methods of translating data for ana lysis 5-1 . 1 Tabulation of trial data 5-1.2 Response types used for ana lyses 5-2 Analysis us ing mean of fi II weights across fi II ing cycles 5-2.1 Mean of fill weights across fi ll ing cycles at High-P 5-2.2 Mean of fill weights across fi ll ing cycles at Low-P 5-3 Ana lys is using mean of fill weights across filler heads 5-3.1 Mean fill of weights across filler heads at High-P 5-3.2 Mean fill of weights across filler heads at Low-P Methods to identify, quantify and minimise variation of net weights in canned foods VII Page 59 59 59 60 60 60 6 1 62 63 63 66 66 66 68 70 70 7 1 73 73 73 79 85 86 87 88 89 90 Contents Contents 5-4 Analysis using variance of till weights across tilling cycles V III Page 9 1 5-4.1 Variance of ti ll weights across til ling cycles at High-P 92 5-4.2 Variance of fill we ights across tilling cycles at Low-P 93 5-5 Analysis using variance of fi ll weights across ti ller heads 5-5. 1 Variances across filler heads at High-P 5-5.2 Variances across filler heads at Low-P 6-0 Discussion of results 6- 1 Analysis of the planned experiments using mean fi ll weights 6-1.2 Projection of designs with mean till weights 6- 1.2.1 Projection designs (23 ) across tilling cycles and across fi ller heads 6-2 Analysis of the planned experiments using variances as responses 6-2. 1 Analysis for variance of ti ll weights across filling cycles 6-2. I. I The contribution from the unstable tilling mechanics 6-2.2 Analysis using variance of til l weights across tiller heads 6-2.3 Factor level combinations required to minimise variation 6-3 Partitioning of variability due to ti lling Heads 7.0 Application of results 7- 1 Relevance to the existing food process 7- 1.2 Applicabi lity of the experimental model 7-2 Combinations required fo r least variance 7-3 Long and short-term variability of product processes 7-4 Application of DoE in different ti lling and measuring environments Methods to identify, quantify and minim ise varia tion of net weights in canned foods 94 95 95 96 96 99 99 100 100 102 103 105 107 109 109 110 113 115 117 IX Contents Contents Page 80 Summary and conclusions 119 8- 1 Summary of investigation 119 8-2 Main findings 121 8-3 Applications 125 8-4 Recommendations 126 8-5 Areas for further research 127 References 128 Appendices 133 Methods to identify, quantify and minimise variation of net weights in canned foods X List of Tables Table D escription Page Table 3-1 Fill variations of vari ous particle sizes using pocket Filler 41 Table 4 . 1 Des ign matri x for 24 factoria l des ign using selected facto rs 62 Table 4.2 Tabular plan for filling process using starch so lution 67 Table 4-3 Trial 1314 b: gross fill weights, means and variances across filling cycles and heads 71 Table 4-4 Factor-level measurements 72 Table 5- 1 Mean fill weights and variances across filling heads 74 Table 5.2 Mean fil l weights & vari ances across fillin g cyc les 75 Table 5-3 Schedule of planned analyses 83 Table 5-4 Summary analysis of trial responses using Mini tab output 84 Table 6-1 Modeling of factor combination fo r fill we ight optimisation 98 Table 6-2 Modeling of factor combinat ion for variance optimisation 106 Table 6-3 Partitioning head variation using range method 108 Table 7-1 Component vari ation of three different Beef Curry types I I I Table 7-2 Compari son of direct phys ical effects versus actual process effects of factors on fill we ights 114 Methods lo identify, quanti fy and minimise variation of net weights in canned foods XI List of Fig1.u·es }...,ioure ~ Description Page Fig. 2-1 Use of statistical methods within responding organisations 8 Fig. 2-2 Three major sources of variation in industries (Crow, 2006) 9 Fig. 2-3 Concept of robust design (Crow, 2006) 10 Fig. 2-4 General model of a process 19 Fig. 3-1 The can seal system comprising rubber gasket and open flange 32 Fig. 3-2 Four major areas of a piston filler operating cycle 37 Fig. 3-3 Cross sections of piton strokes during four major operations 38 Fig. 3-4 Cross-section sketch of a pocket filler 41 Fig. 3-5 Fluid dynamics during pi ston and vacuum filling 43 Fig. 3-6 Head analysis using fill weights agai nst each filler head for filler R2 45 Fig. 3-7 Potential for net weight savings based on variation reduction 48 Fig. 3-8 Cause and Effect diagram for sources of net weight variation in cans 53 Fig. 3-9 Hierarchy of variables in product filling process 54 Fig. 4-1 Can transfer from star wheel to filling station showing the can Labelling method 69 Fig. 5- 1 Run Charts for mean (Original) fill-weights across filling cyc les for each filler head 76 Fig. 5-2 Run Charts for mean (Original) fill-weights across filler heads for each cyc le 77 Fig. 5-3 Run Charts for variance of (original) fill-weights across ft I I er-heads 78 Fig. 5-4 Transfom1ation of filler trial data into four main response variables 80 Fig. 6-1 Projection of 24 into two rep li cates of23 designs using F4 99 Fig. 6-2 Fluctuation of Fill Weights with in each head during filling cycles 101 Fig. 7- 1 Effect of temperature on ft 11 variations 116 Methods to identify , quantify and minimise va ri ation of net weights in canned foods XII Glossa1•y of N otations ANOM ANOVA AQS Brix DIKT Do E HACC P !QR LQ LSL LSM MFw-Ac-Fc MFw-Ac-Fh PPOSE PCA PCOSE PCR PLC Pre-ft II SD SKU SPC SSB SSW TQC TQM UQ USL Var-Ac-Fe Var-Ac-Fh Analys is of Means Analys is of Variance Average Quantity System Percent of soluble solids on w/w basis on sugar scale Data-Information-Knowledge-Technology Design of Experiments (Planned Experiments) Hazard and Criti ca l Control Points Inter Quartil e Range Lower quartile Lower Specification Limi ts Least Square Means Mean Fill weights Across Filling cycles Mean Fill weights Across Filler heads Normal Probabili ty Plot of the Standardized Effects The principal component analys is Pareto Chart of the Standardized Effects Process capability ratio Programmable Logic Control Any filling components invo lved prior to final fi lling Standard Deviation Stock Keeping Unit Stati stica l Process Control Sum of Squares Between Samples Sum of Squares Within Samples Total Quality Contro l Total Quality Management Upper quartil e Upper Specificati on Limits Variance of Fill we ights Across Filling cycles Vari ance of Fill weights Across Filling heads Methods to identi fy , quanti fy and minimise varia ti on of net weights in canned foods Chapter I Introduction I INTRODUCTION Chapter 1- I SCOPE OF RESEARCH Heinz Wattie's Ltd. is a food company which manufactures about 1200 food products in the form of cans, pouches and frozen packs. In 1993, the author introduced Statistical Process Control (SPC), replacing the existing methods of control. The existing methods specified control of individual weights of cans within specification limits. The initial stage of SPC (Statistical Process Control) involves the controlling of process averages within control limits. During the next stage, the variabi lity of some processes was reduced by taking corrective actions for special causes. A total of over NZ$ I .Sm worth of "give-away" has been recovered. The proportion of time for which the process is in control has increased from 23% to 62% (Vithanage, 1994, 68). In spite of the initial success and ongoing improvements, the SPC programme was faci ng ever-increasing challenges from the fo llowing areas. • Incapable production processes due to high variabi li ty of quantity and related quality parameters • Control of weights while maintaining product safety and can integrity • Control of weights whi le conserving product quality characteri stics • Expectations from overseas markets fo r tighter weight perfo1m ance • Increased consumer vigilance and meeting the regulatory requ irements of different countries Methods to identify, quanti fy and minimise variation of net weights in canned foods Chapter I Introduction 2 A significant amount of direct and indirect saving wi ll be made by meeting some of these challenges. For example, the cost of non-standard products due to under- and over-filling is around $ 500,000 per annum (Vithanage, 2003). The majority of quantity and quality related Nonstandard Products are due to the increased variations of output process vari ables. Although the total output variabi li ty could be reduced to some extent, a planned and systematic reduction was not possible for following reasons: a) Lack of tools to identi fy and quantify the sources of this variability b) Lack of knowledge of the optimum combination of the input variables, required to reduce the variation of net contents in canned products Therefore, this study was aimed at deve loping methods to identify and quantify the contributing va riations to the total variability of net weights. These process variables, which appear in the fonn of quali ty and other process parameters, would be identified and used as factors in the Design of Experiment. Appendix 12 illustrates how the major sources, the factors within each of the major sources and the interactions among major sources, contribute to final and measurable net content variation in canned foods using the Japanese Beef Curry Process as an example. Methods to identify, quanti fy and minimise variation of net weights in canned foods Chapter I Introduction 3 1-2 PROJECT PROPOSALS The overall goal of the project was to develop methods to minimise the variation of net weights in canned foods. The following two tasks were used to achieve this goal. a) Identify and quantify the process variables, which contribute to net weight variab ility in canned foods using a liquid medium representing the key variables. b) Develop methods to determine the optimum combination required of the contributing factors to minimise the variation of total net contents. The optimum combination of the variables to reduce net weights variation, once quantified, provides a logica l basis to optimise weights to reduce "give-away" and underweights. These objectives were achieved by studying a representative medium through a se lected filler. The infom1ation was then used to develop methodologies which are applicable to any product fill situation . The scoping of this study was therefore confined to the filler and product variability aspects represented in Appendix 12. Project Stages I) Process and Literature Reviews 2) Project approva l from Heinz Wattie Ltd. 3) Screening of factors for planned experiment 4) Planned Experimentation 5) Data Collection 6) Analysis 7) Submission of Thesis Methods to identi fy , quantify and minimise vari a tion of net weights in canned foods Chapter 2 Literature Survey LITERATURE 2 SURVEY 2-1 REVIEW OF ST A TISTICAL METHODS A literature survey was carried out to establish available statistica l methods to reduce variability and to optimise processes in the food industry. The li terature was reviewed to provide a genera l review of methods applied, the benefi ts of statistical methods, the role of Design of Experiments, building process knowledge using DoE as a tool, useful statistical techniques in DoE, and the application of DoE in reducing food process variability. Finally, the methodology of DoE was summarised as a tool for experimenting with multivariate processes and as a tool of analysis. 2-2 APPLICATIONS AND CASE STUDIES 2-2.1 STATISTICAL METHODS IN FOOD PROCESS IMPROVEMEN T Increased competition in the international market-place has prompted a revolution in quality control systems since 1980. Up to this period, relati vely little was published on the application of SPC in the food industry. They included the use of control charts to monitor food processes, show trend analys is and for obtaining warning signals to control the process as the key statistical applications. Among examples of specific applications was the use of SPC in HACCP (Hazard and Critica l Control Points) analysis, automated SPC monitoring in extruded products, use of control charts with suboptimal sampling in breweries and variance components applications in the biscuit industry. In recent times, Methods lo identify, quantify and minimise variation of net weights in canned foods 4 Chapter 2 Literature Survey the use of sta tistical methods such as multivariate analysis, regression analysis, time series and non-parametric methods has become more common. (Srikaeo el al., 2005, 309-317) In a paper dealing with industrial applications of Statistica l Quality Control (SQC) (Grigg, 2005) provided a summary of academic articles which has been classified to show the subject and relevant industry product or sector. The key quality criteria were food safety and weight control and the industrial sectors involved were dairy, biscuit manufacture, meat, and poultry and drinks production. The lead-time required for the successful launch of a product in the market-place can be greatly reduced by Design of Experiments. An article published by Arteaga & Peres ( 1994, 242-254), contains details of DoE applications available for optimisation work. Among them, Fractional Factorial Designs including Taguchi methods, Response Surface methodology and Mixture Designs, have been cited as effective. The author has pioneered the use of statistical methods in food process improvements at Heinz Wattie, New Zealand since 1993. The methods used included the fo llowing: I) Control charts for feed-back control 2) Modified acceptance sampling procedures for vendor/customer dealings 3) Challenge testing and validation methods for processing equipment 4) Component analysis for estimating measurement error and improving variation of weights during check weighing 5) Use of correlation, regression and ANOVA techniques to improve specific product quality-related issues Srikaeo wrote about the use of ANOV A and ANOM (Analysis of Means) in conjunction with control charts, as means of identifying significant changes to mean process parameters. His efforts to separate the measurement error component of the observed variance from the true process variation, by the use of variance components methods, was a very useful contribution as this error fac tor has been found to be very significant. (Srikaeo el al., 2005, 309-317) Methods to identify, quantify and minimise variation of net weights in canned foods 5 Chapter 2 Literature Survey 2-2.2 USE OF STATISTICAL 1v!ETHODS 2-2.2.1 Role of Statistical Thinking in Food Industry Bjerke & Hersleth (2001 ,49-59) assert that the core elements of statistical thinking have been recognised as the generation of data, extraction of relevant info1111ation from data and uti lisation of this information for optimal decision making. This paper also promoted the view, which was published by Ishikawa, ( 1985) that statistical thinking should be the basis for Total Quality Management (TQM) or Total Quality Control (TQC). Statistical thinking plays a very important role in changing a company culture to accept statistical methods as an effective tool of process control and improvements. A model developed for Total Quality emphasised three levels of activities: strategic, managerial and operational and the responsibilities which are involved at each level. At the strategic leve l, the basic concepts such as acknowledging the universal presence of variation in processes, the interconnectedness of processes and the fact that reduction of variation improves quality, are established. At managerial levels, systems such as SPC, Design of Experiments and Robust designs are developed in order to bring together the tool kits for the application at the next level below. Although the li st contained a random collection of tools, it is important to notice that most of the tools had origins in Designs of Experiments. These arguments further support the author' s attempts to use methods in Design of Experiments to reduce variabi lity in the net weight process in canned foods . (Snee, 1990, I 16- 121 ) An investigation into the existing difficulties in applying statistical methods to the food industry has identified management involvement, systems for distribution of competence and some aspects of corporate cul ture to be among the main factors. (Bjerke & Hersleth. , 2001 , 49-59) The author of thesis also agrees with the idea (Grigg, 2005, 10) of the existence of organisational filters (barriers) which need to be overcome before the use of statistical methods is effectively established. Grigg further presented a list of requirements for each of the six leve ls, which is required to be met in order to overcome each of these organisational filters. Methods to identify, quantify and minimise variation of net weights in canned foods 6 Chapter 2 Literature Survey Once the quality engineers have identified the factors and the optimum levels of those factors to minimi e variability, the process can be redesigned for robustness. Snee (1990) highlights the importance of statistical thinking to make processes more robust. This derives from the quality principle published by W. Edwards Demings, urging for reduction in variation to improve quality. This form of thinking developed based on identifying, characterising, quantifying, controlling and finally reducing variations to provide opportunities for improvement. (Snee, 1990, 116-121) One of the ways to reduce variation is to design and build processes that are not affected by unknown or uncontrollable variations . Snee used the term rugged or robust to recognise these processes. Design of experiments is an important tool to design rugged processes. A rugged product or process reduces variation , because it becomes insensitive to variations in components of manufacture, method of use and conditions of use . This is important in food processes where the net weights or a specific quality is affected by factors not directly involved in the food formulations but which arise from methods of handling such as fillin g. 2-2.2.2 Benefits of Statistical Methods Grigg (2005) examined the existing statistical applications in the food industry and their benefits as well as the work needed to enhance the use of SQC techniques for effective realisation of those benefits. This paper furth er illustrates the benefits in a flow chart through a hierarchical format , which was reported using the research questionnaire (Figure 2-1 ). Although the flow chart looked somewhat complex, the list of benefits derived was a definitive proof of the validity of applications of statistical methods. The five main benefits, which appear at the top level of the hi erarchy, were as follows: a) Improved customer and producer confidence in product b) Improved process visibility and understanding c) Improved competitiveness d) Facilitated continuous improvement e) Cost savings Methods lo identify, quantify and minimise va riation of net weights in canned foods 7 Chapter 2 Literature Survey Further down the hierarchy were reduced process variability, enhanced control of product quality and consistency, reduction of waste and reduced giveaway due to overfi ll ing. Use of statistical methods used at all sed sometirrcs sec! extensively ' Advanced ' methods Basic Analy.,is Process average values Process variation ncisures Process defect rates Basic graphs Trend analy.,is / scattCl])lots Statistical Process Control Methods Meanclk'lflS Range Charts standard deviation chm1s defect count chai1S Proponion defect ivc chai1s U lSlUll 0 k'lf1S Moving averages chai1s Multi variate control charts Capability asscssm::111 AcccptMCc sairpling Quality Engi neering / ln-µovcrrent rrethcx.ls process dcsign/itllJrOVcm::nt through cxpcrirrentation QfD Q.rali ty circle techiqucs 00 60 40 20 0 20 40 60 Figure 2.1 Use of statistical methods within responding organizations (Grigg, The work of Srikaeo (Srikaeo et al. , 200S) demon trated the use of SPC techniques to characterize a food process. He has quoted a definiti on for process characteri sation as the "activity required to identi fy the key inputs and outputs of a process, collect data on their behavior, es ti mate the steady state behav ior at optimal operating conditions and Methods to identi fy , quant ify and min imise variation of net weights in canned foods 8 Chapter 2 Literature Survey build models to describe the parameter relationship across the operating range". Although this is similar in context to the definition given by Montgomery (200 l , Section 2-2 .5), the latter highlights the essential fea tures of the characterisation process. Montgomery's definiti on encompassed the identification of process parameters, both controllable and uncontrollable, that affect the response parameter concerned and the designing of an experiment that enables one to estimate the magnitude and the direction of the fac tor effects. The present author agrees with the latter definition and likes to allocate the activity of "the estimati on of steady state behav ior at optimal conditions and modeling" which Srikaeo quoted into the proper des ign of the experiment stage including the analys is. This acti vity would be more effecti ve in my view, if done with a special emphasis on minimising the process vari ability. 2-2.3 DESIGN OF EXPERIMENTS IN VARIABILITY REDUCTION Crow (2002) describes three sources of vari ation: manufac turing variation , environmental or deterioration vari ation and usage vari ation (Figure 2.2). The product can be des igned to counter these sources of vari ation, depending on the extent to which we are aware of these potential sources of variation. Sources of \/aration Manufacturi ng D?terioration Product Use Product Design Process Des ign Production EJI Countermeas ures poss ible Q Countermeasures imposs ible EEi 0 0 [I] 0 0 Figure 2-.2 Three major sources of variation in industries (Crow, 2006) Methods to identi fy , quanti fy and min imise varia tion of net weights in canned foods 9 Chapter 2 Literature Survey The concept of robust design is based on this. It is achieved primarily by the use of design of experiments to determine which factors (product and process parameters) are most sensitive to variation or noise and which factor level settings (parameter values) minimise the variability in the desired performance parameter (Figure 2.3). Design of Experiments (DOE) can be used to counter all three sources of variation. In addition, other steps can be taken to counter manufacturing variation and usage variation. First, understanding the statistical capability of a process can help to either design within the capabi lity of that process or determine when an improved capability is required. Second, by using SPC, special cause variation can be identified and attacked. Third, over the longer term, process variability from common causes can be systematically reduced through process optimisation, operator training, preventative maintenance, tool monitoring, standardisation of machine settings, climate control, power conditioning, and so on. Fourth, variability in usage can be countered by mistake-proofing, warning labels, easy-to-understand operating manuals and controls, and other, similar, measures. Parameter & Tolerance Design Design of Experiments Process Capability Process Sett ings Figure 2-3 Concept of robust design (Crow, 2006) Variabi lity reduction involves understanding customer needs and developing a product and process design that balances these needs with process capabilities and potential sources of variation. Thus variability reduction is broader than SPC and DoE individually and more proactive than SPC. A robust product is one that works as intended regardless of variation m its manufacturing process, vari ation resulting from deterioration, and variation in use. Methods lo identify, quantify and minimise variation of net weights in canned foods 10 Chapter 2 Li terature Survey Robust des ign can be achieved when the des igner understands these potential sources of vari ation and takes steps to desensitise the product to these potenti al sources of variation . When the operati on of the product or achievement of a performance characteri stic can be mathemati ca lly related to a product or process des ign parameter, optimum product and process des ign parameters can be calcul ated. When these relationships are unknown, des ign of experiments (DoE) can aid in determining these optimum parameter values and , thereby, in deve loping a more robust des ign. Des ign of Ex periments is based on the objective of desensitising a product's perfom1ance characteri sti c(s) to variation in critical product and process design parameters (Crow, 2002). 2-2.4 DoE AS A TOOL OF BUILDING PROCESS KNOWLEDGE A paper publi shed by Grigg & Graham. (2003), provides some va luable guidelines to use organi sational knowledge, which is fo und to be equall y applicable during the intended research. Spec ifi ca ll y, thi s paper dea ls with aspect of building knowledge by the use of dynamics of knowledge crea ti on. Thi s knowledge has been class ifi ed as either ex isting or obtainable through planned experiments. The paper fu11h er examines the deve lopment of process knowledge, which nonnall y ex ists w ithin th e company. The approach described in thi s paper could be effective ly used to challenge and test both the fo rma l and in formal knowledge among company personnel. It also recogni ses the va lue of process know ledge diffused w ithin a manufacturing fa cility simi lar to Heinz Wattie's Ltd . Further, it recogni ses the need to deve lop thi s knowledge to be useful for practical applicati ons using DoE. The di scuss ion on the process of creating process knowledge by using known concepts such as ' knowl edge spiral ', ' knowledge management landscape' and the ' DIKT (data­ information-knowledge-technology) learning process' , is of parti cul ar interest. It further stresses how DoE could provide an effecti ve means of chall enging and testing the competing hypotheses as we ll as ex isting tac it knowledge. Methods to identi fy , quant ify and minimise varia tion of net weights in canned foods 11 Chapter 2 Literature Survey The paper stresses the need for an initial hypothesis that detennines the variables to be included in the screening experiments. The problem is seen as the selection of key variables from a large domain of possible variables. This stage involves tapping into formal and tacit knowledge bases. Quality circles are seen as a means of tapping the tacit knowledge of the work force , with some limitations. 2-2.5 USEFUL STATISTICAL TECHNIQUES IN DOE The process of quantifying variables contributing to the total variability of net weights requires an approach that is more fundamental than complex statistical systems. My approach is to review literature related to designs , in which screening techniques, methods of reducing experimental runs effective ly and simpler but effective analytical techniques have been employed. The suggestion to use simpler techniques is essential; as such techniques could be repetitively applicable as a standard by technologists with low statistical knowledge. One of the major difficulties encountered during planned experiments for multi-stage food processes is the occurrence of outliers. Of the numerous techniques presented, a very basic approach reproduced in the book by Clark and Randal (2004), appeals most. They have defined as an odd value that seems to be lying all by itself away from the main body of data and provided the following equation to identify an outlier. An outlier is an observation a) Above UQ + 1.5 x IQR or b) below LQ - 1.5 x IQR Where UQ and LQ denotes Upper and Lower quartiles and IQR denotes Inter Quartile Range. When the factorials are un-replicated, outliers could affect the analysis . In a different approach, the experimenter does not actua lly have to remove the outliers because the this method helps to keep experimental error rate under control and at the same time Methods lo identi fy, quantify and minimise variation of net weights in canned foods 12 Chapter 2 Literature Survey improves the power to detect active factors. This method involves the combination of the rank transformation of the observation, the Daniel plot and a fo rmal stati sti ca l testing procedure to assess the signi ficance of the effects. (Aguin-e & Peres, 200 I, 63 7-663) It is important to check whether or not a parti cular di stri bution of variables in stati stical applications confirm to nonnal distribution. For example, the quality characteri stics which contro l the beer-brewing proce s were fo und to be non-nonnally di stributed using both Kurtosis and Skewness tests. The skewness measures the symmetry and the Kurtosis measures the tendency to deviate fro m the bell shape of the normal di stribution. The application of transformed parameters using logarithmica ll y transfom1ed vari ables to construct contro l charts was fo und to be effecti ve. (Ozilgen, M., 1998 ,57-60) Gonzalez-Miret et al., (200 I) have used the Kolmogorov-Smirnov-Lilliefors Test to check the nonnality of the transfom1ed data during their validation of microbiological vari ables using univariate and multivari ate stati sti cs. Alternati ve ly, when variables do not fit no1mal di stribution, the non-parametri c Fri edman Test model was used for analysing the effects as an alternative to ANOY A (Gonzalez-M iret et al., 200 1,26 1-268). Split-plot experiments help to reduce the time and cost constra ints as well as the size of the experiments in manufacturing environments having complex and busy schedules. A method of conducting sixteen run des igns using fractional fac tori al and confo unding was described by Kowalski (2002, 399-4 10). One of the simple techniques of obtaining more in formation out of a factor analys is is to co llapse or project the original des ign into another 2k design with fewer variables by dropping one or more original fac tors if those fac tors and their interactions are less significant (Hines et al., 4th ed., 384). Similarly, if one or more fac tors from a fractional fac torial des ign can be dropped the remaining des ign will project into a full fac torial design. The method requires the ana lyst to identi fy the largest estimated effects without regard to sign. If any of the fac tors under study are absent in the effect li st, the analyst can assume that the fac tor has no significant effect or that the factor effect is di storted by error. Consequently, the analyst now could collapse the design into the nex t lower order. (Juran & Gryna, 1974) Methods lo identify , quant ify and min imise va riation of nel weights in canned foods 13 Chapter 2 Literature Survey The present author has experienced difficulties in targeting the process so that all the key responses meet within a desired set of specifications. An article dealing with corre lating multiple responses details the design and analytica l aspects required to achieve this. The procedure involved the modelling of di stributional parameters in terms of experimental factors and finding a factor setting which maximises the probability of being in a specification region (Chiao & Hamada, 2001,451-465). The lead-time required for successful launching of a product in the market place can be greatly reduced by Design of Experiments. An artic le published by Arteaga & Peres ( 1994, 242-254), contains details of DoE applications available for optimisation work. Among the methods presented, Fractional Factorial Designs including Taguchi methods, Response Surface methodology and mixture designs have been cited as effective. It is important to look for corrective action when experiments could not be performed as planned. One of the common etTors is the failure to reset all the factor levels during successive runs, vio lating the assumption of independent observations from run to run. This causes biased estimates of the coefficients in the model during least squares analysis , leading to incorrect conclusions. Methods are avai lab le now to predict (Webb et al., 2004, 1-11) the variance, where the variance derived from the experiment was unrealistic against the expected variance. This paper explains how the failure to reset factor levels in successive runs leads to larger than expected variance and erroneous parameter estimates. Further, the paper provides methods of analysis for this type of experiment and recommendations to conduct DoE when one or more factor levels are not reset. Another common obstacle in industrial experiments is the encountering of factors which cannot be completely randomised due to manufacturing restrictions. For example, complete randomisation of important factors , which affect the net weight variation such as concentration of product, temperature, filler speed and so on, will demand resources not usually available to the experimenter. Methods to identify , quantify and minimise variation of net weights in canned foods 14 Chapter 2 Literature Survey In a paper written about restrictions in complete randomisation , the authors (Arvidsson & Gremyr, 2003, 87-89) present consequences of incomplete randomisation. The paper achieves the objective of developing a pragmatic and user-friendly decision tree. The decision tree helps to select restrictions in complete randomisation. Further, the decision tree shows the importance of understanding the differences between randomisation and resetting, and therefore the consequences for the experimental results. Therefore, the decision tree offers a practica l value in se lecting restrictions. Where we face factors that are hard to change, split-plot designs have become increasingly important. When fac tors cannot be fully randomised, researchers have demonstrated the use of the D-optimal first and second order split-plot designs (Goos & Vandebroek, 2004, 12-26) to overcome the problem. In fact, it has been shown that these designs could perform better in many cases than completely randomised designs. 2-2.6 APPLICATIONS OF DOE IN FOOD PROCESS VARIABILITY A case study of the use of experimental design (Ellekjaer et al., I 996, 29-36), in product development involved a study of the process quality of cheese to improve its sensory quality. A screening experiment was used to identify important variables that affect the sensory quality. Fractional factorial design with resolution IV was used to reduce the number of experimental runs. A OVA and normal plots were used to evaluate the sensory quality due to factors involved. The principal component analysis (PCA) was found to produce similar results and therefore fo und to be important in identifying the improved sensory properti es. The combination of experimental design and multivariate methods is thought to be effective in this situation. The scores of the main principal component constituted the response variables during multivari ate analysis. The design consisted of two levels of the selected seven factors in a 2 7 - 2 fractional factorial design, where main effects and two factor interactions remain un-compounded. ANOV A was used to isolate sensory attributes which could differentiate between samples. Probability plots were used to study the effects of di fferent factors and to isolate signi ficant effects. Multivaria te analys is was found to be Methods to identify, quantify and minimise variation of net weights in canned foods 15 Chapter 2 Li terature Survey more useful than uni vari ate AN OY A for product improvements . Researchers were able to use conditions which have significant effect on quality such as grade of maturity and additi on of dry matter and a coo ling process to improve qua li ty . The use of nested des igns will be considered fo r component analysis of net weight if the des ign can be converted to a nested des ign. We could have two di fferent product temperature within each of the product concentrations and two fi ller speeds w ithin each product tempera tures and so on. Use of a Nested mode l was published by lttzes (200 I, 11 9- 125) , who described how vari ance components were analysed in dry matter content in butter cream. The fac tors consisted of the compound itse lf, sample units chosen and the measurements within the sample. The vari ance of the process is estimated into these components using three control charts fo r averages, the sample-to-sample vari ability and the w ithin-sample vari ability. The Nested model was applied to design the experiments and the ANOV A was used to separate variance components. However, adoption of a net we ight variabili ty mode l into a nested model of analysis must be compared w ith the analysis using factorial designs. Where the analysis proves the mix ture characteri stics of a given product to be a signifi cant fac tor contributing to tota l net weight vari ati on, then more advanced mixture des ign and analysis could prove important. Bjerke at e l. (2000, 22-36) described an applicati on of a more advanced technique to optimi se the food mi xture. Thi s design is ca lled a projection design and takes into account the constra ints among ingredients and is herefore more complex than facto ri al and frac ti onal fac tori al des igns. This paper di scusses and compares the conventional mixture mode l approach with that of the proj ection des ign approach. lt was mentioned that they have used the work published by Box at e l. ( 1990) on thi s methodology as the bas is fo r their work . He concludes that the proj ecti on des ign approach is more useful than mixture model des igns. However, mi xture model des igns were seen as flex ible in the design region and easier to analyse and interpret. The projecti on des ign approach seems to prov ide fractional versions for combined des igns. If the fac tori al des igns provide enough evidence that the recipe mix is a maj or source of vari ati on, such as suspected in the case of White or Demiglace Sauce, then we could apply these techniques to minimise the vari ation. Methods to identi fy , quanti fy and minimise vari ation of net weights in canned foods 16 Chapter 2 Literature Survey Unlike classical fac tori al des igns, the two methods allow for constraints among ingredients, in which ingredients are assumed to sum up to I 00%. The possibili ty of combining mixture vari ables with other vari ables has also been mentioned. However, analysis was fo und to be more complicated and it may be di ffic ult to perform without a sta tistical background. The obj ecti ve of the experiment was to study the effect of different ingredients on the sensory quality of cooked sausages. The ingredients consisted of carrageenan, a type of whey protein, and di ffe rent mi xtures of milk powder, whey protein and sodium caseinate. For the purpose of this paper, only a single sensory quality was selected. The large number of experimental runs required in this experiment was reduced by using the des ign method ava ilabl e in projection des ign methods, rather than using fractional fac torial des igns. Data were also analysed according to projection design methods. The obj ecti ve was to deri ve the regress ion coefficients such as would be obtained from a fractional fac torial des ign in two levels (2 5 - 1 ) . One of the key issues, specifica ll y in the food industry, is the losses that are incurred due to deviation of key quality characteristi cs fro m the desired target. This process could be reversed by reducing the variability, enabling the quality character in question to approach the target value. An application of the Taguchi method in optimising the milling process published by Ghani el a l.,(2004,84-92) provides a good example of creating robustness in processes. The paper presents a case study on the end milling process at three levels. The DoE aims at finding the right combination of milling factors to achieve low cutting force and surface roughness. Although materi als being handled in a milling process are very di ffe rent fro m the food process in our study, the methods applied fit we ll with the intended research in which the reduction of net weight vari ability would lead to signi fica nt reduction of the cost of giveaway and underweights. Much simpler statisti ca l techniques fo und in the Taguchi methods can be used to replace approaches that are more complex. Two major tools are signal lo noise ratio and orthogonal arrays . Signal to noise ratio measures quality with emphasis on variation. The orthogonal arrays can handle many fac tors simultaneously. Taguchi ( 1978 has developed a method for determining the optimum va lues of process vari ables which will minimise the variation in a process whi le keeping a process mean on target. Taguchi 's approach uses statisti ca l des ign of experiments economically to solve specific problems. Methods to ident ify, quanti fy and minim ise variation of net weights in canned foods 17 Chapter 2 Literature Survey It quantifies the fac tor effect helping to identi fy the highest as well as the lowest influencing fac tors. At the same time, it is equall y effective in investigating the effect of multiple fac tors on the process . Ghani (Ghani et al., 2004,84-92) suggests the use of the Pareto A OYA technique, which requires minimum knowledge about ANOV A. It uses a simplified AN OY A in conjunction with the Pareto analysis principle. The signi fi cant fac tors and interactions can be chosen from the left hand side of the Pareto diagram. The Minitab stati stical software, which we intend to use extensively during our ana lysis, is capable of creating a similar diagram called a Pareto Chart of the Standardised Effects. Dr. Don Wheeler, who identi fied the "Four States of a Process", has introduced an analysis based on DoE techniques ca ll ed "Scree Plot Analys is" (Santos & Clegg, 1999). This graphical approach was, in effect, a Pareto diagram similar to the above based on the fac tor sum of squares. The idea was to view the diagram as the cross section of a cliff, identi fy ing the significant fac tors as standing out from the "scree" (rubble), at the foo t of the cli ff. This is a much simpler technique, which avoids the need for calculating vari ances and F-Ratios. Methods to identi fy , quanti fy and min imise variation of net weights in canned foods 18 Chapter 2 Literature urvey 2-3 EXPERIMENT AL MODEL OF DOE The Design of Experiment mode_) as proposed by Montgomery (2nd ed., 199 1, 454) which is illustra ted in Figure2.2, shows its applicability to food manufac turing processes. Similarly, in the process of food manufac ture, input vari ables are transformed into output vari ables, which are measured as quali ty and quantity characteristics. The specific set of output vari ables whi ch characterises the food item during the manufac ture can be grouped according to the uni t operations involved. Fo llowing the trend in the model, some of these input process vari ables are contro llable, while others are uncontrollable. The product and process improvement is usually achieved by improving one or more output responses. They are also ca ll ed dependent variables since they are dependent on input process vari ables. Applying this model to our research theme, the target reduction and associated savings on "give-away" was proposed as the improvement objective and the vari ance and the mean of net weights were identi fied as output process vari ables. X I , X2, X3 ----- Xn (Input Variab les) l l l Process Y l ,Y2,Y3 -------Yn (Output Variables) l Figure 2-4 General model of a process (Montgomery, 2nd ed., 1991, 454) Method of Experimentation Planned experimentation is a test or seri es of tests in which measured changes are made to the input variables of a process and then the corresponding changes occurring in the output variables or responses are measured. Input vari ables that are used during planned Methods lo ident ify , quant ify and minimise variat ion of net weights in canned foods 19 Chapter 2 Literature Survey experimenta ti on are refen-ed to as .factors. Favourable changes to output or response variables are used to identify and quanti fy input process vari ables, fo r process improvement. The aim of des ign of experiments is to identi fy important fac tors and to estimate or quanti fy the effect of those factors on the response vari ables . It is poss ible that during experiments, the effect of background and nui sance variables will be as great as the effect of se lected fac tors on the response variables. The Experimental Des ign must therefore inc lude methods to partit ion the total variati on of the response variable into components due to fac tors, background variables, and nui sance vari ables. Once the process is characterised and effect of fac tors and their interacti ons have been quantified the nex t step in the process is to optimise the important fac tors in order to deli ver the best possible output or response . Depend ing on the nature of the process, the quality engineer may be in terested in either maximum yield in a producti on unit or minimum vari ation of a se lected character. Shewhart (cited in Moen et al. , 1998, 6) has introduced the concept of "degree of belief'to explain the extent to which an ex perimenter has drawn conclusions from an analyti ca l study. The confidence on the effecti veness of the changes implemented as a result of a des igned experimentation constitutes thi s degree of belief. The input variables, the range of their applicati ons, type of DoE, and the too ls of experiments se lected detennine the degree of beli ef requi red. 2-3.1 TOOLS OF EXPERIMENTATION R.A. Fisher (cited in Moen et al., 1998 , 57) has recommended the fo llowing tools to ensure an effective planned experiment. a) Experimental pattern b) Planned grouping c) Randomisati on d) Replication Methods to identi fy , quant ify and mi ni mi se vari ation of net weights in canned foods 20 Chapter 2 Literature Survey 2-3.1 .1 Experimental Pattern This is the plan of fac tor leve ls and combinati ons for the experiment. Thi s plan should ensure that the des ired outpu t responses would be affec ted in a manner necessary to achieve the obj ecti ve. There are severa l types of experimenta l patterns. Factorial design is employed when the multiple fac tor effect needs to be studied. Although it is poss ible to study one fac tor at a time, thi s may lead to incorrect conclusions due to the fo llowing. I) Presence of interactions between facto rs may di stort the effect of a given fac tor on the response, as the response may also depend on other fac tors levels. 2) The study of the effect of one factor at a time takes longer. Nested or hi erarchical des ign deals with the study of factor levels which act within a higher set of fac tor levels in a hi erarchica l order. For example, when head-to-head vari ation of a fi ll er is acting at a higher level then a set of va lves wi thin a given head becomes the second level of fac tors to study. Incomplete designs consist of only a subset of a full y facto ri al or nested design. Composite des igns contain patterns made of combinat ions of fac torial and nested designs. 2-3.1. 2 Randomisation Randomi sati on is used to determine the fo llowing using random numbers: a) The fac tor combinati on of experimental units b) The order of execution of certai n aspects of the tudy. Randomisati on prevents the effect of unknown factors on the response variables. As explained prev iously, these un known fac tors are call ed nuisance .factors. During DoE, fac tors and background variab les are built into the des ign to explore their effect on response variables. Randomisation separates the effect of fac tors and background variables from that of nui sance vari ab les on response variables. Without randomisation, it is not poss ible to mask the effect of nui sance variables on response variables which could lead to the drawing of wrong conclusions. Random number tables should be used to generate numbers fo r the purpose of selection and ordering (Moen et al., 1998, p. 63). Methods lo iden tify, quantify and min imise variat ion of nel weights in canned foods 21 Chapter 2 Literature Survey 2-3.1.3 Planned grouping (blocking) Planned grouping or block des ign is important in manipulating background variation to achieve the obj ecti ves of planned experiments. Objectives include controlling of background vari ation to minimise their effect on response vari ables of interest. The three ways of controlling background variation are as fo llows. (Moen el al., 1998, pp 60-6 1) I) Keep background variables constant 2) Measure and make adjustment to eliminate their effects 3) Use planned grouping to block their effects. 2-3.1.4 Replication Replication involves repetition of se lected parts of an experiment, whi ch is used to gain increased confidence or degree of belief in the outcome of an experiment. Replication is carri ed out in di fferent fo rms, which include: repeating of measurements, use of multiple experiments fo r each factor combination, partial replicati on and complete replication. Replication causes nuisance vari ables to be averaged out and allows the experimenter to study the interaction between the factors. Methods to ident ify, quant ify and min im ise va riation of net weights in canned foods 22 Chapter 2 Literature Survey 2-3.2 THE TOOLS OF ANALYSIS IN DOE The aim of design of experiments is to identi fy important factors and to estimate or quantify the effects of those factors on the response variables. The classical method used to separate effects of these factors is ca ll ed analysis o_f variance (ANOV A). It is a means of looking at the total variation in the data, breaking it into components, and running statistical tests to find out which components influence the experiment (Lorenzen & Anderson, 1993, 3 1) More precise ly, the analys is of variance serves to verify whether data from different sampl es provide sufficient evidence to indicate a difference among the populations (means) from which the samples were taken . When the vari at ion among the sample means is large compared to the within-sample variation, it is apparent that a real difference does exist among population mean . However, when the variation among samples is not too large compared to variation within it is not easy to make such a decision. The method of analysis of variance compares between sample variation (S 8 2 ) with within-sample variation (Sw2 ) to determine whether or not the difference is significant. The Sw2 represents the pooled es timate of the common variance 62 for the samples invo lved. The method invo lves the use of ratio S8 21 Sw2 as a test statistic to test the hypothesis of equal variance. The region above the tabulated va lue in the upper tail of the di stribution thus becomes the reject region . lf the calculated value F = S8 21 Sw2 falls in the reject region it is concluded that at least one of the population means is different from the others. The Regress ion analysis is the other important stati stical technique for estimating parameters which relate a parti cular vari able to another variable or set of variables. This way regression quantifies a relationship between two quantitative variables, where a quantitative variable Y is related to a set of random quantitative variables X. Therefore, during the experiment it is possible to change one factor at a time and observe the effect on the response vari able. With the results obtained, it is also poss ible to see the degree of effect on the response variable on an individual bas is. The next step is to change the factors together in different combinations and observe whether or not such a change Methods lo identify, quanti fy and minimise variation of net weights in canned foods 23 Chapter 2 Literature Survey produces more favourable results. These experiments are called screening experiments. Since the Y variable {Factor) is dependent and predictable based on its relationship to x variables (Responses) it is also tenned as predictant (Juran & Gryna, 1974, 23.97). The ultimate objective of regression analysis is to predict the quantity of dependent variable Y, in relation to independent variables, using a prediction equation derived using the Least Square method. The prediction equation finds a predicted va lue for Y by using estimated coefficients "p0 and " P1. The error of prediction is Y -" y and is called residual. "Y = "Po +" PI x There are some important inferences about constants Po. P1. with which confidence intervals and statistical tests for slope (P 1) and intercept (Po), can be performed. The confidence intervals for these constants reveal the reliability of the prediction equation. The Statistical tests about the slope in the fonn of p1 = 0 are a useful tool to check the predictability of Y using X. The test statistic for the statistical test about the slope, p1 is: If the computed value above exceeds the table value, the Null Hypothesis (P 1 =0) is rejected, to conclude the slope p1 is pos itive and hence variable X is usefu l in predicting Y (Ott & Mendenhall , 1990, 399). The Minitab software, which employs both ANOV A as well as the regression approach, wil l be extensively used in the analysis of factorial designs by this author. In the regression analysis section of the output, the effects of blocking, factors and interactions ("Te1ms") are estimated, along with the respective coefficients of the regression for the slope ("Coef") standard error of the coefficient ("SE Coef'), t-value for the statistical test ("T") and the corresponding significance ("P"). The latter section generates an ANOV A table in which the Degrees of Freedom (OF), Sums of Squares (Seq SS), Means of Squares ("MS"), F-ratio ("F"), and the Level of Significance (P) are tabulated against ma in effects, their interactions, blocks, and the residual error. The "N01mal Probability Plot of the Effects" helps us to determine which te1ms are important. The terms that do not lie on the straight line connecting the points are probably important. Most importantly, the "Pareto Chart of the Effects" is used to identify the order of importance of the effects and interactions and to separate nonsignificant effects (Minitab for Windows, Release 14). Methods to identi fy, quant ify and minimise variation of net weights in canned foods 24 Chapter 2 Literature Survey 2-4 CONCLUSIONS OF LITERATURE SURVEY The literature review section commenced by looking at the type of statistical methods applied and the key process areas published in the literature. It was revealed that apart from simple SPC applications, the methods such as variance component, regression, multivariate and time series analyses were used to improve product processes. The experimental designs used included Factorial Designs, Response Surface methods and mixture designs and combinations of control chart applications with DoE. The discussion on the role of statistical thinking showed how to implement the stati stica l methods in food process improvement within the corporate environment. The discussion broadened into establishing the philosophies such as robust designing, and reduction of process variability as key drivers for process improvement. The results of the surveys published in the literature prove that application of statistical methods, including DoE, has contributed significantly towards process improvements, savings, customer satisfaction and the profitability of food industries. Specifically, use of these applications in process and product capability assessments, and follow-up corrective actions and in process characterisation to improve processes proves to be very useful. The discussion of the role of SPC in DoE demonstrated the use of SPC as a screening test as well as means of controlling and holding the ga ins of DoE. Once the DoE delivers factor level combinations and directions to minimise variation of net weight, the author must not only implement them into relevant processes but must also hold these gains in future production acti vities using the effecti ve control tools in SPC. Holding gains of the results of DoE could be achieved using the methods detailed under building of process knowledge. The discussion shows how we could utilise formal and informal knowledge using techniques such as "DlKT" and verify competing knowledge claims through DoE. The literature provided case studies and some statistical techniques certain to prove invaluable in solving problems arising during experimentations. The reader will observe that some of these techniques had actually been applied during data collection and analysis. The survey presented techniques to remove outliers, methods of checking Methods to identify, quantify and minimise variation of net weights in canned foods 25 Chapter 2 Literature Survey nom1ality assumption , alternatives to ANOVA when data did not fit normality assumption, and split plot designs to reduce time and costs. The projection design concept in particular was applied extensively throughout the analysis to obtain more information for optimisation efforts . The correlation of multiple responses, estimation of correct variance when there was a failure to reset factor levels as planned and the decision tree type approach to restrictions m randomising were among other useful techniques referred to . The case studies provided the author with an insight into how to combine various designs and analysis combinations to find an appropriate pathway in dealing with net weight variability. They included conventional techniques, which commenced with screening experiments and used a combination of PCA and fractional factorial (2 7- 2 ) for the design, and ANOVA and its multivariate version for the analysis . An experiment which employed a nested design using means and averages from control charts as responses provided an another option . The survey further provided a selection of tools to identify and quantify significant factors . They included the Pareto ANOV A technique and Scree Plot analysis. The Pareto ANOV A technique was found to be very effective during net weight analysis to find the optimum combination required to minimise variations. The literature survey provided some useful information, guidelines and problem solving tools to empower the intended research . The review of the DoE model asserted its suitability as an experimental and analytical tool in the proposed research . However, the survey failed to reveal any published work done to reduce variation in net weight (pack weights) through control of contributing factors . Therefore the proposal to build and apply an appropriate research model based on the knowledge presented in this survey is justified. The intended research should contribute to our understanding of the variability of the net weight process, by satisfactorily providing answers to our research questions . Not only should we attempt to answer, "Where does the net weight variation come from?" but also "What can we do to reduce this variation to optimise the net weight process to deliver the expected results" Methods to identify, quantify and minimise vari ation ofnel weights in canned foods 26 Chapter 3 Sources of Variability 27 SOURCES OF VARIABILITY 3-1 SOURCES AND STATE OF CONTROL The sources of process vari ables on the net we ight vari abil ity and the means and extent to which such sources of variations are controlled are important in des igning experiments. The quality and quantity control methods a lso serve as the means of establi shing the levels of factors as revea led by DoE to optimise processes. The following section ex plores the dynamics of the food process, the composition of foods , packaging, filling and the weighing mechani sms for related variabl es, status, and methods of control. 3-1 .1 VARIATIONS RELATED TO PROCESS DYNAMICS The food manufacturing processes consist of unit operations, fo r example ingredient measuring, size reduction, mi xing, preheating, fillin g, and thermal processing (A ppendices 3- 1 and 3-2). Although the bulk of ingredi ents are mi xed up during the initia l stages, further ingredients and fu ncti ona l additi ves may be added during subsequent operati ons. During unit operati ons, the mi xture is subj ected to chemica l and phys ica l treatments such as: ac idifi cation, size reducti on, and steam inj ecti on and so on, during specific po ints of the process. Such treatmen ts lead to specific phys ical or chemical transformati ons such as ge latini sation of starch, caramelisa tion of sugars, and hydration of beans. These transformations in tum result in detectab le or measurable changes in characteri stics such as v iscosity and ¾ Brix levels. (The Brix is defin ed as the percent Methods to identify, quantify and minimize variation of ne t weigh ts in canned foods Chapter 3 Sources of Variability 28 soluble solids measured using refractometers). These measurable changes manifest either as a shift in the process characteristics or as a change in the level of variation. The operational fluctuations of uni t-operations also create variations in these quality characteristics. While some of the unit operations are common to most foods such as batch weighing, mixing and preheating, other operations may be product specific such as steam injection or de-aeration . Appendix 2 shows a typical process layout on the factory floor with some of the common unit operation stations which are inter-linked via transfer lines . Appendix 12 illustrates, usmg the Beef Curry Process, how the author had conceptualised the origin of net weight variability by using his own experience as well as the existing knowledge among operational staff The net weight variability was seen as the net result of variations contributed by key process components. These components were identified as raw materials, processing methods, products, filler and containers. The variability arising from each of these components in turn is thought to be contributed by series of input variab les characteristic of the particular component. Similarly, filler variability was shown to be caused by "come up time" for optimum filler operation, head-to-head variation, variations due to filler setup parameters and variation due to filler mechanics. Some of the filler mechanistic variations arise mostly from product uptake and a few from delivery related variables . They include degree of plug and valve fit, mechanical changes in the stroke guide assembly and so on. The net weight filling process is a result of the combination of two separate processes namely, the product fornmlation and the filling where the fonner runs into the latter. Some of the variables in the fillin g process have their origins in the formulation process, while the rest of the variables are unique to the filling process itself as shown in Appendix 12. For example, the key input variables such as viscosity, concentration and temperature begin with product formulation while the speed of the filler and piston stroke length are unique to the filling process. Methods lo identify , quantify and minimize vari ati on of net weights in canned foods Chapter 3 Sources of Variability 3-1.2 VARIATION RELATED TO PRODUCT COMPOSITION 29 The major components used in the manufacturing opera tions are class ified into several groups. These groups mainly consist of Fresh or Frozen Vegetables, Meat and Fish, Dairy-based products, Condiments and seasonings, fl avours and colouring agents, starch-based products, sugar-based products , oleoresins, oil s and fa t and functional additives. The functional additives commonly used in Wattie's canning sector include emulsifiers such as lecithin, pH modifiers such as citric ac id, and bicarbonates. During the food process, blends and mixtures of various ingredients are added to the main recipe mix. Those frequently used incl ude slurries made from starch, Roux made out of butter and starch, liquid sugars and tomato paste. The variation of ingredient particle sizes, consistency, post-harvest age, climatic and topographical characters are the pri mary contributors to the product variabili ty in food processes. A common example is the seasonal vari ation of % Brix level in Tomato paste and variation of fruit tex ture levels from the start to the end of the season. The original variation may be further increased by the variation contributed from the weighing and measuring of these ingred ients prior to processing. 3-1 .3 VARIATION RELATED TO PACKAGING Besides being the deciding fac tor for quantity, can size contributes a significant amount of vari ability to gross weights due to variation of its weight and volume. The net weight of a product unit is ca lculated by subtracting the mean tare weight of the production lot fro m the gross weight. Therefore, this method produces only an estimated net weight of the package and, as a resul t, the tota l expressed variation of net weights in the production lot has a error component bui lt in to it. A constant headspace created by filler heads in a vacuum-filler contro ls fi lling vo lumes to a constant can height. Therefore the natural variation of can volumes during vacuum fi lling also could produce a net volume Methods to identify, quantify and minimize variation of net weights in canned foods Chapter 3 Sources of Variabi li ty 30 or weight variation, under a constant headspace. It is also possible to have several different can types under a single can size which create a weight or volume variation depending on the can neck type, lid type, beading characteristics, body thickness and the coating type. These variations could be characteri sed by their specific can codes. The variability of net content inside the can produces a headspace variation above the fill level. The headspace variation causes variation in vacuum inside cans, under constant steam pressure exhaustion. In extreme cases overfilling creates a variation of positive pressure instead of a vacuum. This illustrates how the net weight variation could affect can integrity and safety related factors. Methods to identify, quantify and minimize variation of net weights in canned foods Chapter 3 Sources of Vari ability 31 3-1.4 VARIA TIONS RELATED TO FILLING OPERATIONS The discussion below explains the mechani sms involved in vacuum and piston filling operations and how the specific mechanisms contribute to the variation of fill weights. Finally, a model for fill er operations is presented combining the knowledge of those operation mechanisms and the knowledge of the poss ible fo rces invo lved from fluid dynamics. 3-1.4.1 Vacuum Filling operations Thi s machine is suitable for the filli ng of low viscous products into glass, metallic, and plastic containers. The cans, prior to fillin g, may be either empty or may already contain other product components in solid or liquid fo rm. As vacuum filling eliminates most of the air present in the container, it ensures elimination of trapped air and maintains consistent filling. The effect of a vacuum combined with the headspace creating mechanism further ensures constant level fill. Vacuum fill ers are the most effi cient, accurate, and least vari able of the fo ur main types of fill ers under di scussion. As ev ident from the basic mechanism, the fill er perfo rms best when the food is liquid or when used as a final liquid fill of a multi -filling operation. 3-1.4.1.1 Mechanism of Vacuum Filling A tabletop chain conveys the contai ners to a synchronisation screw, with variable pitch to fac ilitate can entry. This provides spac ing between the cans and synchronizes them with a transfer star. The tar then transfers the containers fro m the synchronisation screw to the filling station or va lve tank where the cans are fill ed with the product. At the in-feed to the filling station, the cans directl y acti vate the "no can-no fill " mechanism, which pem1its the first of three phases of the filling cycle to begin. All these phases involved in fi lling take place during one fi lling cycle of the fill er operation. One filling cycle is a single rota tion of the filling table to which all the filling heads are attached. The phases of fi lling are controlled by a lever in contact with the Methods to identi fy , quanti fy and minimize variation of net weights in canned foods Chapter 3 Sources of Variability 32 cams located on a ring concentric to the filling station. Once the filling phase 1s completed, the containers leave the machine through a tabletop chain belt, which 1s tangential to the machine. The filling station is essentiall y a series of filling valves (heads) attached to the bottom of a circular filling bowl through a valve, in the form of a circle at the periphery. The filler bowl with its filling heads is variously referred to asfiJling station, valve tank, or .filling table. The filling table rotates above a can race that runs at the same speed, with all the filling heads directed at the open cans. ,_, ,\.~ /I" - .. /, ... r- .,.- ..;_ ' I 1 ,I ' • "tJ:. 1;'\ Ct/~ ~r '- .. J l t- • I " " =.:.._ I I ...__,__ '-1'---t-.l Figure 3-l Can sea l system comprising rubber gasket and open can flange In the beginning, the can top is temporarily sea led, and the filling head is lowered into the can. The temporary sealing of the can top takes place when a rubber gasket, which is located around the filer head, comes into contact with the open flange of the can as shown in Figure 3-1 . The headspace created inside the can is proportional to the depth to which the head is lowered inside the can. This is because during the filling process, liquid fills the can except for the volume occupied by the head. During the next stage, a vacuum is created inside the can and then the filler head opens the va lve leading to the filler bowl containing filling medium. Methods to identify, quantify and minimize variation of net weights in canned foods Chapter 3 Sources of Variability Operating Cycle l fil phase - Valve Cleaning The rubber gasket presses the can and a jet of steam cleans the valve passages. 2_nd phase - Vacuum application 33 After the drainage of the valve , the "no can-no fill" device turns the di stributor to a position which permits the evacuation of all the air present in the container. The container seat plates push the containers upwards until the flan ge is pressing against the gasket, assuring an airtight hold of the container. The filler creates a vacuum inside the can by connecting the inside of the can to a vacuum reservoir through a series of valves and pipes. The valve in the filler head opens up mechanica ll y during rotation to enable thi s connection. Once the connection is estab li shed, the vacuum will stay until the head moves out of the can. ~_r4 phase - Filling The di stributor rotates further, opening the valve leading to the filler bowl. At thi s point the inside of the container meets the liquid in the filler bowl. The dual action of suction due the vacuum inside the can and the gravity of the liquid head fills the can. The filler bowl maintains a steady supply of fi lling liquid at a controlled height. 4th I V . _ - p 1ase - entmg The di stributor rotates another fraction of the filling cycle, allowing the inside of the container to meet the outside atmospheric pressure. Once atmospheric pressure inside the can has been restored, the container is released from the filling va lve. The vacuum filler is connected to a vacuum pump through a vacuum tank. The vacuum tank is connected to the vacuum sliding block inside the filler through one or more plasti c pipes. The sliding block is designed to draw the vacuum from inside the cans while the cans rotate past the block during the vacuum phase. A contact is made between the sliding block and the individual filler heads through a short pipe. The vacuum slide system is water lubricated by a connection from the main water network (Zacmi , 1999, M 0255). Methods to identify, quantify and minimize variation of net weights in canned foods Chapter 3 Sources of Variability 34 3-1 .4.1.2 Head Space adjustment Gross Headspace Adjustment: The gross fill volume adjustment above I mm is achieved by changing the spacer rings (shims) between the di stributor and gasket to move the gasket up or down alongside the filler neck. Medium Headspace Adjustment: For headspace adjustment below I mm, the filling valves must be lifted or lowered by regulating the head-wheel that moves the filling station verti cally. This changes the thickness of the rubber gasket, making finer adjustments to the ex isting headspace. 3-1.4.1.3 Fine-tuning of fill-weights A vacuum-breaking valve located on the outside of the vac uum tank can adjust the level of vacuum created inside cans. The va lve is controll ed to bleed air into the tank at different leve ls, thus changing the degree of vacuum inside the tank. The e lectronic or manual control of this valve enables fine adjustment of fill weights for a given set of spacer rings and the gasket. The tank helps to maintain a uniform degree of vacuum inside the cans and facilitates separation of any product which may have been sucked in w ith the air. The tank is equipped w ith a level probe, which signal s to the volumetric pump to start to remove product in the tank. 3-1.4.1.4 Selection Criteria A vacuum fill er is recommended w here constant volume and headspace are preferred to constant weight. The vacuum filler produces consistent weights in cans where product density is uni form . Thi s is due to the capability of the vacuum filler to produce a consistent headspace , and a consistent vo lume as a result. The vacuum filler is particul arly good in delivering final liquid fill in a multi -fi lling situation. The use of Methods lo identify, quantify and minimize variation of net weights in canned foods Chapter 3 Sources of Variabil ity 35 other type of fi llers fo r the final fill , in a multi-filling situation, may produce increased vari ability in the total net weight due to the variation of previous fill weights and volumes. However, the final net weight variation tends to come down when vacuum filling is used as the means of final fill , due to the final fill level is being maintained to a constant fill vo lume. Another advantage would be the de-aeration of the product due to vacuum action of the fill er. Additional net weight variation due to unintentional incorporation of air could be significantly reduced by the use of vacuum filling. 3-1.4. 1.5 Limitations of vacuum fillers The limitation of filling capability ari ses mainly due to failure to ensure that the required vacuum or variation of vacuum is achieved during filling. There are several reasons for variation of vacuum inside cans: I) Vacuum loss due to malfunction of the vacuum pump 2) Inability to create the required vacuum due to leaking of air into an already- created vacuum from loose rubber gaskets used for sealing can tops 3) Improper valve seating 4) Product blocks in ducting which draw air from cans to the reservoir 5) Overfi lling of products in vacuum reservoir due to breakdown of product­ emptying mechanism 6) Temporary leak of air into reservoir during valve opening 7) Loss of horizontal alignment of the filler heads around the fi ller table, which causes loose fitting of rubber seal on the seam top 8) Worn or highly compacted spacer rings which pack the space between the rubber seal and the fill er head 9) Inadequate water for lubrication of vacuum sliding block The vacuum fill er capacity is limited by its speed, and the number of heads and can size the filler has been designed to handle. Vacuum fillers are designed to work best in terms Methods to identi fy, quanti fy and min imize variat ion of net weights in canned foods Chapter 3 Sources of Variability 36 of fill variations with a specified range of speed, which is usually expressed in terms of cans per minute. The throughput of vacuum fillers depends on their speed and the can size. 3-1.4.2 Piston Filling operations 3-1.4.2.1 Operation Mechanism A piston fill er with vertical piston va lves is used to fill liquid or semisolid products volumetrically. This machine is able to fill a wide variety of products into tinplate or plastic cans, and glass jars. The most commonly used products include tomato paste, tomato sauce, brine, jams, rnannalades, fruit and creams, mineral and vegetable oils. This type of filler is capable of handling thin liquids with minimum leak as opposed to the older generation of piston fill ers with rotary piston valves, which are in operation on manufacturing lines such as Recipe-RI and Seasonal-HI lines. A tabletop chain conveyer conveys the containers to a synchronisation screw. The screw separates and synchronises the containers through a transfer star. Then the star transfers the containers from the worm screw to the filling zone. When the holding tank rotates, major operation phases follow one another. The phases are controlled through the position of the roller on the piston-lifting earn . The circle through which the pistons are rotating is divided into four parts, which correspond to the four phases of the operating cyc le (Figure 3-2). The height of the product-holding tank is adjustable by means of motorised screw columns, which are controlled by the control panel. The vertical closing valve with the piston-controlled earn ensures the tota l absence of dripping, and max imum closing Methods to identify, quantify and minimize va riation of net weights in canned foods Chapter 3 Sources of Variability 37 3rd Area 2nd Area 1' 1 Arca P-------H Cans in D Cans out Figure 3-2 The four major areas of a piston filler operating cvcle Operating cycle (Figure. 3-3) l ~ area - level earn The valve rotates by a predetermined angle in order to close the connection between the filler bowl and piston cylinder. At thi s point, the passage between the piston cylinder and container opens up. 2_nd area - downward sloping earn The piston, whi lst carrying out the downward stroke, transfers the previously sucked product into the container. 3@ area - level earn This stops product transfer from the cylinder to the container and allows passage of the product from the filler bowl to the intake piston cylinder. Methods to identify , quantify and minimize va riation of net weights in canned foods Chapter 3 Sources of Variability 38 :-- 'a -_, -, - - I -, t- ,, __ / .J1_, - - - ) I.'' 5 • L ~ • ~) Figure 3-3 The cross-sections of piston strokes during the four major operations Methods to identify, quanti fy and minimize vari ation of net weights in canned foods Chapter 3 Sources of Variab ility 39 4th d I . _ - area - upwar s oping cam The piston , during its ascending stroke, sucks a quantity of product which corresponds to the volume of the cylinder detem1ined by the maximum height to which the piston plug ascends inside the cylinder. The height adjustment level of the piston plug carrier guides determines the height to which the plug ascends. Each of the piston plugs is connected to a plas tic roller through a long vertical slit in the upper part of the piston cylinder. Plas tic rollers are secured within two guide rail s and are able to ascend and descend through the lowest and highest points to which the rail s have been adjusted in the beginning. The ra il s are adjusted verti cally by turning a fill er adjustment wheel manuall y. The duration of the upward and downward strokes, (fourth and second areas) corresponds to approximately half a revolution of the fill er bowl. The table is moved up or down to accommodate the height of can plus the space for the cans to be lifted up to the filling head during the initial set up. This is done electrically. (Zacmi Food Process ing Plants, 200 I, Model 0290) 3-1.4.3 Pocket Fillers The pocket fill er is essentially a seri es of telescopic tubes where the inside volume of each tube can be adjusted to vary the fill vo lume or weight. Since the accuracy of the fill depends on the consistency of packing ins ide the pockets, fill weights and volumes are highly variable compared to vacuum or piston fill ers. Since the pockets are not properly sealed, they cannot be used fo r liquid fi lling. Pocket fill ers are generally used for meat varieti es, baked beans or any other ga rnish fillings of larger parti cle size, such as vegetable mixtures, spaghetti shapes and ravioli . However, they prov ide a practi ca l and economical way of weighing solid and chunk food as a part of a multiple filling process. Methods to identi fy , quanti fy and mi nimize variation of net weights in canned foods Chapter 3 Sources of Variabil ity 40 3-1.4.3 .1 Mechanism of Filling The pocket fill er consists of a seri es of telescopic tubes (pockets) that open up to the top of the filling table. Each one of the telescopic units consists of a larger diameter top tube and a smaller diameter bottom tube that slide in to each other during height adjustment. Top tubes are welded to the periphery of the circular filling table and the bottom tubes are connected to a bottom table which has the same diameter as the top table. The top table serves as a container for the filling material. Filling materi als are transfen-ed to the container by an appropriate conveyer system such as a slat be lt conveyer. When the table rotates, the filling material is swept aero s the pocket entrances. The fi lling into pockets is guided by a system of converging partitions and a set of brushes as illustrated in Figure. 3-4. Although the fill er adjustment is incremental, fill we ights are produced in di screte steps due to significant differences between weights of parti cles. For example, fill weights during sausage filling vary in steps of about 7 grams, due sausage-to-sausage weight di ffe rence. Variation within given adjustment depends on uni fo rmi ty of packing. The vari ation usually decreases with decreasing parti cle size. Other than the effect of particle size, the degree to which the frozen materi als have been thawed also contributes to packing vari ation. Properly thawed material s such as sausages and meat become fl ex ible enough to create a consistent pattern of filling compared to rigid frozen materials. Table 3- 1 shows the relationship between the parti cle size of di ffe rent meat fi ll s and the esti mated standard deviations for sample size = I 0. Methods lo identi fy , quantify and minim ize variat ion of net weights in canned foods Chapter 3 Sources of Variability 41 Separator guide ---t ____ ,- : : -----f, ; . ; o·p~~~le ... . - "'.'