CaseGriggetal.1998SPCinFishPackaging.pdf

ELSEVIER PII: SO956-7135(98)00018-S

Food Control, Vol. 9, No. 5, 289-297, 1998 PP. 0 1998 Elsevicr Science Ltd

PAPER

Case study: the use of statistical process control in fish product packaging

Nigel P. Grigg,“’ Jeannette Daly* and Marjorie Stewart?

The DTI publication ‘Code of Practical Guidance for Packers and Importers’ provides detailed advice on the establishment of effective statistical process control (SPC) systems to ensure ejficient and effective compliance with the requirements of the Average System for food and drink organizations. Since this document is non-mandatory, however; many organizations opt not to adopt such systems, and rely instead upon their checkweigher as a last line of defence. In many cases, this is because of the apparent complexity of the systems to the non-statistically trained. This paper presents a case study which demonstrates that simple, manual SPC systems involve very little statistical knowledge to establish and operate. Such systems, it is argued, can reduce unnecessary checkweigher rejections and product giveaway, assist with Trading Standards inspections and improve customer confi- dence. 0 1998 Elsevier Science Ltd. All rights reserved Keywords: The Average System; statistical process control; fish product manufacture

NOMENCLATURE

DTI Department of Trade and Industry. Non-Standard package Any food package having a weight below the absolute tolerance limit, T,. Inadequate package Any food package having a weight below the toler- ance limit, T,. Qn Nominal (declared) Quantity (or weight)

under the Average System. The quantity marked on a container of package.

Qt Target Quantity (or weight). The average quantity to which a packing or filling line operation is intended to produce.

SD Text abbreviation for standard deviation.

*Department of Consumer Studies, Glasgow Caledonian University, Park Campus, 1 Park Drive, Glasgow, U.K. G3 6LP and +Swankie Food Products Ltd, Baden- Powell Road, Kirkton Ind. Est., Arbroath, U.K. DDll 3LS. *Corresponding author. Tel: (0141) 337 4000; Fax: (0141) 337 4420; e-mail: [email protected]

(5 s so SP SPC TNE

TSO T

Population standard deviation. Sample standard deviation. Short-term standard deviation. Medium-term standard deviation. Statistical process control. Tolerable Negative Error. The negative error in relation to a particular nominal quantity, as defined by the Weights and Measures Act, 1979. Trading Standards Officer. Tolerance Limit. A defined quantity below which, according to the requirements of the Average System, no more than 2.5% of package weights may legally fall. It’s value is equal to the nominal quantity minus the tolerable negative error: ie Q, – 1TNE. Absolute Tolerance Limit. A defined quantity below which, according to the requirements of the Average System, no package weight may legally fall. It’s value is equal to the nominal quantity minus twice the tolerable negative error: ie Q. – 2TNE.

Food Control 1998 Volume 9 Number 5 289

Case study: N. P. Grigg et al.

INTRODUCTION

Quality control and statistical process control (SPC) have, for several decades, been synonymous in most high volume manufacturing environments, yet rela- tively little is written on the successful application of SPC within the food industry. This is surprising, given the benefits which SPC can provide in terms of economic, predictive and systematically documented process control (eg Gaafar and Keats, 1992; Xie and Goh, 1993; Wu, 1994; Cartwright, 1995; Holmes, 1996). In a recent Food Control paper, Hayes et al. (1997) highlighted the role which SPC can play in relation to food safety. This paper aims to highlight the role of SPC in packaging control.

Typically, weights and measures (W&M) control has less of a high media profile than food safety, but the underfilling of packages is also a significant issue in consumer protection, since the consumer takes it on trust that purchased food products are of the stated weight. In addition, for the food producer, there are potentially significant financial considera- tions. Quality costs can be considered under three main headings, namely, prevention, appraisal and failure (Crosby, 1979; BS6143, 1992; Oakland, 1997). Weight control falls into two of these categories: appraisal, in terms of monitoring product and compo- nent weights during process control, and failure, in respect of product giveaway and underweight products. Examples of failure costs would be the cost per unit of giveaway, the cost of any packaging material discarded in a rejection/repack situation (eg the bursting out of sachets of dehydrated quick soups, in an attempt to salvage the contents where inade- quate packages have been manufactured), or the cost of any litigation under the Weights and Measures Act, 1985.

Whilst failure costs can, in some cases, be traded off against the costs of establishing and maintaining effective control systems (prevention costs), there are other, less tangible, benefits associated with SPC which make the investment worthwhile. Such benefits include the systematic recording of quality data, the possibility of predictive control, allowing corrective action to be taken proactively, and the provision of confidence to customers (eg large retailers) that an effective system is in place.

In recognition of the benefits which SPC systems can accrue to food packing environments, the DTI outlines the application of such systems in it’s Codes of Practice for packers (DTI, 1979a) and inspectors (DTI, 1979b). These recommendations are designed to enable the packer of food products to routinely, systematically and demonstrably meet the require- ments of the Average System. In spite of this, however, many packers do not make use of the tech- niques described, many relying instead upon check- weighers to detect underweight items. In a recent pilot survey of 200 U.K. food manufacturers, under- taken at Glasgow Caledonian University, out of 71

responding organizations representing a variety of product types, 46 (65%) had a formal documented quality system covering W&M control, but only 24 (34%) made use of SPC in this area.

This paper presents a case study of a medium- sized fish product producer based in Scotland, Swankie Food Products Ltd. The organization uses a pre-checkweigher SPC system, designed around the guidelines presented in the DTI (1979a) Codes of Practice. The system is manual, and is simple and effective, requiring neither specialist statistical know- ledge nor any computer software. Only the equivalent input of one member of the staff is required to operate the system, but there are quantifiable cost benefits in terms of the control of product overfill and the less easily quantified costs of the conse- quences of product underfill. This paper aims to demonstrate the practical simplicity of the system used, and present the benefits that such systems may thus accrue, regardless of the complexity or sophisti- cation of the system. It is hoped that this might encourage more organizations to consider adopting such a system.

LEGISLATION AND ENFORCEMENT

In the U.K., weights of packaged goods are controlled according to the Average System, the requirements of which are encapsulated under the ‘three rules for packers’, which state that for any given sample of packages taken for testing by a Trading Standards Officer (TSO):

Rule 1

Rule 2

Rule 3

The actual contents of the packages shall not be less, on average, than the declared nominal quantity. Not more than 2.5% of the packages may be non-standard, ie have negative errors larger than the Tolerable Negative Error (TNE)* specified for the nominal quantity <Q.>. No package may be inadequate, ie have a negative error larger than twice the speci- fied TNE.

Figure I is a schematic representation of the average system, showing the relationship between Qn and the tolerance limits T, and T2 for a Normally Distributed filling or packing process*. Table 1 shows the TNE values for associated Q. values.

*TNE values for given nominal or declared quantities are tabulated in DTI (1979a). *Note that in the parent distribution of packages, there is a limit of one package in 10,000 applied to the permitted number of inadequate packages, as opposed to none at all, as in the three rules. This is to compensate for the theo- retically infinite tails of the normal distribution and to allow for pragmatic enforcement.

290 Food Control 1998 Volume 9 Number 5

Case study: N. P. Grigg et a/.

Non-standard

Inadequate items

b 2TNE

Q" Nominal Quantity

(max 0.01%) Figure 1 Relationship between package weight distribution and Average System tolerance limits

THE CODE OF PRACTICE

This document (DTI, 1979a) sets out the require- ments of the Average System, and the attendant responsibilities of packers and importers. Appendix C of the Code is a detailed exposition on the applica- tion of effective SPC on packaging operations. The system and theoretical background are rigorous, and can be daunting to those without prior training in statistical theory. In recognition of this, the document also presents an alternative ‘Off-the-Peg Control System’, designed to be straightforward with a minimum of theory, and applicable to ‘widely varying products and processes’ (Appendix F, DTI, 1979a). The following case describes the practical operation of this system, and the benefits which have resulted for the organization.

CASE STUDY: W&M CONTROL IN FISH PRODUCT MANUFACTURING

The sequence of activities required to establish the off-the-peg system come under the following broad headings:

(1) Collection of data. (2) Assessment of process characteristics. (3) Calculation of target fill quantity.

Table 1 Tolerable negative errors (DTI, 1979a,b)

Nominal quantity Tolerable negative error (TNE) (Q.1 g or ml

as % of Q. gorml

5-50 9 – 50-100 4.5 100-200 -4.5 – 200-300 9 300-500 -3 – 500-1000 – 15 1000-10000 1.5 10000-15000 150 above 15000 1

Reproduced with the permission of the controller of Her Majesty’s Stationery Office.

(4) Construction of control charts. (5) On-line control using charts.

Collection of data

The first stage in establishing this or any other such system is to accurately determine the average fill level and the inherent variability of the packing process under current manufacturing conditions.* This stage is normally referred to in statistical quality control as an initial study. The data for this Case Study are shown in Table 2. A minimum of 200 packages are required, to be collected in subgroups of size 3 or 5. The subgroup size and sampling rate are left to the packer to select on the basis of production levels. Swankie opted to use subgroups of size 5, taken at half-hourly intervals, making a total of 40 subgroups. As shown, for each subgroup (i), the mean (denoted X), the standard deviation (hereafter abbreviated to SD and denoted by the symbol ‘s’), and the range (the difference between the largest and smallest subgroup weights, and denoted ‘R’) have been obtained.

Calculation of process characteristics

The overall mean fill is the average =of the 40 subgroup means. This value is denoted x (mean of means), and its value for this data is 363.54 g. The individual subgroup means will be required later for plotting on the X chart.

The variability of a process is normally measured using SD. In the detailed system (Appendix C of DTI, 1979a), two measures are required; short-term SD (denoted s,,) and medium-term SD (denoted sp). The value for s,, represents normal piece-to-piece

*Although an accurate assessment is sought, the data should be taken under conditions as ‘ideal’ as possible. It is not recommended to set up control charts on the basis of out-of-control data, as this will render the process charac- teristic estimates inaccurate for ‘normal’ production.

Food Control 1998 Volume 9 Number 5 291

Case study: N. P. Grigg et al.

variation, whereas s, incorporates the effects of any additional shift-to-shift fluctuations. In the off-the- peg system, for practical simplicity, one measure of variability suffices. All necessary calculations can be carried out using either average subgroup SD (S), or average subgroup range (R). For the case study dataset (Table 2) it can be seen that the two values are respectively 5.06 g and 12.80 g. Since in this system process variability can be estimated using R, then it is not always necessary to calculate SD at all. For some manual systems, this will come as good news to the process controller, since this is not a trivial calculation, and can lead to errors (such as simple miscalculations or failure to apply the correction for small samples).

The process parameter estimates thus obtained are required for the construction of control charts, and the determination of the target quantity described in the next section. All remaining calculations are there- fore based upon these measures. The appropriate calculations will depend upon which process varia- bility measure (s or R) has been calculated. For the purposes of this article, calculations using both

measures shall be used, to give the reader an indica- tion of the degree of error involved between the two methods.

Determination of minimum target quantity (Q,)

The minimum ‘target quantity’, denoted Qt, is the minimum level at which the filling or packing process should be set in order to consistently meet the requirements of the Average System regulations (ie the three rules for packers). As shown in Figures 2 and 3, the value of Qt will depend upon the process SD. For any given Qn, there is an associated ideal value of SD which will exactly meet the requirements for non-standards and inadequates. The ideal distri- bution with this SD is referred to in the diagrams as the ‘W&M requirements distribution’. As shown in Figure 2, for any actual process with a SD at or below this value, Qt can be set equal to Qn (its lowest possible value).

As Figure 3 shows, however, the larger the process SD, the more packages will cross the lower limits of T, and T,. In this case, Q, must be set larger than Q.,

Table 2 Net weight data set (Q. = 350 g, all weights in g)

Batch Sample i

Mean Std Dev ii si

Range Ri

364 361 360 361 360 362 368 361 364 364 358 361 361 362 361 367 358 365 361 369 361 364 362 360 359 359 360 361 365 351 362 368 364 361 365 361 363 363 358 362 360 369 360 363 362 358 363 362 362 364 361 362 365 371 363 362 362 358 359 364 357 380 361 348 364 354 377 362 353 365 351 367 363 361 365 363 363 371 364 367 365 361 360 372 366 374 370 367 372 377 374 364 371 375 374 310 362 362 370 367 367 316 377 362 382 367 358 366 350 356 352 360 357 360 351 353 359 361 355 365 357 364 362 356 360 365 377 372 351 371 365 364 313 366 370 365 367 370 360 364 365 361 365 361 357 357 360 370 363 382 354 361 373 361 358 372 377 358 356 358 360

361 1.6 3.2 2.8 3.4 6.7 3.0 4.0 4.1 3.0 2.0 3.5 3.2 3.1

4 8 2 360

3 357 363 360 1

364 362 9 18

8 9

11 8 5 9 9 8 9

351 356

361 360 1 6

358 362 363 361 361 364 362

8 365 9 357

10 362 11 12

366 367 357 369 360

1 2 13 362

14 15

365 4.3 360 2.7

16 367 17 352 18 376 19 366 20 369 21 354 22 374 23 376 24 365 25 364 26 377 27 354 28 363 29 369 30 369 31 361

364 11.5 32 360 10.5 25 364 9.0 25 364 367 363

1.9 2.9 6.8 3.0

1s 371 372 5.2 13 369 5.6 13 3 366 3.1 8 375 359

7.5 1.4

20 17

358 358 361 361 367 367 367 363 362 362 365 368 361 : = 363.54

4.2 I1 7.1 18 5.7 14

30.0 8 26 13

5 13 IO

5 32 371 5 33 360

10.0 5.3

34 365 357 358 358 357 370 371

2.1 5.0 4.3

5 5

35 36 37 38 39 40

5.2 13 11.8 28 5

5 5

7.9 19 6.0 15

s = 5.06 R=l 2.80

292 Food Control 1998 Volume 9 Number 5

Case study: N. P. Grigg eta/.

Figure 2 Q,, = Q,: 3 Rules Met

Qn= Q, /'

b = W&M requirements distribution

and its lowest possible value must be mathematically obtained.

Theoretically, based upon the requirements of the packers rules, for a known medium term SD,* Q, is equal to the largest of the values resulting from the following expressions:

Q,# (1) T,+I .96s,, (2) T2+3.72s,, (3)

Expression (1) ensures that the target value is at least equal to the nominal quantity. (2) ensures that

*The medium term SD figure s, is used in this instance because it will allow for shift-to-shift variation.

no more than 2.5% of packages will fall below T,, and (3) ensures that no more than 1 in 10,000 packages will fall below T,, for any given s,. The multiplying factors used (1.96 and 3.72) are fixed constants under the normal distribution, representing the number of SDS limiting 2.5% and 0.01% of the curve. To aid clarity, the situation is represented schematically in Figure 4.

In the off-the-peg system, because practical simpli- city is sought, the multipliers involved in formulae l-3 are changed so that S or l? can be used to obtain Qt, instead of s,, which is harder to obtain. The new multipliers are given below in Table 3. For the sample data, Q. is 35Og, and therefore (from Table 2) the TNE is 3% of 350 g = 10.5 g. Hence T, = 339.5 g, and

Qn – Q,?

a = Filling Distribution

b = W&M requirements distribution

Figure 3 Q, > Q,,: Q, to be obtained

Food Control 1998 Volume 9 Number 5 293

Case study: N. P. Grigg et al.

Ts,-W Target

I 3.72 sp

b ~ Quantity

(Unknown) Figure 4 Finding Q, from T, and T,

Tz = 329.0 g. Table 3 shows the resulting values for Q, calculated using both S and R.

The optimum value for (2, is therefore 352.76 g (based upon S), or 353.07 g (based upon R). Once the optimum target quantity is thus established, this then becomes the optimum value for the filling process, and is used as the basis for the centreline of the mean control chart.

Use of on-line process control charts

mean and range or SD (eg Gaafar and Keats, 1992; Xie and Goh, 1993; Wu, 1994). ‘Mean’ charts are used to control the average value of a key process variable such as package weight, and ‘Range’ or ‘S’ charts to control the variability of the same measure- ment. The charts (of which Figures 5 and 6 are examples) consist of a centreline, relating to the overall average value of the sample statistic, and between one and four control limits, representing (broadly) minimum and maximum allowable values for that measurement.

The Code of practical guidance describes three types of on-line control chart which can be used by the packer or filler. These are:

(1) Original value plots; (2) Mean and Range Control Charts; (3) Cusum Charts.

This paper will discuss only mean and range control charts. The reader is referred to the Code of Practice for a discussion of the other methods.

These charts, standard in much of industry, are used to monitor and control sample statistics such as

For average weight charts, lower limits control underweight items and upper limits control overfill. As stated in the introduction, although underweights are of primary legal importance, overfill is to be avoided in food packing for economic reasons. It is therefore of benefit to the packer to use both upper and lower limits for this chart.

Table3 Formulae for obtaining Q, from s or R (adapted from DTI, 1979a)

Using & Using J7

Rule 1 Formulae Q,,+0.49 S Q,,+O.2 l? Value = 35og+o.49 (5.06g) =35Og+O.2 (12.8Og)

= 352.48 g = 352.56 g Rule 2 Formulae T,+2.62 S T,+1.06 f? Value = 339.5 g+2.62 (5.06 g) = 339.5 g+1.06 (12.80 g)

= 352.76 g = 353.07g Rule 3 Formulae T2+4.45 9 Tz+1.8 R Value = 329 g+4.45 (5.06g) =329g+1.8(12.8Og)

= 351.52 g = 352.04 g

For weight variability charts, upper limits are of more relevance, since the variability of values should be kept from becoming too great, but low values are to be welcomed. Weight variability is dependent on a number of factors and can be extremely product specific. The most common factors are those of equipment capability and specific product character- istics, eg sauce viscosity, the presence or absence of discrete particulate materials and the number of discrete components within the finished product.

Both types of chart are normally necessary in prac- tice, because, whilst average weight might be under control and stable, process variability could be simul- taneously increasing, resulting in increasing overfill and underfill, which would be undetectable from the mean chart alone. To exemplify this situation, consider two samples of size 3, the first consists of the values 249 g, 250 g and 251 g, and the second consists of the values 240 g, 250 g and 260 g. Both samples

294 Food Control 1998 Volume 9 Number 5

Case study: N. P. Grigg et al.

380

345

Figure 5 .? chart for net weight data set

Centreline

Upper/Lower Control Limits -ae-v–*-_

clearly have the same mean, which if treated in isola- tion fails to account for the underlying variation.

Apart from alerting the packer to actual violations of control limits, the other main function of these charts is predictive. They can be used to reveal trends in subsequent samples, which alert the controller to potential violations, and non-random variation. W. Edwards Deming, the pioneer of such methods, splits variation into common causes and special causes (Deming, 1986). Common causes are random and

unavoidable, whereas special causes are due to correctable faults such as machine drift or shift- to-shift variation. Using control charts, acceptable limits can be established on common causes, and special causes can be identified for corrective action.

Construction of charts

In the off-the-peg system, charts are constructed using only one control limit. On mean charts, this is a

Centreline

Upper Control Limit -c——-

Figure 6 Range chart for net weight data set

Food Control 1998 Volume 9 Number 5 295

Case study: N. P. Grigg et al.

Table 4 Control limit multipliers and values (adapted from DTI, 1979a)

Using 5

Upper/lower control limit for mean (x) chart Formulae Q,,,.,., 2 1.43 3 Value for dataset = 352.76 g + 1.43

(5.06 8) Upper limit = 360.00 g Lower limit = 345.52 g

Upper control limit for range (R) chart Formulae 2.29 S Value for dataset = 2.29 (5.06 g) Limit = 11.59 g

Using li

Q,,.,,., kO.58 ti = 353.07 g + 0.58

(12.80 g) = 360.99 g = 345.65 g

2.36 I? = 2.3h (12.80 g) = 30.2 g

lower limit to detect underfill, and on variability charts, this is an upper limit to detect high variation only. The authors of this article recommend using both upper and lower limits on the mean chart, and so both are used here.

Table 4 shows the formulae used to derive the limits, and the associated values, from the dataset. Again, the value for s or l? is merely substituted into the appropriate equation to obtain these limits, depending upon which statistic has been calculated.

The reader will observe that the limits for the mean chart do not differ significantly, regardless of which measure of variability is used. The difference in the range chart limits are merely due to the fact that the range is a larger value than the SD. This, therefore, requires a higher limit for its control.

The final stage of the analysis is the setting up of the control charts, using the limits obtained above, and the sample data in Table 2. The sample data are plotted on the appropriate chart, as shown in Figures 5 and 6. In the charts shown, R has been used to obtain the limits.

Comments on charts

The mean chart shows that a high level of giveaway is present in this product. Since the product is battered fish, it is the less expensive batter which is being oversupplied. Under normal circumstances, the packer might decide to adjust the mean level, resample and recalculate all the values described above before beginning real-time control with the charts. The chart also shows an apparent shift varia- tion from sample 21, after which means fluctuate more than previously. Such non-random or unusual variation should be investigated before continuing to use the chart, in case non-standard data has been included.

The reader will note the presence of an out-of- control value on the Range chart (Figure 6). This value relates to sample number 16 (Table 2), which shows a SD of 11.5 g, and range of 32 g. The lowest value in the sample is 348 g, which does not represent a non-standard nor inadequate item, but there is a high value of 380 g, representing 30 g of giveaway. This is followed by two further high values, suggesting

that it is unlikely to be a purely random occurrence. This should represent a significant concern to the producer of a relatively expensive product like fish. Normal procedure under such conditions would be to investigate the non-conforming sample in order to ascertain the reason for non-conformance, to then remove the sample from the data set, and finally, to repeat all stages and calculations of the initial study. The reason for this is that a sample with a mean or range value beyond normal production values will tend to influence the calculated values for overall mean and SD or average range. This will then influ- ence all subsequent calculations of target value and control or limits. A recalculation has not been carried out for this article but would, in normal, circum- stances be recommended strongly.

Use of charts

With the charts established, sample data are now taken on an on-line basis until limits require to be recalculated. The organization has to decide upon what action should be taken in the event of an out of control signal being detected. It is not sufficient to merely record non-conforming sample values. Such action should be written into a documented procedure as part of the organization’s quality system.

Benefits of the system

The benefits which Swankie Food Products Ltd experience as a result of these activities include reduced giveaway and unnecessary rejections at the checkweigher stage, since package weights are controlled prior to this stage. The systematic and explanatory records which are kept help to facilitate Trading Standards inspections, and can be shown to customers as evidence of Good Manufacturing Practice.

CONCLUSION

The case study in this article is designed to demon- strate the simplicity of application of the off-the-peg system. It is hoped that food organizations not presently applying SPC may by reading this article and become aware of the ease with which they might apply a working system, regardless of their particular specialism, through reference to the Code of Practice.

By establishing and using such an SPC system, the food packer can provide confidence to their customers that they are unlikely to violate any of the packers rules, fail the reference test, and thus be liable to prosecution under the Act. The organisation can also assist TSOs with routine inspections through the provision of high quality records (datasheets, calculations and control charts). Such a system can also reduce giveaway and help the packer to maintain the lowest possible legal target quantity.

296 Food Control 1998 Volume 9 Number 5

Case study: N. P. Grigg et al.

ACKNOWLEDGEMENTS DTI, (1979a) Code of Guidance for Puckers and Importers. Depart- ment of Trade and Industry, HMSO, London

The authors gratefully acknowledge the comments and suggestions received from the two reviewers of this paper.

DTI, (1979b) Manual of Practical Guidance for Inspectors. Depart- ment of Trade and Industry, HMSO, London

Gaafar, L. and Keats, B. (1992) Statistical process control: A guide for implementation. International Journal of Quality and Relia- bility Management 9(4), 9-20

REFERENCES Hayes, G. D., Scallan, A. J. and Wong, J. H. F. (1997) Applying

statistical process control to monitor the hazard analysis critical control point hygiene data. Food Control S(4), 173-176

BSb143, (1992) Guide to the Economics of Quality. British Stand- ards Institute, London

Cartwright, G. (1995) Measuring up for success. Qua/i& World 21(l). 16-19

Crosby, P. B. (1979) Quality is Free. McGraw-Hill, New York

Holmes, B. (1996) The true value of SPC. Quali@ World 22(h), 414-417

Oakland, J. (1997) Total Quality Management: Text with Cases. Butterworth-Heinmann, Oxford

Wu, Z. (1994) Single x control chart scheme. Internntionnl Journal of Quality and Reliability Management 11(Y), 34-42

Deming, W. E. (1986) Quality Productivity and Competitive Posi- Xie, M. and Goh, T. (1993) Improvement detection by control tion. Cambridge University Press, Massachusetts Institute of charts for high yield processes. International Journal of Quality Technology and Reliability Management 10(7), 24-31

Food Control 1998 Volume 9 Number 5 297