Ch 6 BB Improve

download Ch 6 BB Improve

of 105

Transcript of Ch 6 BB Improve

  • 7/28/2019 Ch 6 BB Improve

    1/105

    BB ProgramImprove Phase

  • 7/28/2019 Ch 6 BB Improve

    2/105

    Five Broad Improvement ApproachesExit from Analyze Phase

    Available?

    Check Availability of

    Useful Y = f (X)

    A1

    Yes

    No

    Effect Established?

    Examine Nature of the Xs

    Control Factors?

    Examine Status of the Xs

    Yes

    Yes

    No

    No

    Examine Feasibility

    of ExperimentationFeasible?

    Examine Nature of

    the SolutionKnown?

    A3

    A4

    A5

    A2

    Optimize using

    graphical or quantitativeoptimization techniques Adopt Data BasedInnovative Approach

    Conduct DOE

    Obvious Solution

    Adopt Innovation-Prioritization Approach

    Yes

    Yes

    No

    No

  • 7/28/2019 Ch 6 BB Improve

    3/105

    Innovation-Prioritization Approach (A2) Most popular. Frequently used for improving service

    processes

    But the approach is risky, since we are dealing with a

    poorly characterized process.

    So, even if the problem is solved successfully, usually

    It becomes difficult to hold on to the gains

    Very little is learned about the process

    Often requires sizeable capital investment or the solution

    involves control procedures which are difficult to follow

    on a routine basis

  • 7/28/2019 Ch 6 BB Improve

    4/105

    Innovative (A4) and Obvious (A5)

    Approaches In many cases it is found that the root cause of the problem is

    a control factor and its optimal level is either known(specified standard) or obvious. Implementation of the knownor obvious solution will solve the problem. Note however that

    even though the solution is obvious, identification of the rootcause may not be so.

    But what to do when the root causes identified are found tobe within their specified operating tolerances or if theproblem remains unsolved even after maintaining the Xswithin their operating tolerances?

    If the process is stable, then tightening operating tolerancesmay solve the problem. However, in case of an unstable

    process, tightening of tolerances is not likely to be effective.So what do we do? The answer to this question is discussedin the next section.

  • 7/28/2019 Ch 6 BB Improve

    5/105

    A Difficult Problem Consider a highly unstable continuous chemical process. The

    team takes a 30000 ft view of the process, somehow estimatesthe approximate sigma level and proceeds to the analyze

    phase. The control factors are identified easily. However the

    team now finds it extremely difficult to establish the effectsof the control factors. The team also observes that the levelsof some of the control factors are adjusted quiet frequently inresponse to the unstable behavior of the process. Althoughthe standard operating ranges are available for the controlfactors, at times these limits are violated. More surprisingly,such violations are found to have no bearing on process

    performance. In fact, in many cases the process performancedid not improve even after making the necessary processadjustments. Given this background, what should be theimprovement approach of the team?

  • 7/28/2019 Ch 6 BB Improve

    6/105

    A Difficult Problem (Contd.) The above situation indicates presence of strong interactions

    in the system.

    So we have a situation where the root causes have been

    identified but their effects are yet to be established. Also,experimentation is not feasible since we are dealing with acontinuous chemical process. The team needs to beinnovative enough to take care of the interactions.

    Note that, there is no point in rejecting such problems on the

    ground that these are primarily control problems (as identifiedduring the measure phase) and hence not ideal forbreakthrough. This is because, one finds it hard to controlsuch processes, particularly when the inputs vary a great dealdue to their natural or agricultural origin.

  • 7/28/2019 Ch 6 BB Improve

    7/105

    Improve Phase Road Map

    Find Optimal SolutionBrainstorm to Generate

    Potential Solutions

    Evaluate Potential Solutions

    and Select the Optimum

    Plan and Conduct DOE

    Find Optimal Solution

    Conduct Confirmatory Trials

    Solution Confirmed?

    Compare Expected Benefits

    and Project Goals

    A1 A2 A3

    Go back to

    appropriate stepof the Analyze orImprove phase

    B

    No

    Yes

    Find Optimal Solution

    Formulate Innovative data

    Collection and Analysis Plan

    A4

  • 7/28/2019 Ch 6 BB Improve

    8/105

    Improve Phase Road Map (Contd.)

    Set Operating Tolerances

    Assess Risks

    Risk Acceptable?

    Pilot Solutions

    Results Satisfactory?

    Develop Implementation Plan

    Results Acceptable?

    Go back to appropriatestep of the Analyze or

    Improve phase

    Go back to Definephase or close projectand document failure

    B

    Exit to Control Phase

    Yes

    Yes

    Yes

    No

    No

    No

    A5

  • 7/28/2019 Ch 6 BB Improve

    9/105

    Learning and Experimentation

    Chinese Proverb

    I hear and I forget

    I see and I remember

    I do and I understand

  • 7/28/2019 Ch 6 BB Improve

    10/105

    Theory and Experiment Experiments complement, supplement and

    stimulate theoretical research

    Conceptualization

    Experimentation

    No sense of measurement

    Thought experiment

    Creation of instruments Experimental set up

    (hardware and software)

    Measurement

    Theory

    Propositions

    Hypotheses

    Observations

    (Data)

    Experiments

    may betriggered by a

    specific

    societal need.

    There may be

    no concern foror relevance

    to theory

  • 7/28/2019 Ch 6 BB Improve

    11/105

    Role of Experimentation

    Three Broad Objectives

    Exploration/Discovery

    Results are compared against the whole body ofknowledge

    Investigation/Characterization/Optimization

    Comparison of a set of treatment results

    Verification/Confirmation/Demonstration

    Results are compared against

    standard/known/predicted values

  • 7/28/2019 Ch 6 BB Improve

    12/105

    Black Box Model of Experimentation

    PROCESSControlFactors

    Responses

    X1X2X3X4

    Xp

    .

    .

    Y1

    Y2

    Y3

    Yr

    .

    .

    . .Z1 Z2 Z3 Z4 Z5 Z6 Z7 Z8 Zq

    Noise Factors

    (Y1, Y2, ...., Yr) = f (X1, X2, ..., Xp; Z1, Z2, , Zq)

  • 7/28/2019 Ch 6 BB Improve

    13/105

    How to ExperimentAn Example Consider a simple experiment to find the soaking temperature

    that will maximize hardness of a steel component. Assumethat the experimenter selects three levels of soakingtemperature9500C, 10000C and 10500C forexperimentation.

    Comparing with our Black Box model Control factor: Soaking temperature

    Response: Hardness

    Noise factors: Soaking time, Chemical composition, Heating curve,Ambient condition, Measurement and a host of other factors

    Even in case of such a simple experiment, it is necessary toplan the experiment carefully

  • 7/28/2019 Ch 6 BB Improve

    14/105

    Heat Treatment Experiment

    Planning Questions

    How many trials at each soaking temperature?

    How many test pieces from each trial?

    How many measurements on each test piece?

    Should the number of trials / test pieces / measurements besame for each level of soaking temperature?

    How to deal with the potential noise factor like chemicalcomposition?

    How to deal with the potential noise factors like soaking time

    and heating curve? How to deal with variation in ambient temperature?

    Assuming three trials per level, what should be the order ofthe nine trials?

    Answers are by no means trivial!

  • 7/28/2019 Ch 6 BB Improve

    15/105

    Three Principles of Experimentation

    Answers the planning questions

    REPLICATION

    RANDOMIZATION BLOCKING OR LOCAL CONTROL

    Common to all experimentswhether single factoror multifactor

    Multifactor experiments raise additional designquestions

  • 7/28/2019 Ch 6 BB Improve

    16/105

    Replication

    Replication means repetition of a trial

    Why replicate?To obtain an estimate of

    experimental error Why estimate experimental error?To test

    statistical significance of the observed effects

    Why test?To gain confidence that thepredicted effects will be realized in practice

  • 7/28/2019 Ch 6 BB Improve

    17/105

    ReplicationPrimary, Secondary,

    Soaking temperature 9500C

    TP1 TP2M1 M2

    TP3M3 M4 M6M5

    Primary or Experimental Error

    9500C

    TP4 TP5M7 M8

    TP6M9 M10 M12M11

    Similar set up for 10000C and 10500C

    Source df

    Soaking temperature 2

    Trial - Primary error (e1) 1 x 3 = 3

    Test piece - Secondary error (e2) 4 x 3 = 12

    Measurement- Tertiary error (e3) 6 x 3 = 18

    TOTAL 35

    Test pieceMeasurement

    T

    E

    S

    T

    Genuine

    repetition

    of trials

    Heat Treatment Example

  • 7/28/2019 Ch 6 BB Improve

    18/105

    Randomization - 1Heat Treatment Example

    Trial # Temp.

    1 9502 950

    3 1000

    4 1000

    5 1050

    6 1050

    TP2

    TP1

    TP3

    M1

    M2

    M3

    M4

    M5

    M6

    Assume eighteen test pieces are cast

    from the same heat

    How should we allocate the test piecesto the trials?

    Randomly! To protect against anyunforeseen bias resulting from the

    condition of the test pieces

    Having made the allocation, in what order should we conduct the trials?

    Randomly! To protect against any unforeseen bias resulting from ahost of uncontrolled factors like furnace and environmental condition

  • 7/28/2019 Ch 6 BB Improve

    19/105

    Randomization - 2Heat Treatment Example

    Randomized trial order and

    allocation of test pieces

    Trial # Temp.

    4 950

    2 950

    5 10001 1000

    6 1050

    3 1050

    TP7

    TP10

    TP16

    M1

    M2

    M3

    M4

    M5

    M6

    How to randomize?

    Mechanical procedureDrawing

    of numbered chips from a lot

    Published random number table

    MathematicalRAND function

    in Excel

    Avoid mental randomization!

  • 7/28/2019 Ch 6 BB Improve

    20/105

    Randomization - 3

    Gain

    Protection from unforeseen bias Valid estimate of experimental errorif more than one replicate

    is available

    Higher the non-homogeneity of experimental material (e.g. onetest piece from each heat), the larger is the gain

    Loss Simplicity of logistics

    Protection from systematic bias only - Despite our besteffort, outliers may creep in!

    Gain and loss from randomization

  • 7/28/2019 Ch 6 BB Improve

    21/105

    Blocking Or Local Control - 1

    How should we deal with the noise factors likesoaking time and heating curve?

    Of course, as far as possible these are to be kept fixedat their standard operating level. Care should beexercised, so that they do not vary from trial to trial.

    This is called local controlobjective is to block the

    effect of the noise factors. Note that, these two can be control factors in other

    experiments

    Heat Treatment Example

  • 7/28/2019 Ch 6 BB Improve

    22/105

    How should we deal with the noise factorchemicalcomposition?

    We need 6 x 3 = 18 test pieces. How do we select them? All the eighteen test pieces from the same heat?

    Local control of a noise factor should be avoided.Reproducibility of the results may be compromised

    All the test pieces may not be obtainable from a single heat

    One piece each from eighteen different heats? Results may vary too much on account of variation in

    chemical composition itself. Effect of soaking temperature,the control factor of interest, may not be detected

    Solution for both the above problems follows:

    Blocking Or Local Control - 2Heat Treatment Example

  • 7/28/2019 Ch 6 BB Improve

    23/105

    Blocking Or Local Control - 3

    H1 H2 H3

    950C TP12 TP24 TP36

    1000C TP13 TP21 TP32

    1050C TP15 TP22 TP35

    Heat Treatment Example

    Select three different heats (H1, H2, H3)covering as muchmaterial variation as possible

    Select six test pieces randomly from each heat and allocate these(TP11, .., TP16, , TP31, , TP36) to the trials as follows

    H1 H2 H3

    950C TP14 TP23 TP31

    1000C TP11 TP26 TP34

    1050C TP16 TP25 TP33

    Replicate 1 Replicate 2

    Fair comparison of temperature levels is possible - WHY?

    H1, H2 and H3 are called BLOCKS. Hence the name blocking

  • 7/28/2019 Ch 6 BB Improve

    24/105

    Blocking Or Local Control - 3 How does blocking affect data analysis?

    For simplicity, consider only one replicate and onemeasurement / test piece in our heat treatment example

    H1 H2 H3

    950C (TP12) Y11 (TP23) Y21 (TP33) Y31

    1000C (TP13) Y12 (TP21) Y22 (TP32) Y32

    1050C (TP11) Y13 (TP22) Y23 (TP31) Y33

    Separate heat treatment ofeach test piece and conductingthe nine trials in random order

    Two-way ANOVA with oneobservation per cell

    Occasionally the block factor impose a restriction on randomizationof trial order

    For example, if day is chosen as a block factor then randomizationof trial order is possible only within a day

    ANOVA should reflect such restriction on randomization

  • 7/28/2019 Ch 6 BB Improve

    25/105

  • 7/28/2019 Ch 6 BB Improve

    26/105

    Types of Experiment Observational

    Past QC records or observation of the factor levels and thecorresponding outcome without making any intervention

    (1) Limited blocking/local control, if any (2) No randomization of

    trial orderhence no valid estimate of experimental error (3)Approximate replicates

    Correlation among the independent variables and autocorrelationof the dependent variables

    Even if a factor is found significant, one must be careful to infer

    cause and effect relationship Manipulative

    Factor levels are changed as per plan and responses noted

    Follows the three principles of experimentation

    Should be preferred, wherever possible

  • 7/28/2019 Ch 6 BB Improve

    27/105

    Types of Observational Studies Prospective

    Sampling points are either predetermined randomly (Y yet to beobserved) or determined randomly from all the available records

    Autocorrelation may be avoided by spacing out the sampling

    points. Blocking of the timeline and drawing samples randomlyfrom each block is a good practice

    Regression analysis

    Retrospective

    Y values are always available. Observations are classified

    depending on the value of Y, say defectives and non-defectives. Random samples are drawn from each group and then the X

    values of each group are compared to find the significant Xs, if any

    Rare event or the X values are generated after sampling

    Discriminant analysis

  • 7/28/2019 Ch 6 BB Improve

    28/105

    Manipulative Experiment Henceforth by an experiment we shall mean a manipulative

    experiment

    Observational studies are usually conducted during theANALYZE phase

    Usually, multifactor experiments play the most important roleduring the IMPROVE phase

    By a multifactor experiment we shall mean an experimentinvolving more than one control factors

    An experiment involving one control factor and one or moreblock factors is not a multifactor experiment

    Rest of the material is devoted solely to multifactor experiments

  • 7/28/2019 Ch 6 BB Improve

    29/105

    Multifactor ExperimentsAn ExampleAn experiment to increase hardness of a die cast engine component

    Factor Code Factor Level 1 Level 2

    A % Cu 0.10 0.20

    B % Mg 0.05 0.07

    C % Zn 0.03 0.06

    D Water Cooling On Off

    E Air Cooling On Off

    Which factors affect hardness?

    What is the rank of the factors withrespect to their impact on hardness?

    What is the best factor level combination?

    How much improvement can we expect atthe best factor level combination?

    Response: Rockwellhardness (B scale)

    Qualitative factor

    How to designthe experiment?

  • 7/28/2019 Ch 6 BB Improve

    30/105

    Traditional Approach

    (One-factor-at-a-time)

    Trial No. A B C D E Hardness Remark

    1 1 1 1 1 1 56

    2 2 1 1 1 1 63 A2 is better

    3 2 2 1 1 1 68 B2 is better

    4 2 2 2 1 1 69 C1 is better

    5 2 2 2 2 1 72 D2 is better

    6 2 2 2 2 2 75 E1 is better

    OptimumCombinationA2B2C1D2E1

    To what extent are the results of this experiment reproducible?

    A2 is found better while the other factors are at B1C1D1E1. Is itreasonable to assume that A2 will be better even at B2C1D2E1, the

    recommended optimum?

  • 7/28/2019 Ch 6 BB Improve

    31/105

  • 7/28/2019 Ch 6 BB Improve

    32/105

    Response GraphsA B AB

    A1 A2

    B1

    B2

    Avera

    geresponse

    A B AB

    A1 A2

    B1

    B2Avera

    geresponse

    Averagerespon

    se

    A B AB

    A1 A2

    B1

    B2

    Averageresponse

    A B AB

    A1 A2

    B1

    B2

    Averagerespon

    se

    A B AB

    A1 A2

    B1

    B2

    A B AB

    A B AB

    ?

    ?

    Three levelfactors ?

  • 7/28/2019 Ch 6 BB Improve

    33/105

  • 7/28/2019 Ch 6 BB Improve

    34/105

    Analysis of 2kExperiment

    An Example

    Single replicate of a 24

    experiment

    Response: Filtration rate of

    a chemical produced in apressure vessel (to be

    maximized)

    Factors: Temperature (A),Pressure (B), Reactant

    concentration (C), Stirringrate (D). Pilot plant

    experiment

    Level codes: Low (1), High(2)

    Standard

    order

    A B C D Filtration

    rate (Y)

    (1) 1 1 1 1 45

    a 2 1 1 1 71

    b 1 2 1 1 48

    ab 2 2 1 1 65c 1 1 2 1 68

    ac 2 1 2 1 60

    bc 1 2 2 1 80

    abc 2 2 2 1 65

    d 1 1 1 2 43

    ad 2 1 1 2 100

    bd 1 2 1 2 45

    abd 2 2 1 2 104

    cd 1 1 2 2 75

    acd 2 1 2 2 86

    bcd 1 2 2 2 70

    abcd 2 2 2 2 96

  • 7/28/2019 Ch 6 BB Improve

    35/105

    The 24 ExperimentANOVA Table

    ?15TOTAL

    00Error

    ?1ABCD

    1BCD

    1ACD

    1ABD?

    1ABC

    1CD

    1BD

    1AD

    1CA

    1BC

    ?

    1AB

    1D

    1C

    1B?

    1A

    SSdfSource

    Eight Y valuescorrespondingto level 2 of A

    Eight Y valuescorrespondingto level 1 of A

    A2A1Form four

    one way

    tables Cell Total Cell Total

    Form six

    two way

    tablesFour Y valuesFour Y values

    B2

    Four Y valuesFour Y valuesB1

    A2A1

    Cell Total

    Cell Total

    Cell Total

    Cell Total

    SSA =

    (Cell Total)2 / 8 - CF

    SSAB =

    (Cell Total)2 / 4

    CFSSA - SSB

    Form four

    three way

    tables

    SSABC =

    (Cell Total)2 /2CFSSASSBSSC

    SSABSSBC - SSAC

    By Subtraction

    TSS = Y2

    - CF

  • 7/28/2019 Ch 6 BB Improve

    36/105

  • 7/28/2019 Ch 6 BB Improve

    37/105

    2kExperimentComputing SS

    Computing interaction SS from the 2-way,3-way,..,

    k-way tables is obviously very tedious

    Yates method is very helpful for easy computationof the SS

    We wont discuss Yates method, since MINITAB

    is available

    However, it will be instructive to compute the SS

    from the effect contrasts using Excel

  • 7/28/2019 Ch 6 BB Improve

    38/105

    The 24 Experiment - Computing

    Factorial Effects and SS

    Trial No. A B C D Y

    1 - 1 - 1 - 1 - 1 45

    2 +1 - 1 - 1 - 1 71

    3 - 1 +1 - 1 - 1 484 +1 +1 - 1 - 1 65

    5 - 1 - 1 +1 - 1 68

    6 +1 - 1 +1 - 1 60

    7 - 1 +1 +1 - 1 80

    8 +1 +1 +1 - 1 65

    9 - 1 - 1 - 1 +1 43

    10 +1 - 1 - 1 +1 100

    11 - 1 +1 - 1 +1 4512 +1 +1 - 1 +1 104

    13 - 1 - 1 +1 +1 75

    14 +1 - 1 +1 +1 86

    15 - 1 +1 +1 +1 70

    16 +1 +1 +1 +1 96

    Y TotalL1 474 548 521 5021121

    Y TotalL2 647 573 600 619

    Recoded Design Matrix

    1 -1, 2 +1 Let the factor-level totalsgiven at the last two rows ofthe table be denoted by F

    i

    Effect of A = EA = A2A1

    = (547/8)(474/8) = 21.625

    SS due to A = SSA= (A

    2)2/8 + (A

    1)2/8CF

    = 4742/8 + 6472/811212/16

    = 1870.563Effect and SS due to otherfactors are obtained similarly

  • 7/28/2019 Ch 6 BB Improve

    39/105

    Effect Contrast Contrast for main effect of B

    = Column B x Column Y= - Y1 - Y2 + Y3 + Y4 -Y5 - Y6 + Y7

    +Y8 - Y9 - Y10 + Y11 + Y12 - Y13- Y14 + Y15 +Y16 = 573548 = 25

    Esource = 2*Contrastsource/n

    SSsource = (Contrastsource)2/n

    n = Total number of observations

    EB = (2*25)/16 = 3.125

    SSB = (252

    )/16 = 39.0625 We can verify thatSSB = (B1

    2+B22)/8 - CF = 39.0625

    (A12 +A2

    2)/n(A1+ A2)2/(2*n)

    = (A1A2)2/(2*n)

    ContrastAB= ? . ContrastABCD= ?

    Trial No. A B C D Y

    1 - 1 - 1 - 1 - 1 45

    2 +1 - 1 - 1 - 1 71

    3 - 1 +1 - 1 - 1 48

    4 +1 +1 - 1 - 1 65

    5 - 1 - 1 +1 - 1 686 +1 - 1 +1 - 1 60

    7 - 1 +1 +1 - 1 80

    8 +1 +1 +1 - 1 65

    9 - 1 - 1 - 1 +1 43

    10 +1 - 1 - 1 +1 100

    11 - 1 +1 - 1 +1 45

    12 +1 +1 - 1 +1 10413 - 1 - 1 +1 +1 75

    14 +1 - 1 +1 +1 86

    15 - 1 +1 +1 +1 70

    16 +1 +1 +1 +1 96

    Y TotalL1 474 548 521 5021121

    Y TotalL2 647 573 600 619

  • 7/28/2019 Ch 6 BB Improve

    40/105

    Contrast and SS of InteractionsA B C D AB ABC ABCD Y

    - 1 - 1 - 1 - 1 +1 - 1 +1 45

    +1 - 1 - 1 - 1 - 1 +1 - 1 71

    - 1 +1 - 1 - 1 - 1 +1 - 1 48

    +1 +1 - 1 - 1 +1 - 1 +1 65

    - 1 - 1 +1 - 1 +1 +1 - 1 68

    +1 - 1 +1 - 1 - 1 - 1 +1 60

    - 1 +1 +1 - 1 - 1 - 1 +1 80

    +1 +1 +1 - 1 +1 +1 - 1 65

    - 1 - 1 - 1 +1 +1 - 1 - 1 43

    +1 - 1 - 1 +1 - 1 +1 +1 100

    - 1 +1 - 1 +1 - 1 +1 +1 45

    +1 +1 - 1 +1 +1 - 1 - 1 104

    - 1 - 1 +1 +1 +1 +1 +1 75+1 - 1 +1 +1 - 1 - 1 - 1 86

    - 1 +1 +1 +1 - 1 - 1 - 1 70

    +1 +1 +1 +1 +1 +1 +1 96

    474 548 521 502 560 553 5551121

    647 573 600 619 561 568 566

    ColAB = ColA x ColBColABC = ColAB x ColCColABCD = ColABC x ColD

    Columns for other interactions are

    obtained similarly ContrastAB = ColAB x ColY

    = 561560 = 1

    SSAB = 12/16 = 0.0625 (same as

    obtained before from two way table)

    SSABC = (568 - 553)2/16 = 14.0625

    SSABCD = (566555)2/16 = 7.5625

    SS for the other components can beobtained similarly

  • 7/28/2019 Ch 6 BB Improve

    41/105

  • 7/28/2019 Ch 6 BB Improve

    42/105

  • 7/28/2019 Ch 6 BB Improve

    43/105

    The 24 ExperimentFinal ANOVASource df SS

    MS F%

    A 1 1870.5625 1870.562595.86** 32.3

    AC 1 1314.0625 1314.062567.34** 22.6

    AD 1 1105.5625 1105.562556.66** 19.0

    D 1 855.5625 855.562543.85** 14.6

    C 1 390.0625 390.0625

    19.99** 06.5ABD 1 68.0625

    68.0625B 1 39.0625

    39.0625BCD 1 27.5625

    27.5625BC 1 22.5625

    Pooled

    F.05,1,10 = 4.96F.01,1,10 = 10.04

    % =SSSourcedfError x MSError

    TSSx 100

    % indicates

    importance of the

    component

    Higher the F ratio,

    the higher will be

    %. But F ratio is

    difficult to interpret

  • 7/28/2019 Ch 6 BB Improve

    44/105

    The 24 ExperimentNPP

    Effect

    Percent

    20100-10-20

    99

    95

    90

    80

    7060504030

    20

    10

    5

    1

    Effect Type

    Not Significant

    SignificantAD

    AC

    DC

    A

    Normal Probability Plot of the Effects(response is filtration rate, A lpha = .05)

    Lenth's PSE = 2.625

    MINITAB output. We can computethe effects from the contrasts asdescribed before

    Insignificant effects are smaller inmagnitude and tend to be centered

    around zerothe fitted line

    Significant effects are larger inmagnitude and fall far away fromthe fitted line

    The same five effects as found bypooling are found to be significant

    NPP should not be viewed as an alternative to pooling, since we need an estimate oferror for constructing prediction band for the best combination. Lengths PSE(Pseudo Standard Error) in MINITAB output appears to be too conservative

    NPP can be used for having better judgment on pooling

    In many other cases, finding significant effects from NPP will not be as easy as above

  • 7/28/2019 Ch 6 BB Improve

    45/105

  • 7/28/2019 Ch 6 BB Improve

    46/105

    Practical Limitations of Full Factorials

    Let us consider a situation where we wish to

    investigate 13 factors, each at three levels

    A FF experiment will require 313 trials

    313 = 1594323

    Even if it takes only 10 minutes to conduct a

    trial, the complete experimental results will

    be available only after 60 years!!

    Industrial experiments involving ten or

    more three level factors is common

  • 7/28/2019 Ch 6 BB Improve

    47/105

    The Alternative Two practical alternatives

    Fractional Factorials (2k-p, 3k-p designs)

    Orthogonal Arrays (OA designsL4, L8, L16, L18, )

    Above classification is conventional, but somewhatarbitraryFFs are orthogonal and OAs are saturated FFs

    Here we shall discuss only OA designs developed by G.Taguchi because of their ease of construction and

    flexibility Taguchis OA designs are available in MINITAB

    3k-p and mixed level FFs are not available in MINITAB

  • 7/28/2019 Ch 6 BB Improve

    48/105

    Orthogonal Array L8

    1 2 3 4 5 6 7

    1 1 1 1 1 1 1 1

    2 1 1 1 2 2 2 2

    3 1 2 2 1 1 2 2

    4 1 2 2 2 2 1 1

    5 2 1 2 1 2 1 2

    6 2 1 2 2 1 2 1

    7 2 2 1 1 2 2 18 2 2 1 2 1 1 2

    One factor is assigned to one column -Maximum of seven two level factors

    1s and 2s in each column - Two levelsof the factor assigned to the column

    Each row constitute a trial

    Assume four two level factors (A-D)are assigned to the first four columns

    First trial: A1B1C1D1 Eighth trial: A2B2C1D2

    Vacant columns are used for estimatingeither interactions or experimental error

    Columns 1, 2 and 4 constitute the 23 design

    1/16th fraction of the 27design or 27-4 design

  • 7/28/2019 Ch 6 BB Improve

    49/105

    Orthogonal Array Series

    Two level series: L4(23), L8(2

    7), L16(215),

    L16(215): 16 rows, 15 columns, each column consists of 1s

    and 2s only. Similarly for other arrays

    Three level series: L9(34), L27(313),

    L9(34): 9 rows, 4 columns, each column consists of 1s, 2s

    and 3s.

    Mixed level series: L18(2

    1

    x 3

    7

    ), L36(2

    11

    x 3

    12

    ), . L18(2

    1 x 37): 18 rows, one 2-level column and seven 3-

    level column

    In short, OAs are represented as L8, L16, L18,

  • 7/28/2019 Ch 6 BB Improve

    50/105

    Meaning of Orthogonality

    Orthogonal means balanced or seperabe or unbiased

    The idea of balancing for clean separation of alternativesis used in many forms of experiments with which we are

    familiar In football games, the changing of field and each team

    getting a chance to kick-off is an act of balancing to avoidbias

    Technically, orthogonality implies that in any pair ofcolumns the all possible factor-level combinationsappear with the same frequency

    To explain, consider the array L9

  • 7/28/2019 Ch 6 BB Improve

    51/105

  • 7/28/2019 Ch 6 BB Improve

    52/105

  • 7/28/2019 Ch 6 BB Improve

    53/105

    Designing a Simple OA Experiment

    Step 1: Identify the control factors, noise factors and theresponse

    Control factors are those factors whose optimal level can be fixedand monitored during experimentation as well as in actual practice

    In case of field experiments, select about 5 -12 control factors.Reproducibility of results will be poor with less than five factors.Experimental error and cost of experimentation may be very largewith more than twelve factors.

    The Cause and Effect Diagram can be very useful in identifyingimportant control and noise factors.

    Identification of the noise factors is important for having properrandomization and blocking schemes.

    Selection of the response is not trivial, although it may seem so.However, considerations for selecting proper response is beyond thescope of this course.

  • 7/28/2019 Ch 6 BB Improve

    54/105

  • 7/28/2019 Ch 6 BB Improve

    55/105

    Designing a Simple OA Experiment

    Factor

    Code

    Factor Level 1 Level 2

    A % Cu 0.10 0.20

    B % Mg 0.05 0.07

    C % Zn 0.03 0.06

    D Water Cooling On Off

    E Air Cooling On Off

    Die Casting

    Experiment

    Response:

    RockwellHardness(B scale)

    Example: Factors, levels and response

    Levels can be created using any scalecontinuous, discrete, ordinal or nominal

    All the factors need not have the same number of levels. Such situations will be

    discussed later

  • 7/28/2019 Ch 6 BB Improve

    56/105

    Designing a Simple OA Experiment

    Step 3: Decide on the interactions to be estimated

    Experimentation keeping provisions for estimating too many

    interactions (say more than three) is not a good practice.

    Interactions are best dealt with through proper selection of factors and

    their levels (Steps 1 and 2). Use supplementary measurements and

    sliding level technique to deal with interactions. These techniques will

    be explained later through examples and case studies.

    Keep provisions for estimating only important interactions, which

    cannot be tackled otherwise.

  • 7/28/2019 Ch 6 BB Improve

    57/105

    Designing a Simple OA Experiment

    Step 4: Draw the required linear graph

    Required linear graph of the die casting experiment)

    A BAxB

    C D E

    Poor Construction of Linear Graphs

    Both the following representations are poor

    A

    B

    A x BC

    D

    EA B

    A x B

    C D E

  • 7/28/2019 Ch 6 BB Improve

    58/105

    Designing a Simple OA Experiment

    Step 5: Choose an appropriate OA

    Compute the total degrees of freedom (TDF) needed to estimate all

    the main effects and the desired interactions. Compute Minimum Run

    Size (MRS) and then the Desirable Run Size (DRS).

    MRS = TDF + 1 DRS = MRS + 4 (error degrees of freedom)

    Choose an OA from the appropriate series and of size >=DRS or of

    size >=MRS and replicate. If the number of factors involved is ten or

    more then use the MRS criteria.

    For our die casting experiment, TDF = 5x1(five main effects) +1x1(assuming A x B is also desired) = 5 + 1 = 6. MRS = 6 + 1 = 7.

    DRS = 7+4 = 11. Thus we have to choose either L16 or replicate L8.

  • 7/28/2019 Ch 6 BB Improve

    59/105

    Designing a Simple OA Experiment

    Choosing the right OAThe MRS criteria may fail

    Die casting example: MRS= 7. So one more interaction can be

    accommodated in L8. Can we accommodate one of AC, AD, BD etc?

    YES. Can we accommodate one of CD, DE etc? NO.

    Lesson: Two or more independent edges cannot be estimated usingL8. We have to use the array L16 in such cases.

    Similarly, independent interactions involving two 3-level factors

    cannot be estimated using L27 (see the standard LG of L27).

    Another extension (2-level factors): Interactions forming two or

    more independent triangles cannot be estimated using L16 . The

    array L32 is necessary for this purpose (see standard LGs).

  • 7/28/2019 Ch 6 BB Improve

    60/105

    Designing a Simple OA Experiment

    Choosing the right OAWhen the DRS

    criteria fail

    Modify the required LG by dropping a few interactions so

    that it becomes possible to use a smaller array.

    Using a larger array for estimating one or two specific

    interactions is likely to be unnecessary wastage of

    resources.

  • 7/28/2019 Ch 6 BB Improve

    61/105

    Designing a Simple OA Experiment

    Step 6: Allocate the factors to the columns of the chosen

    OA and construct the OA layout

    If no interactions are present and complete randomization of the trials

    is possible then any factor can be allocated to any of the columns.

    When interactions are present, it will be convenient to proceed as

    follows:

    6.1: Choose a standard linear graph and modify it to match the required linear

    graph

    6.2: Allocate the factors to the nodes of the modified LG by comparing the

    modified LG with the desired LG. 6.3: Construct the OA layout of the experiment

    Alternatively, the interaction table of the chosen OA can also be used

    for the purpose of allocation

  • 7/28/2019 Ch 6 BB Improve

    62/105

    Designing a Simple OA Experiment

    6.1: The required linear graph and its modification

    A BAxB C D E

    Required Linear Graph (Die casting experiment)

    2 46

    1

    7

    3 5

    4 5 6 72

    1

    3

    Standard Linear Graph (L8) Modified Linear Graph

  • 7/28/2019 Ch 6 BB Improve

    63/105

  • 7/28/2019 Ch 6 BB Improve

    64/105

    Designing a Simple OA Experiment

    6.3: OA layout of the experiment

    1 2 3 4 5 6 7

    1 1 1 1 1 1 1 1

    2 1 1 1 2 2 2 2

    3 1 2 2 1 1 2 2

    4 1 2 2 2 2 1 1

    5 2 1 2 1 2 1 2

    6 2 1 2 2 1 2 1

    7 2 2 1 1 2 2 1

    8 2 2 1 2 1 1 2

    eEDAxB CBA

    Column No.

    Source

    Preserve the

    layout. It will

    be useful

    during data

    analysis

  • 7/28/2019 Ch 6 BB Improve

    65/105

    Designing a Simple OA Experiment

    Step 7: Decide the number of replications and repetitions

    Strictly speaking, to be decided based on expected experimental error

    (process capability) and the least effect size that we want to detect

    But rarely done in practice. Mostly decided based on cost of

    experimentation and experimental budget

    A block factor may be associated with the replicates

    Recall that the experimental error is variation between trials and

    repetition error is the variation within a trial

    It is always advisable to have repetitions since repetitions are usuallyeasy to make more than one part/sample from each trial and more

    than one measurement on each part

    Measurement error is a common to both primary and secondary error.

    So, it is extremely important to check for measurement adequacy

  • 7/28/2019 Ch 6 BB Improve

    66/105

    Conducting the Experiment

    Step 1: Construct the physical layout of the experiment

    Delete the interaction and error columns from the OA layout

    Substitute the actual levels of the control factors in place of the coded

    levels of the OA layout

    Trial

    No

    (1)

    %Cu

    (2)

    %Mg

    (4)

    %Zn

    (5)

    WC

    (6)

    AC

    1 .1 .05 .03 ON ON

    2 .1 .05 .06 OFF OFF

    3 .1 .07 .03 ON OFF

    4 .1 .07 .06 OFF ON5 .2 .05 .03 OFF ON

    6 .2 .05 .06 ON OFF

    7 .2 .07 .03 OFF OFF

    8 .2 .07 .06 ON ON

    Physical layout

    of our diecasting example

  • 7/28/2019 Ch 6 BB Improve

    67/105

  • 7/28/2019 Ch 6 BB Improve

    68/105

    Conducting the Experiment

    Step 3: Design the experimental log sheet

    The physical layout + the data column + the provisions for recording

    special experimental conditions

    Step 4: Conduct the trials and record results Exercise local control

    Be careful in setting the levels of the control factors correctly. But no

    special care should be exercised

    Preserve the experimental units, input material etc., wherever feasible

    Step 5: Repeat erroneous and missing trials, if possible Statistical treatment of missing values is difficult

    Misplacement and breakage of parts before measurement may lead to

    missing values

  • 7/28/2019 Ch 6 BB Improve

    69/105

  • 7/28/2019 Ch 6 BB Improve

    70/105

    Analysis of Experimental Data and

    Confirmation of Results

    Step 1: Identify significant effectsANOVA

    Step 2: Find the best factor-level combination

    Effect curves, production cost, operating cost

    Step 3: Predict expected response at the best

    combinationSimple prediction formulae or

    regression analysis

    Step 4: Confirm predictionsA small confirmationrun and comparison with the predictions

  • 7/28/2019 Ch 6 BB Improve

    71/105

  • 7/28/2019 Ch 6 BB Improve

    72/105

    Analysis of the Die Casting Experiment

    Column # 1 2 3 4 5 6 7 Interaction AB

    Source A B AB C D E e B1 B2 Effect

    Level 1 549 573 579 555 537 591 567 A1 288 261 - 6.75

    Level 2 576 552 546 570 588 534 558 A2 285 291 + 1.50

    T = Total of the 16 observations = 1125

    RSS = Raw sum of squares of the 16 observations = 712 + 732+ + 722 = 79685

    CF = T2/16 = 11252 / 16 = 79101.56, TSS = RSSCF = 583.44 (df = 15)

    A1 = Total of 8 observations at A1 = 549, A2 = Total of 8 observations at A2 = 576

    SScol1 = SSA = (A1)2/8 + (A2)2/8CF = 45.56 (df = 1)

    Similarly for the other columns / sources SSe1 = SScol7 = 567

    2 /8+ 5582/8CF = 5.07 (df=1)

    SS (pure error) = SSe2 = TSS - j Sscolj = 583.44(45.56+ +5.07) = 57.50 (df=8)

    SSe2 can also be computed as ?

    SSreplicate need not be computed since the 16 trials have been randomized completely

    ANOVA table follows

  • 7/28/2019 Ch 6 BB Improve

    73/105

    Analysis of the Die Casting Experiment

    Source df SS MS F %A 1 45.6 45.6 6.55* 6.6

    B 1 27.6 27.6 3.97 3.5

    C 1 14.1 14.1 2.02 1.2D 1 162.6 162.6 23.39** 26.7

    E 1 203.1 203.1 29.21** 33.6

    AB 1 68.1 68.1 9.79* 10.5

    e1 1 5.1 5.1 - -

    e2 8 57.5 7.19 - -

    (Poolederror)

    (9) 62.6 6.95 - 17.9

    Total 15 583.4 - - 100

    ANOVA Table

    % =SSsourcedfsource x MSerror

    TSSx 100

    Factors D and E are most

    important. Factor A and

    the interaction AB also

    play significant role

    * Significant at 95% level of confidence

    ** Significant at 99% level of confidence

  • 7/28/2019 Ch 6 BB Improve

    74/105

  • 7/28/2019 Ch 6 BB Improve

    75/105

  • 7/28/2019 Ch 6 BB Improve

    76/105

    Prediction of expected response Main effect (significant): Expected response at A1 and A2 are

    .)()( 2211 lyrespectiveTATAandTATA

    If the factor A is insignificant then both are

    zero and hence the expected response is just T-bar.

    )()( 21 TAandTA

    Interaction effect (significant): jijiji BABATBAofEffect )(

    Assuming the effects of A, B and AB are additive, the expected

    response at AiBj is then given by

    E(Y) = (T-bar + Effect of Ai + Effect of Bj + Effect of AiBj)

    Case 1: A, B and AB are significant: jijijiji BATBABATBTATYE )()()()(

    Case 2: Only A and AB are significant: )()()()( jjijijii BBATTBABATATYE

    Case 3: Only B and AB are significant: )()( iji ABATYE

    Case 4: Only AB is significant:jiji BABATYE 2)(

  • 7/28/2019 Ch 6 BB Improve

    77/105

    Die Casting Experiment: Prediction of

    Mean Response at the Optimum

    Optimum combination: A2B1C1D2E1

    69.7831.7088.7350.7563.7125.71

    )()()()()(

    12112

    121212212112

    TEDBBA

    TETDTBABATATYE EDCBA

    Existing combination: A1B1C1D1E1

    07.7131.7088.7313.6763.7100.72

    )()( 1111111111

    TEDBBAExistingYE EDCBA

    Gain: (78.69 -71.07) RC = 7.6 RC

    + GREATER ROBUSTNESS

  • 7/28/2019 Ch 6 BB Improve

    78/105

    i i i fid

  • 7/28/2019 Ch 6 BB Improve

    79/105

    Die Casting Experiment : Confidence

    Interval

    From F table, F0.05, 1, 9 = 5.12

    From ANOVA table, Ve = 6.95

    ne = 16/(df of T + df of A + df of AB + df of D + df of E)

    = 16/5 = 3.2 Thus 95% of the individual observations of the

    confirmation run should lie within 78.69 6.82 RC.

    Confirmation failure indicates that either the factorial

    effects are not additive or important interactions havebeen ignored.

  • 7/28/2019 Ch 6 BB Improve

    80/105

  • 7/28/2019 Ch 6 BB Improve

    81/105

    Optimization: Complex Situations

    Example 1: Consider an experiment involving two

    responses Y1 and Y2. Assume A2 is better than A1

    with respect to Y1 but A1 is better than A2 with

    respect to Y2. How should we choose the best levelof A?

    Example 2: Consider three factors A, B and C. The

    combinations A1

    B1

    and B2

    C1

    are found to be the

    best. How can we find the overall best combination

    from the average response table?

  • 7/28/2019 Ch 6 BB Improve

    82/105

  • 7/28/2019 Ch 6 BB Improve

    83/105

  • 7/28/2019 Ch 6 BB Improve

    84/105

    Analysis of 3-Level OA Experiments

    Each main effect has two df and each interaction has 2 x 2 =

    4 df.

    SSA (for example) = (A12 + A2

    2 + A32)/rCF, r = No. of

    observations in the total Ai. If the two columns representing the interaction between two

    3-level factors are available in the OA layout then the

    interaction SS may be obtained by adding the SS of the two

    columns. Alternatively, the interaction SS may be obtained

    from the two-way table.

    If needed, each SS can be partitioned further into components

    of 1 df each.

  • 7/28/2019 Ch 6 BB Improve

    85/105

    Analysis of 3-Level OA Experiments

    For example, SSA (2df) = SSAL (1df) + SSAQ (1df)

    SSAL = (A3 totalA1 total)2 / (2r), r = No. of observations in

    Ai total.

    SSAQ = (A1 total + A3 total2*A2 total)2

    / (6r) SSAB (4df) = SSALBL (1df) + SSALBQ (1df) + SSAQBL (1df) +

    SSAQBQ (1df)

    Partitioning as above will be useful when the available error

    df is very limited Note that partitioning of SS into components of 1df is

    meaningful only for quantitative factors

  • 7/28/2019 Ch 6 BB Improve

    86/105

  • 7/28/2019 Ch 6 BB Improve

    87/105

  • 7/28/2019 Ch 6 BB Improve

    88/105

    Modification of OA: Mixed-Level Factors

    Collapsing of Columns

    To create a higher level column within a

    lower level array Creating a 4-level column in a 2-level array

    Creating a 8-level column in a 2-level array

    Creating a 9-level column in a 3-level array

    Creating a 6-level column in L18.

  • 7/28/2019 Ch 6 BB Improve

    89/105

  • 7/28/2019 Ch 6 BB Improve

    90/105

  • 7/28/2019 Ch 6 BB Improve

    91/105

  • 7/28/2019 Ch 6 BB Improve

    92/105

    Compound Factor Method

    A1B1 (AB)1, A2B1 (AB)2, A1B2 (AB)3

    Main effect of A = (AB)2(AB)1

    Main effect of B = (AB)3(AB)1

    SS(AB) = [{(AB)12 + (AB)22 + (AB)32}/(3r)]CF Assuming the assignment is to L9 and r is the no. of replicates

    SSA = [(AB)1(AB)2]2/(6r)

    SSB = {(AB)1(AB)3]2/(6r)

    Note that SS(AB) SSA + SSB, since A and B are not orthogonal

    If any of the two factors (say B) is found insignificant, then we may

    consider (AB)1 = A1, (AB)2 = A2 and (AB)3 =A1 and recalculate SSA as

    applicable for a dummy technique

  • 7/28/2019 Ch 6 BB Improve

    93/105

  • 7/28/2019 Ch 6 BB Improve

    94/105

    Effect Curves: 3 x 2 Experiment

    111.67

    129.33

    118.33

    110

    122.33

    132.33

    105

    110

    115

    120

    125

    130

    135

    A1 A2 A3

    BHN

    B1 B2

    How can we avoid the interaction?

  • 7/28/2019 Ch 6 BB Improve

    95/105

  • 7/28/2019 Ch 6 BB Improve

    96/105

    Background

    The bulk gas transmission and distribution network

    consists of thirty two flow meters, of which two are

    turbine meters and the rest are orifice meters.

    The whole network can be partitioned into severaloverlapping segments. The gas balancing equation

    for each such segment is given by

    Unaccounted Gas (UAG) = Gas OutGas In

    Gas In is the total flow measured by the input meters

    Gas out is the total flow measured by the output meters

    plus the Line Pack component

  • 7/28/2019 Ch 6 BB Improve

    97/105

    Background

    The company monitors daily UAG (as % of Gas

    In) in all the segments. This project was undertaken

    when the %UAG dropped to -0.6%.

    This amounted to a loss of about Rs. five millionsper month.

    It should however be mentioned that the real loss

    may not be as high as above, since all the gas outpoints are not billing points

  • 7/28/2019 Ch 6 BB Improve

    98/105

    Intermediate Studies

    The study consisted of systematic investigation of the process

    to identify the root causes for high variation of UAG. The

    details of this investigation, consisting of many small studies,

    will not be discussed here.

    The main result that was obtained at the end of these studies

    is that the UAG% could be kept near zero if the gas

    temperatures at two particular stations could be controlled in

    a particular fashion.

    Such a result was surprising since the meters are supposed to

    record flow in standard condition.

    This led us to the flow validation study, discussed next.

  • 7/28/2019 Ch 6 BB Improve

    99/105

    Flow Validation

    Two main sources of errorError in on-line measurement of

    gas composition, pressure, temperature and specific gravity

    AND error in flow computation by the flow computers.

    Detailed scrutiny of the calibration record eliminated the first

    possibility.

    Thus, although all the flow computers connected to the

    SCADA (Supervisory Control And Data Acquisition) are

    AGA -3 and AGA-7 compliant, it was decided to validate the

    flow computations by the flow computers.

  • 7/28/2019 Ch 6 BB Improve

    100/105

    Flow Validation Experiments

    The array L25 was used to study the effect of six factors on

    the error in flow computation.

    The levels of the factors were chosen to cover the entire range

    of operating conditions.

    Five levels were selected for each factor since the operating

    range was large and the theoretical flow computation

    equations involved fourth order terms.

    For each of the 25 experimental conditions, flow was

    computed by each of the five computers and the same

    compared with the corresponding standard value provided by

    a standard laboratory.

  • 7/28/2019 Ch 6 BB Improve

    101/105

    Factors and Levels

    Factor UnitLevel

    1 2 3 4 5

    Tube Diameter (TD) mm 40 110 180 250 320

    Diameter ratio (DR)* - 0.3 0.4 0.5 0.6 0.7

    Absolute pressure (PR) KPa 500 2000 3500 5000 6500

    Differential Pressure (DP) KPa 0.2 25 50 75 100

    Temperature (TE) C 10 20 30 40 50

    Specific Gravity (SG) - .50 .55 .60 .65 .70

    * Orifice Diameter (OD) = TD*DR

    Note that the factor OD has sliding levels. For example, the levels of OD

    are 12, 16, 20, 24 and 28 when TD = 40 mm but the levels are 96, 128,

    160, 192 and 224 when TD = 320 mm.

  • 7/28/2019 Ch 6 BB Improve

    102/105

    Experimental Results

    Trial

    #

    Flow (SCMH)

    Bristol

    Babcock

    Flow Boss

    600

    Flow Boss

    503ROC 809

    Turbo

    2500

    Apex

    (Standard)

    1 41.5 37.7 42.261 42.049 42.258 42.248

    . . . . . . .

    . . . . . . .

    18 24150.2 19323.07 21847.946 21867.01 21873.03 22054.52

    . . . . . . .

    25 243562.6 245153.01 247495.359 247428.3 247917.4 247924.7

  • 7/28/2019 Ch 6 BB Improve

    103/105

    Analysis of Experimental Data Let Fi be the flow measured by the i

    th computer and S be the

    corresponding standard value. The linear model [log(Fi) = i +

    *log(S)] was developed for the five computers. Ideally we should

    have i = 0 and = 1. It was found that in all the five cases was

    nearly unity. Accordingly, further analysis was carried out using the difference Z

    = log(S)log(F). Performance summary is given below

    Flow Computer Average Bias Mean Square Error Performance Rank

    Bristol Babcock -0.01405 0.00072 3 (Worst)

    Flow Boss 600 0.01735 0.00066 3

    Flow Boss 503 0.00294 0.000063 2

    ROC 809 0.00247 0.0000094 1

    Turbo 2500 0.00217 0.0000089 1

  • 7/28/2019 Ch 6 BB Improve

    104/105

  • 7/28/2019 Ch 6 BB Improve

    105/105

    Corrective Action and Benefits

    The matter was taken up with the supplier of the flow

    computers (Bristol Babcock and Flow Boss)

    Meanwhile these two computers were taken out of the

    system and the flow meters were connected to the other

    computers. Such a temporary corrective action resulted inmarked improvement in the variation of UAG.

    Apart from reducing variation in UAG, the company now

    had a methodology for identifying erratic meters (not

    discussed here)

    Further, it was realized that the present practice of flow

    validation based on measurements made at a particular

    condition is inadequate