You are on page 1of 17
1E 380 — Quality Control and Improvement Instructor Emre Nadar Unit 12 Notes 1. Unit 12 Goals Introduce acceptance sampling. * Discuss technical aspects of lot sampling, including modeling, OC-curves, and determination of parameters. # Calculate costs of a representative sampling plan. 2. Sampling There are three types of sampling: * Sampling for data collection Slashes! Posess Corl Deve canted e Investigative /Hypothesis testing. Design oP experimen Find cook causes 4 Iprove pravess # Lot sampling, 2.1 Lot Sampling Lot sampling is generally viewed as a lower life form: it is believed that as compa- nies mature, quality control will move from sampling to process control to design of experiments. To quote Montgomery: “Acceptance sampling can be an impor- tant ingredient in any quality assurance program; however, remember that it is an activity that you try to avoid doing.” 1 Lot or acceptance sampling (including 100% sampling) is used to: Ensure the quality of jincoming|material — When might you want/need to do this, and when not? Yes No Critical component Teiviol componestt Valuable material Commadtly New supplie Trusted supplier Inherent variation ia caw material Regular qeally Safely Cool, [Deshucdive to. scomple * Ensure quality of foutgoing|material — When might you want/need to do this, and when not? Yes No New prow Gal [Dashade Tsao Chaly probln High ily Variation Tne to market Safely / Liabslihy / Reg lenpossble Casts demand © It provides protection ~NOT PREVENTION - Use: We test all or a portion of a lot, of incoming or outgoing goods. We then sentence the lot; decide what action should be taken # Accept lot. # Reject lot « Do further testing/send for rework. 2.2. Aspects of Lot sampling 1. The purpose of lot sampling is to sentence lots, not to estimate lot quality (although this can be a side benefit). In sentencing we are in fact actually estimating the quality by using our gathered data, but the goal is an up/down or yes/no decision . THIS IS NOT A DIRECT FORM OF QUALITY CONTROL. Lots of variable quality still will get in. But we can reduce the numbers of bad lots, thus ensuring acceptable raw input to our process nN 3. We are not “inspecting quality into the product.” Rather we are auditing to see that our requirements are being met. 2.3 Approaches 1. Accept with no inspection. When? Ten pot Trusted supplier / process High [regglee qealhy 2. 100% inspection. When? JIT No margin for erro Catical pe! ~ Set engine Sofely / Lichikty Easy / Adel Masimize dota | 3. Acceptance sampling. Select a portion to test, and based on this decide about the lot. When? Work to speck bot 100% npection is. bee ay too cootly foo JPR boo tne-vleme bee dsr hoe bring Advantages and disadvantages of 100% sampling Pros Cons Mere cerkin Less cerkin - Fake Find 4 bras al deel Toe coy in tne, Morinize doe Destuclve Cuts © Weskd fl 2.4 Lot and Sample Selection Lot formation: This is often not considered, but it should be given some thought. We want lots which are: ‘* Homogeneous. Pacfedly nixed Same fype / bishry/ process / sepplir © Large Economies of scale Easily handled. Sample selection: This is crucial, we want a representative sample. We can choose © Numerically. Rardomly by serial # * Randomly by location. © Through stratification. 3 Select 3 from each ‘level” 3 3 3 There are different sampling techniques for different objectives. Often a com- pany will switch techniques as their needs and their processes change. This should be a dynamic process. 3. Sampling Schemes There are many different types of acceptance sampling techniques used: Single sampling, double sampling, multiple sampling and sequential sampling. There are also specific sampling frameworks which have been adopted, such as MIL STD 105E for US military work, and the Dodge-Romig and ANSI/ASQC Z1.4 for civilian use Over the lifetime of a firm, it is expected that they will vary the sampling methods they use, for example moving from single to double sampling, or from 100% inspection to a less rigid sampling plan. We will concern ourselves with studying the characteristics of basic single and double sampling schemes, with the belief that further knowledge can be obtained once the basics are understood. 4. Single Sampling for Attributes In single sampling for attributes, one sample of n units is selected (randomly!) from a lot (of size NV), and the number of defectives in the sample (d) are counted. If the number of defectives exceeds a prespecified acceptance number c, the lot is rejected. Otherwise it is accepted. nde: primary decision variables As with all sampling plans, there must be an underlying assumption as to the model which generates our input. In this case, we assume that each item in the shipment is defective, independent of the other items, with a fixed (unknown) probability p. Therefore if we have a shipment of V pieces, the number of defective pieces will follow the distribution of: J ~ Binomial (N, p) = defectives in lob J ~ Binomial (5, p) : dePctives in sample The problem is that we don’t know the parameter p of the distribution. It seems reasonable to assume that if the proportion defective is above some threshold, we will reject the shipment as unacceptable. 6 The question is how do we estimate the proportion defective, and how sure can we be of our estimate? How can we control the accuracy of our estimate? nf: occuracy f 4.1 Operating Characteristic (OC) Curves To answer these questions, we use the OC curve of the sampling plan, which plots the probability of accepting a lot versus the fraction defective it actually contains. Remember that sampling is a random process, so we can be absolutely sure of rejecting a lot with a higher proportion of defectives (or accepting a lot with a lower proportion of defectives) than we want only by counting the entire lot. We wish to avoid this; so by a judicious choice of n and c we can hope to be “reasonably” sure of our decision without 100 percent sampling. For a given probability of being defective (p), the probability of accepting a lot based on a sample of size n with acceptance number c, is just the probability that a binomial random variable with parameters n and p will generate c or less defectives (d). Therefore the probability we accept the lot, P,, is given by: YS aet a i=0 ° = BINOMDIST(<, n, p, TRUE) P, * F(c) = Pd < Example: If we are going to implement a sampling plan with n = 89 and c = 2, what is the probability that we accept a lot of size 10,000 if it has 1% defectives? 2 80! - <2) = Sv _onyic.ggy 8? P, = P(d<2) x Feo yi OD (99) _ So 0/.g9)89 .. 32 1 ss, 89! 27 gq)87 — 93 FrmgT OL) 99)"? + Fae aqO4)'(-99)** + srry (-01)(.99)*" = 9307 = BINOMDIST (2, 89, .01, TRUE) By calculating P, for various values of p, we can construct an OC-curve for this example, as seen on the provided spreadsheet. Question; Assume we wish to accept all lots with two percent or less defectives. What would our ideal OC curve be? Fecagh t | ACCEPT REJECT nr 2% If this is the case, how can we try to make the curve in the example look more like our ideal curve? Increase n. What other effects can we have on it? cP + Becoming less. strict Graphically 7 curve shiPls up 4 ight This is illustrated in the other two curves on the provided spreadsheet What about this curve with c = 0? Very. shit Veedore = ect happy 4.2 How Do We Determine n and c? To specify n and c, we need how many equations? Two What data do you think we can use to establish these? “Acceptable lead of quahy Error cos east oP ecephog, “bad” bokeh cost aP ejeling "gel" babeh Sampling eas We wish to construct a sampling plan such that the type | error, or the proba- bility that we reject when we should accept (if p = Paooa) is a, and type II error, or the probability that we accept when we should reject (if p = pgaa) is 8. In other words, we will have a probability of acceptance of (1 — a) for a lot which is sufficiently “good,” and rejection with probability (1 — 3) for a lot which is sufficiently “bad.” Having specified a, Peoods 8 and Prrad, we can solve for n and c from: nt it ni Daa = Yo pPliot(t — Paet) = a! 8 = Yael — Piae)™ n-7 d~ Binomial (n, p) ~ NC np, npll-p)) Why do we need two p values (Pood and Ppaz) instead of just one, and how do their values affect our sampling scheme? Pech. sngle warts ACCEPT ee Reset defective disk oP % defective > dich oP % defective T find if p=Pyed T find fp Pld | | | | | 1 Pred Pal oP OVERLAP IS TOO BIG! INCREASE “n" A LOT! LIN 2% 201% 2% 201% How do we find the sampling parameters? l-a@ PGood B PBad n a C2 Test1 Test2 0.975 0.010 0.005 0.040 50 1.879 -1.569 0.975 0.010 0.005 0.040 100 2.950 -1.048 0.975 0.010 0.005 0.040 200 4.758 0.862 0.975 0.010 0.005 0.040 300 6.378 3.257 0.975 0.010 0.005 0.040 400 7.900 5.905 0.975 0.010 0.005 0.040 500 9.361 8.713 0.975 0.010 0.005 0.040 600 10.777 11.636 0.980 0.005 0.975 0.010 0.005 0.040 550 10.073 10.162 0.975 0.003 0.975 0.010 0.005 0.040 550 10.000 10.000 0.975 0.003 0.975 0.010 0.005 0.040 545 10.000 10.000 0.977 0.003 0.975 0.010 0.005 0.040 540 10.000 10.000 0.978 0.003 0.975 0.010 0.005 0.040 530 10.000 10.000 0.981 0.005 Method: Fix the four parameters, and use the distribution function. A spreadsheet similar to this one has been provided/can be found on the MOODLE. Often peiood will be the acceptable quality level, or AQL, while pizaa will be the lot tolerance percent defective, or LTPD. The AQL is the poorest level of quality which the purchaser will accept on average, while the LTPD is the poorest level normal approximation to find the appropriate c values for a given n. Once the c values approximately agree (they have the same integral term), move up to the next integer, and vary 7 until you reach the accuracy you want. To test the accuracy, use the actual binomial of quality which the purchaser will accept for an individual lot. What do we do after inspection, especially if a lot fails? + Rework oc send it beck or Yeow i auey © Sang ogo 4 balled + Chede. san ry ia! Example: Sampling A company estimates that the quality of an incoming product is such that the percent defective (p) is normally distributed between zero and three percent per batch. As a first approximation, they model this percentage as discrete, with P(p = .005) = P(p = .025) = 0.25 and P(p = .015) = .5. These are considered “good,” “poor” and “fair” batches. The company estimates that they suffer a $50 penalty for rejecting a good batch, and a $250 penalty for accepting a bad batch. (Due to supplier contract and poor quality, respectively.) A sampling plan has been proposed having n= 100 and c = 2. It costs $.10 per item to sample. . What is the probability of accepting each category of batch? nv Therefore what is a and {3 for this plan? w . What are the expected error costs associated with each class of batch? Therefore what is the expected cost per batch sampled? = . What proportion of batches are accepted? o What is the sampling cost per batch accepted? (Why do we use batch accepted, rather than batch sampled?) ey What is the total expected cost per batch accepted of this plan? N What proportion of accepted batches are actually bad? ~ . Given the above data, if you were to change the plan, what changes might you consider? 2 . If you wanted a finer approximation of the cost, what would you do? 1. P(Accept | Good) = BINOMDIST (2,100, .005, TRUE) = 986 P(Accept | Fair) = BINOMDIST (2,100, .015, TRUE) = .8I0 PlAccept | Bad) BINOMDIST (2,100, .025, TRUE) = .542. 2. «=.0I4 pe 5h2 3. E [error cost| Good) = 50 (.014) = .70 E error cost |Fair] = & E [error cost |Paor} = 250 (.542) = 135.95 ELercor cost per batch sampled] =E [Cost | Good] P(Good) +E [Cost |Faic] P(Fai-) + E [Cost | Poor] P(Poor) =(50)(\/u4) + BU) + (135:55)('4,) = $ 34.06 L. PlAccept) = PlAccept| Good) P(Gued) + P(Accept | Fair) P(Fair) + PlAccept |Poor) PlPoor) = 187 5. Sampling cost per bolch tested = $10 Somplng cost per batch accepled = 410/787 = $129 This cePlects the “hidden’ inspection cost on what T produce. 6. 34.06 + 10 may «49606 PlAccept |[Poor) PlP oa P (Pa Acad coe = LMI 1795) 784 ” B. Be more shri! fra, be 9. -Cost/proftt for “far | ~ Better extinute for dict of p. Question: The above example costs out a sampling plan for a given n and c, while the spreadsheet helps us find n and c for a given @, 8, Pood and Pad. But how do we decide on a good, or better still, the “optimal” a, 3, péoou and PBad? Do we even need these parameters? Find the bet a te that mae profi 4 thn «4p os oof 5. Double, Multiple, and Sequential Sampling We will now briefly discuss other sampling methods. 5.1 Double Sampling A double sampling scheme is defined by four values, rather than just two, as in single sampling: my: The size of the first sample. The acceptance number of the first sample. We accept the lot after the first sample if the number of defective pieces in the first sample, di, is no more than cy ng: The size of the second sample. cg: The acceptance number for both samples. We accept the lot if the number of defective pieces in both samples, di + dz, is no more than cz haan ¢ Bese AO Example: Say we have a double sampling scheme where n, = 200, c = 6, ny = 100, ¢ = 15. What will a flow chart of actions for this scheme look like? Choose. n= 200 ‘es Cant d del 3 Accept Ne Yes Com" Yes No Reject (_ — Choose n,=100 Count dy dtd sis, Questions: # Let us compare the double sampling plan above with a single sampling plan with n = 300 and c= 15. Why would we want to do double sampling? “May hae fo do las song “Save time, save cost * Why would we not want to do double sampling? Double sampling - Less accurabe, mere complicaled Single sampling - More accurate, less complicated 5.2 Multiple and Sequential Sampling Multiple TMG Ma Sey 5 ES Sequeatiel | Sample one item ot a tine d Reject

You might also like