0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Statistical Design Methods For Engineers Lecture 3: Probability Density Functions Lecture 3: PDFs 1 0.8 0.7 0.6 Objectives 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 1. Introduce concept of probability, probability density functions (pdf) and cumulative probability distributions (CDF). 2. Discuss central limit theorem & origin of common pdfs: uniform, normal, Weibull. 3. Introduce JFIT tool to generate pdfs from simulation or experimental data. 4. Use a fitted pdf to calculate a probability of non compliance (PNC). Lecture 3: PDFs 2 0.8 0.7 0.6 0.5 Probability of Events: Multiplication Rule 0.4 0.3 0.2 0.1 o o o 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Let A = an event of interest Let B = another event of interest Let C = Event A occurring and event B occurring. – What is the probability of compound event C occurring? – P(C)=P(A and B)=P(A)P(B|A) P(A and B)=P(A∩B)=probability of A occurring multiplied by the probability of B occurring given that event A has occurred. – P(B|A) is called a conditional probability – P(C)=P(B and A)=P(B)P(A|B) – If events A and B are independent of one another then P(B|A)=P(B) and similarly P(A|B)=P(A) so P(C)=P(A)P(B). – If events A and B are mutually exclusive then P(A|B)=0=P(B|A) o Therefore P(A and B)=P(A)P(B|A)=P(B)P(A|B) which implies P(A|B)=P(A)P(B|A)/P(B) (Bayes’ Theorem) Lecture 3: PDFs 3 0.8 0.7 0.6 0.5 Probability of Events: Addition Rule 0.4 0.3 0.2 0.1 o o o 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Let A = an event of interest Let B = another event of interest Let D = Event A occurring or event B occurring. – What is the probability of compound event D occurring? – P(D)=P(A or B)=P(A)+P(B)-P(A and B) P(A or B ) is defined as an “inclusive OR” which means = probability of event A or event B or both events A and B – If events A and B are independent of one another then P(D)=P(A) +P(B)-P(A)P(B). – If events A and B are mutually exclusive then P(D)=P(A)+P(B) A B A∩B Lecture 3: PDFs 4 0.8 0.7 0.6 0.5 Example 1 0.4 0.3 0.2 0.1 o o o o 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Have two shirts in closet, say one red and the other blue. Have 3 pants in closet, say one brown, one blue and one green. What is probability that randomly picking a shirt and pants that one chooses a red shirt (event A) and blue pants (event B)? P(A and B)=P(A)P(B)=(1/2)(1/3)=1/6 (independent events) Lecture 3: PDFs 5 0.8 0.7 0.6 0.5 Example 2 0.4 0.3 0.2 0.1 o 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Have two parts in series in a circuit and both parts have to operate in order for circuit to operate. (series reliability) R1 o o R2 R1=probability of part 1 operating, R2 = probability of part 2 operating. If the reliability of part 1 is independent of the reliability of part 2 then P(part 1 AND part 2 operating)=R1*R2 Lecture 3: PDFs 6 0.8 0.7 0.6 0.5 Example 3 0.4 0.3 0.2 0.1 o 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 What if two parts are in parallel in the circuit and at least one of the two parts must operate for the circuit to operate. (parallel reliability) R1 R2 o P(part1 operating OR part2 operating)=R1 +R2 – R1*R2 if the reliability of part1 is independent of the reliability of part2. – How likely is this independence to be true? – Might the reliability of one of the parts depend on whether or not the other part is operating? Lecture 3: PDFs 7 0.8 0.7 0.6 Defining a PDF 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Probability Density Function, f(x) Definition - a mathematical function, f(x), such that f(x)dx = the probability of occurrence of a random variable X within the range x to x+dx. i.e. f(x)dx = Pr{x< X < x+dx} A typical Probability Density may appear as f(x) Freq. Of Occurrence X Lecture 3: PDFs 8 0.8 0.7 0.6 0.5 Properties of a PDF 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Probability of Occurrence - is the area under a pdf bounded by any specified lower and upper values of the random variable X. Note That The Value f(x) for all values of x of f(x) Can Be > 1 Total area under f(x) = 1.0 P( xL x xU ) xU xL f(x)=freq of occurrence per unit of x P( x ) 1.0 P( xL X xU ) xL f ( x)dx xU x Lecture 3: PDFs 9 0.8 0.7 0.6 0.5 Properties of a CDF 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 18.2 19.4 Cumulative Distribution Function, F(x) - a function, called a CDF, that returns the cumulative probability that a random variable, X, will have a value less than x. Maximum value = 1.0 x f (u )du for x F ( x) P( X x) Cumulative Distribution Function, CDF 1.0 F(x) P( X x) Probability 18.8 17 17.6 0 x X Lecture 3: PDFs 10 0.8 0.7 0.6 0.5 Expectation 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Expected value of a random variable X. – If X has only discrete value {xi, i=1,2,…,n} and pi is the probability of having the value X=xi. The expected value of X, written E[X], is determined by n E[ X ] pi xi i 1 If pi =1/n for all i = 1,2, ,n. i.e. all n possible values have same probability of being chosen. E[ X ] 1 1 x1 x2 n n x x2 1 xn 1 n n xn arithmetic average – If X has a continuous distribution of possible values then f(x)dx = the probability that X is between x and x+dx. The expected value of X is then given by weighting each value of X by its probability of occurrence f(x)dx. E[ X ] xf ( x)dx e.g. f ( x) e x , E[ X ] x e x dx 0 1 Lecture 3: PDFs 11 0.8 0.7 0.6 0.5 Measures of Central Tendency 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Mean x=+ 1 n Mean=X = xf(x)dx, Sample Mean = x = xk n k 1 x=- Median (middle value) m F(median)=0.5= f(y)dy=Pr(X m) - Mode (most probable) df 0, mode dx X=mode m Lecture 3: PDFs 12 0.8 0.7 0.6 0.5 Measures of Dispersion 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Variance x=+ 1 n 2 Variance (x-X ) f(x)dx, Sample Variance = s x x k n 1 k 1 x=- 2 X 2 2 X Standard Deviation 1 n 2 Standard Deviation= Variance X ,Sample Stdev = sX x x k n 1 k 1 3 Range 2 1 1 2 3 Range R Max x Minx Mean Absolute Deviation Normal pdf x=+ 1 n MAD |x-X |f(x)dx, Sample MAD xk x n k 1 x=- Lecture 3: PDFs 13 0.8 0.7 0.6 0.5 Variance 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Discrete distribution n n Var[ X ] pi xi E[ X ] pi xi2 [ E[ X ]]2 E[ X 2 ] [ E[ X ]]2 1 2 1 1 1-p e.g. geometric distribution f(x) = (1-p) p, E[X]= , Var[X]= 2 p p x-1 Continuous distribution Var[ X ] ( x E[ X ])2 f ( x)dx x 2 f ( x)dx E[ X ] 2 2 1 1 - x e.g. exponential distribution f(x)= e , Var[ X ] 2 2 2 Lecture 3: PDFs 14 0.8 0.7 0.6 0.5 Measure of Asymmetry 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Skewness (3rd Central Moment) x=+ Skewness (x-X )3 f(x)dx E[( x X )3 ], x=- Coefficient of Skewness Sk Skewness StandardDeviation 3 x=+ n 1 3 Sk 3 (x-X ) f(x)dx, Sample Sk 3 xk x x=- s (n 1)(n 2) k 1 1 3 Sk>0, positive Skewness Mean > Median > Mode Lecture 3: PDFs 15 0.8 0.7 0.6 Measure of “Peakedness” 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Kurtosis (4thx=+ Central Moment) Kurtosis (x-X ) 4 f(x)dx =E[(x-X ) 4 ], x=- Coefficient of Kurtosis Ku Ku 1 4 x=+ Kurtosis StandardDeviation 4 (x-X ) 4 f(x)dx, x=- (n 2n 3) (2n 3)(n 1) xk x Sample Ku 3 (n 1)(n 2)(n 3) k 1 s n(n 2)(n 3) 2 Ku>3, more peaked than Normal distribution n 4 Ku < 3, less peaked than Normal distribution Lecture 3: PDFs 16 0.8 0.7 0.6 0.5 Weibull distribution 0.4 0.3 0.2 0.1 23 3.25 3.00 2.75 2.50 2.25 2.00 1.75 1.50 1.25 1.00 0.75 0.50 0.25 0.00 -0.25 -0.50 -0.75 -1.00 -1.25 3.75 Z≡(X-) / Z 3.50 0.04 0.035 0.03 1 Mode = 7.6 Median = 20.1 Weibull 0.025 0.02 Mean = 25 x f ( x) 1 e x 0.75 0.5 CDF 22.4 21.8 21.2 20 20.6 19.4 Probability Distribution f(x), pdf 18.2 18.8 17 17.6 0 0.015 0.01 0.25 0.005 0 stdev = = 20 0 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 10 0 pdf CDF random variable, X Lecture 3: PDFs 17 0.8 0.7 0.6 0.5 Useful Properties 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Expected value of sum of variables is sum of expected values E[ x1 x2 xn ] E[ x1 ] E[ x2 ] E[ xn ] Variance of a constant, c, times a random variable x is c2 times the variance of x. Var[ x] E[( x E[ x]) 2 ] E[ x 2 ] ( E[ x]) 2 Var[cx] E[c 2 x 2 ] ( E[cx]) 2 c 2 E[ x 2 ] (cE[ x]) 2 c 2Var[ x] Variance of sum of two variables is sum of variances of each variable IF the variables are independent of one another. Var[ x y ] Var[ x] Var[ y ] cov[ x, y ] cov[ x, y] E[( x E[ x])( y E[ y ])] E[ xy ] E[ x]E[ y ] If x is independent of y, then cov[x,y]=0 (converse is not true, Why?) Var[ x y ] Var[ x] Var[ y ] Lecture 3: PDFs 18 0.8 0.7 0.6 0.5 Types of Distributions 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Discrete Distributions Continuous Distributions Bernoulli Distributions, f(x)=px(1-p)1-x Binomial Distributions, f(x)=(n!/[(n-x)!x!])px(1-p)n-x Poisson, f(x)=(x/x!) exp(-)*, Uniform f(t) = 1/(b-a) fpr a<t<b, zero otherwise Weibull Distributions, f(t)=(/q)(t/q)1exp(-(t/q)) – Exponential Distribution f(t)=exp(-t) Logistic Distribution, f(z)=z/[b(1+z)2], z=exp[ (x-a) /b]* Raleigh Distribution, f(r)=(r/2)exp(-½(r/)2)* Normal Distribution, f(x)=(1/2p1/2)exp(-1/2x/2 Lognormal Distribution • f(x)=(1/x2p1/2)exp(-1/2lnx)/2* Central Limit Theorem –Law of Large Numbers (LLN) *(not discussed here) Lecture 3: PDFs 19 0.8 0.7 0.6 Another Way to Pick Distributions: Distributions can be determined using maximum entropy arguments. 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Entropy is a measure of disorder or statistical uncertainty. The density function which maximizes the entropy* expresses the largest uncertainty for a given set of constraints, e.g. – If no parameters of the data are known f(x) is a Uniform Distribution (max uncertainty says all values equally probable) – If only the mean value of the data is known f(x) is an Exponential Distribution – If the mean and variance are known f(x) is a Normal Distribution (Maxwellian distribution if dealing with molecular velocities) – If the mean, variance and range are known f(x) is a Beta Distribution – If the mean occurrence rate between independent events is known: (mean time between failures) f(x) is a Poisson Distribution * Ref Reliability Based Design in Civil Engineering, Milton E Harr, Dover Publications 1987,1996. Lecture 3: PDFs 20 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Discrete Distributions Bernoulli, binomial Lecture 3: PDFs 21 0.8 0.7 0.6 0.5 Bernoulli Trial & binomial distribution 0.4 0.3 0.2 0.1 o 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 What if perform test and outcome can either be a pass or a fail (binary outcome). This is called a Bernoulli trial or test. o Let p = probability of a failure Bernoulli distribution f ( x) P{ X x} p x (1 p)1 x , x 0,1 P{ X 1} p probability of failure in test, P{ X 0} 1 p probability of passed test E[ X ] p, Var[ X ] p(1 p) Since many experiments are of the pass/fail variety, the basic building block for such statistical analyses is the Bernoulli distribution. The random variable is x and it takes on a value of 1 for failure and 0 for passed. Lecture 3: PDFs 22 0.8 0.7 0.6 0.5 Multiple Bernoulli Trials 0.4 0.3 0.2 0.1 o 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Perform say 3 Bernoulli trials and want to know the probability of 2 successes. R=1-p where p = probability of failure. – P(s=2|3,R) = P(first trial is success and second is success and third is failure OR first success and second failure and third success OR first fails and second is success and third is success) – R*R*(1-R) + R*(1-R)*R + (1-R)*R*R = 3R2(1-R) – General formula P(s|n,R)=nCs * Rs * (1-R)n-s (binomial distribution) General formulation of binomial distribution n x n x f ( x) P( X x) p 1 p x where n n! x n x ! x ! E[ x] np 2 E[ x 2 ] ( E[ x]) 2 npq np(1 p) np(1 p) npq q p 3 1 npq 1 6 pq 4 2 3 npq Lecture 3: PDFs 23 0.8 0.7 0.6 0.5 Probability Mass Function, b(s|n,p) 0.4 0.3 0.2 0.1 o 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Graph of f(x) for various values of n. (p=0.02) 0.6 f(x|n=20,p) 0.5 f(x|n=100,p) f(x|n=500,p) 0.4 begins to look like a normal distribution with mean = n*p 0.3 0.2 0.1 0 0 1 2 3 4 5 6 7 8 9 10 12 14 16 18 20 Lecture 3: PDFs 24 0.8 0.7 0.6 0.5 Example with binomial distribution 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Suppose we performed pass / fail test (Bernoulli trial) on a system x=1 if fails x=0 if it passes . Perform this test on n systems, the resulting estimate of the probability of non compliance = sum of x values / number of trials, i.e. <PNC> = (x1+x2+…+xn) / n = # noncompliant / # tests. binomial p x / n estimate of PNC E[ x] np,Var[ x] np(1 p) E[ x / n] np / n p Var[ x / n] Var[ x]/ n 2 np(1 p) / n 2 p(1 p) / n Stdev[ x / n] p(1 p) / n p(1 p) / n This information will be used later Lecture 3: PDFs 25 0.8 0.7 Continuous Distributions: Plots for Common PDFs 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17.6 17 0 0.3 0.2 0.4 Uniform 0.2 0 Normal 5 Log-Normal 1 0.1 0 2 4 0.2 0 0.5 Rayleigh 0 1 2 Weibull 1 10 1 0 Exponential 1 5 2 3 2 0 6 0.4 2 1 2 3 1 3 Gamma 6 2 0.5 4 Beta 1 2 0 0.5 1 0 0.5 0 2 4 1 See following charts for details → Lecture 3: PDFs 26 0.8 0.7 0.6 0.5 Uniform & Normal Distributions 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Uniform distribution 1 a b 2 1 2 2 variance= b a 12 b-a std dev= 2 3 mean= pdf Both the uniform and normal distributions are used in RAVE 1 b-a , a x b f ( x) 0 ,otherwise CDF 0 ,x<a 3 0 1 coef. of skewness x-a F ( x) ,a x b b-a 9 2 coef. of kurtosis 1 ,x>b 4 5 Normal distribution (Gaussian distribution) 1 f ( x) e 2p 1 x 2 2 mean 2 variance std dev 3 0 1 coef. of skewness Standard Normal distribution 4 3 2 coef, of kurtosis 1 1 z2 f ( z) 2p e 2 define z = (x - ) / (Unit Normal distribution) z mean=0 2 variance=1 std dev=1 3 0 1 coef. of skewness 4 3 2 coef, of kurtosis Lecture 3: PDFs 27 0.8 0.7 Central Limit Theorem and Normal Distribution 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 The Central Limit Theorem says: – If X= x1+x2+x3+…+xn and if each variable, xj, comes from a its own distribution whose mean is j, stdev j, then the variable Z, defined below obeys a unit normal distribution, signified by N(0,1). n Z n x j 1 j j 1 j 1 -2s/√n -1s/√n +1/√n +2s/√n +3/√n j ~ N (0,1) n -3s/√ 2 j Mean of x, <x> – This is true as long as no single xj value overwhelms the rest of the sum What does this mean? Using a simpler form for Z this means 1 n xj x The pdf of the mean value, x , of a variable is n j 1 Z n n ~ N (0,1) a normal distribution independent of what distribution characterized the variable itself. Lecture 3: PDFs 28 0.8 0.7 0.6 0.5 Generating PDFs from Data 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Commercial codes Crystal Ball http://www.decisioneering.com/ Codes developed by J.L. Alderman – Uses Johnson family of distributions. – 4-parameter distributions. – Parameters chosen to match first 4 central moments of data. EasyFit version 4.3 (Professional) http://www.mathwave.com/ Need some goodness of fit (gof) test. Kolmogorov-Smirnov (K-S) Anderson-Darling (A-D) Cramer-von Mises (C-vM) Chi squared (c2) others JFIT (quick) GAFIT (slower, but more accurate fit) – Uses Johnson family – Parameters chosen using Rapid Tool. Lecture 3: PDFs 29 0.8 0.7 0.6 0.5 Methods Of PDF Fitting 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Temperature 2.82124992 The ‘Crystal Ball’ Tool and the ‘EasyFit’ Tool both Fit the Actual Data To Each of The Various Density Functions In their Respective Libraries Lecture 3: PDFs 30 0.8 0.7 0.6 0.5 Methods Of PDF Fitting 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Temperature The PDF That Produces The Best Match, According To User Selectable Criteria (Chi Squared, KolmogorovSmirnoff(KS), Anderson-Darling (AD) ) Is Selected To Represent the Data. 2.82124992 The ‘Crystal Ball’ Tool and the ‘EasyFit’ Tool both Fit the Actual Data To Each of The Various Density Functions In their Respective Libraries Lecture 3: PDFs 31 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 N. L. Johnson Method Of PDF Determination Temperature 2.82124992 The Johnson family of distributions are fitted in the excel tool, JFIT and commercial tool EasyFit™ by matching the first four standardized central moments of the data to the expressions for the moments of the distributions. GAFIT uses GA for optimized fit. Lecture 3: PDFs 32 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 N. L. Johnson Method Of PDF Determination The Johnson family of distributions is composed of four density functions (normal, lognormal, unbounded and bounded) These members are all related to the unit normal distribution by a mathematical transformation of variables Computations are simplified because they can be done using the unit normal variable distribution f(Z)~N(0,1). Lecture 3: PDFs 33 0.8 0.7 0.6 0.5 Calculation of Moments from Data 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Standardized Central Moments NAME MOMENT 1st (noncentral) MEAN VARIANCE 2nd (central) SYMBOL x s 2 EXCEL FUNCTION AVERAGE(Range)* VAR(Range)* DIRECT CALCULATION (Unbiased) 1 n xi n i 1 1 n 2 x x i n 1 i 1 n 3rd (central)/s3 SKEWNESS Sk SKEW(Range)* n xi x 3 i 1 (n 1)(n 2) s 3 n (n 2n 3) xi x 2 4th (central)/s4 4th (central)/s 4 KURTOSIS* KURTOSIS* Ku Ku * KURT(Range)+3 i 1 (n 1)(n 2)(n 3) s 4 4 3 (2n 3)(n 1) n(n 2)(n 3) [(n2-2n+3)/(n(n+1))]*KURT + 3(n-1)/(n+1)* *Unbiased values Lecture 3: PDFs 34 0.8 0.7 0.6 0.5 Regions of Johnson Distributions 0.4 0.3 0.2 0.1 22.4 21.8 21.2 20 20.6 19.4 23 Skewness = Normalized Third Central Moment ForRegions Any PDF, Kurtosis Distributions > Skew2 +1 for Johnson This Defines The Impossible Area 0 1 Forbidden Region 2 =kurtosis 18.2 18.8 17 17.6 0 2 3 Unbounded, SU 4 Kurtosis = Normalized Fourth Central Moment 211 5 6 Bounded, SB Two ordinate lognormal 7 8 0 1 2 1=(skewness) 2 3 4 21.561.5123 Lecture 3: PDFs 35 0.8 0.7 0.6 JSB, Bounded Density Function 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 f ( x) 2p y (1 y ) e y 1 ln 2 1 y 2 x y , x y F ( x) ln 1 y The Johnson bounded distribution can fit many different shapes because it has 4 adjustable parameters Lecture 3: PDFs 36 0.8 0.7 0.6 JSU Unbounded Density Function 0.5 0.4 0.3 0.2 0.1 f ( x) 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 2p y 1 2 e F ( x) ln y y 1 2 2 1 2 ln y y 1 2 y x Again using 4 parameters allows for better fitting. Lecture 3: PDFs 37 0.8 0.7 0.6 0.5 JFIT Tool 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 Steps: Enter Data in Either C6-9 OR C13-17 not both! 2. Set Plot Range (D22-D24) 3. Enter Limits Or Target Yield Below PDF chart 4. Press compute Compute Type Probs? z -3.000000 -2.49 -2.28 -2.13 -2.03 -1.94 -1.86 -1.79 -1.73 -1.68 -1.63 -1.59 -1.54 First Four Central Moments of The Data 4.00000 Mean 1.00000 Variance 0.18226 Skew 3.05910 Kurtosis Do NOT insert data in both C6:C9 and C13:C17 Output Parameters of Johnson Curve if data inserted in C6:C9. -46.19231651 Types: Problems: 16.49535715 JSL=LogNormal 0 None 1 Variance <= 0 1 JSU=Unbounded -12.48020369 JSB=Bounded 2 Kurt < Skew^2 + 1 3 Sb did not converge JSN=Normal JSL 4 Inconsistent Input Data 0 JST=Two-Ordinate 5 SU did not converge Data Will Be Plotted For zmin = -3 zmax = 3 # of steps = 200 x P(z) 1.2343E+00 4.4318E-03 1.663702784 1.7907E-02 1.848182597 2.9812E-02 1.972738181 4.0850E-02 2.0687069 5.1266E-02 2.147687153 6.1187E-02 2.215323077 7.0697E-02 2.274808354 7.9849E-02 2.328134588 8.8687E-02 2.376631603 9.7241E-02 2.421233645 1.0554E-01 2.46262341 1.1359E-01 2.501315693 1.2143E-01 PDF(x) 5.3305E-03 2.0884E-02 3.4321E-02 4.6623E-02 5.8124E-02 6.8999E-02 7.9355E-02 8.9268E-02 9.8790E-02 1.0797E-01 1.1682E-01 1.2539E-01 1.3370E-01 if you use the Excel Function 'Kurt' to compute the Kurtosis you must add 3 to get the correct value for entry into cell c9 CDF(x) 1.3499E-03 6.3615E-03 1.1373E-02 1.6385E-02 2.1396E-02 2.6408E-02 3.1419E-02 3.6431E-02 4.1442E-02 4.6454E-02 5.1465E-02 5.6477E-02 6.1489E-02 0.4008 Raytheon PDF 0.3611 0.3214 0.2817 PDF(x) Var sk k 0.2419 0.2022 0.1625 0.1228 0.0831 0.0434 0.0037 1.2343 1.8359 2.4376 3.0392 3.6409 4.2425 4.8442 5.4459 6.0475 6.6492 7.2508 LL 2.1268 Yield(%) = 95.00% UL= 6.0452 Press ENTER if you change Lower or Upper Limits or Change Targeted Yield Right Click on/near the axis values to change the number formats 1.0000 CDF 0.9000 0.8000 0.7000 CDF(x) 20 20.6 19.4 18.2 18.8 17 17.6 0 0.6000 0.5000 0.4000 0.3000 0.2000 0.1000 0.0000 1.2343 1.8359 2.4376 3.0392 3.6409 4.2425 4.8442 5.4459 6.0475 6.6492 7.2508 Lecture 3: PDFs 38 0.8 0.7 0.6 0.5 JFIT Tool(Mode 1) 0.4 0.3 0.2 0.1 23 Compute Var sk k Type Probs? Insert values theC6-9 Steps: for the four moments Enter Datainto in Either 2. Set Plot and Rangethe (D22-D top set of cells then press compute 3. Enter Limits Or Targe results appear in the lower set ofcompute cells. 4. Press First Four Central Moments of The Data 4.00000 Mean 1.00000 Variance 0.18226 Skew 3.05910 Kurtosis Do NOT insert data in both C6:C9 and C13:C17 Output Parameters of Johnson Curve if data inserted in C6:C9. -46.19231651 Types: Problems: 16.49535715 JSL=LogNormal 0 None 1 Variance <= 0 1 JSU=Unbounded -12.48020369 JSB=Bounded 2 Kurt < Skew^2 + 1 3 Sb did not converge JSN=Normal JSL 4 Inconsistent Input Data 0 JST=Two-Ordinate 5 SU did not converge 0.4 0.3 0.3 0.2 PDF(x) 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 0.2 0.2 0.1 0.1 0.0 0.0 0.0 Lecture 3: PDFs 39 0.8 0.7 0.6 0.5 JFIT Tool (Mode 2) 0.4 0.3 0.2 0.1 23 22.4 Remove values for the four moments in the top set of cells. known values for Johnson parameters First Four CentralInsert Moments of The Data Mean in the lower set of cells. 0.2319 Variance 0.2088 Press compute and the pdf and CDF graphs Skew 0.1856 Kurtosis appear for that particular distribution 0.1625 Do NOT insert data in both C6:C9 and C13:C17 PDF(x) Type Probs? Output Parameters of Johnson Curve if data inserted in C6:C9. 3.193742536 Types: Problems: PDF 0.2319 1.195191141 JSL=LogNormal 0 None 0.2088 1 Variance <= 0 45.23630054 JSU=Unbounded 0.1856 -0.776525532 JSB=Bounded 2 Kurt < Skew^2 + 1 0.1625 3 Sb did not converge JSN=Normal JSB 0.1394 4 Inconsistent Input Data JST=Two-Ordinate 0.1162 5 SU did not converge 0.0931 PDF(x) Var sk k 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 0.1394 0.1162 0.0931 0.0699 0.0468 0.0236 0.0005 -0.5239 1.5 0.0699 0.0468 0.0236 0.0005 -0.523 1.5297 3.5833 5.6370 7.6906 9.7442 11.797 13.851 15.905 17.958 20.012 9 9 5 1 8 4 Lecture 3: PDFs 40 0.8 0.7 0.6 0.5 JFit TABS (Bottom of JFit Screen) 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Johnson ST, Two Ordinate Distributions Johnson SN, Normal Distributions Johnson SB – Bounded Distributions Johnson SU – Unbounded Distributions Johnson SL – Logarithmic Distributions Lecture 3: PDFs 41 0.8 0.7 0.6 0.5 Johnson Type 1: Log Normal (JSL) 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Lecture 3: PDFs 42 0.8 0.7 0.6 0.5 Johnson Type 2 Unbounded (JSU) 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Lecture 3: PDFs 43 0.8 0.7 0.6 0.5 Johnson Type 3 Bounded (JSB) 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Lecture 3: PDFs 44 0.8 0.7 0.6 0.5 Johnson Type 4 Normal (JSN) 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Lecture 3: PDFs 45 0.8 0.7 0.6 0.5 Johnson Type 5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Johnson ST Fit: Two Ordinate Distribution This is the Histogram corresponding to g=0 h = . 80 e = .022 l = .078 Lecture 3: PDFs 46 0.8 0.7 0.6 0.5 Using JFIT To Estimate PNC or Yield 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Lecture 3: PDFs 47 0.8 0.7 0.6 0.5 Acronyms and Abbreviations 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 AD - Anderson Darling CLT - Central Limit Theorem JFIT - Johnson Fitting Tool KS - Kolmogorov Smirnoff LSL - Lower Specification Limit PDF -Probability Density Function PNC - Probability of Non Compliance USL - Upper Specification Limit Lecture 3: PDFs 48 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 End of PDF Fitting Module (Lecture 3) Back up slides Lecture 3: PDFs 49 0.8 0.7 0.6 0.5 Distribution Fitting - Preliminary Steps 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 This article covers several steps you should consider taking before you analyze your probability data and apply the analysis results. – Step 1 - Define The Goals Of Your Analysis The very first step is to define what you are trying to achieve by analyzing your data. You should have a clear understanding of your goals as this will help you throughout the entire data analysis process. Try answering the following questions: What kind of information would you like to obtain? How will you obtain the information you need? How will you apply that information? Lecture 3: PDFs 50 0.8 0.7 0.6 0.5 Example: 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Robert is the head of the Customer Support Department at a large company. In order to reduce the customer service times and improve the customer experience, he would like to do the following: Determine the probability that a customer can be served in 5 minutes or less. To solve this problem, Robert needs to: – Perform distribution fitting to sample data (customer service times) for a selected period of time (e.g. last week) – Select the best fitting distribution – Calculate the probability using the cumulative distribution function of the selected distribution If the probability is less than 95%, consider hiring additional customer support staff Lecture 3: PDFs 51 0.8 0.7 0.6 0.5 Step 2 - Prepare Data For Distribution Fitting 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Preparing your data for distribution fitting is one of the most important steps you should take, since the analysis results (and thus the decisions you make) depend on whether you correctly collect and specify the input data. – Data Format Your data might come in one of the generally accepted formats, depending on the source of data and how it was collected. You need to make sure the distribution fitting software you are using supports the data format you need, and if it doesn't, you might need to convert your data to one of the supported formats. The most commonly used format in probability data analysis is an unordered set of values obtained by observing some random process. The order of values in a data set is not important and does not affect the distribution fitting results. This is one of the fundamental differences between distribution fitting (and probability data analysis in general) and time series analysis where each data value is connected to some timepoint at which this value was observed. – Sample Size The rule of thumb is the more data you have, the better. In most cases, to get reliable distribution fitting results, you should have at least 75-100 data points available. Note that very large samples (tens of thousands of data points) might cause some computational problems when fitting distributions to data, and you might need to reduce the sample size by selecting a subset of your data. However, in many cases one has only 10 to 30 data points. Expect larger variances! Lecture 3: PDFs 52 0.8 0.7 0.6 0.5 Step 3 - Decide Which Distributions To Fit 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Before fitting distributions to your data, you should decide which distributions are appropriate based on the additional information about the data you have. This can be helpful to narrow your choice to a limited number of distributions before you actually perform distribution fitting. – Data Domain - Continuous or Discrete? The easiest part is to determine whether your data is continuous or discrete. If your data can take on real values (for example, 1.5 or -2.33), then you should consider continuous distributions only. On the other hand, if your data can take on integer values (1, 2, -5 etc.) only, then you might want to fit both continuous and discrete distributions. The reason to use continuous distributions to analyze discrete data is that there is a large number of continuous distributions which frequently provide much better fit than discrete distributions. However, if you are confident that your random data follows a certain discrete distribution, you might want to use that specific distribution rather than continuous models. – The Nature of Your Data In most cases, you have not just raw data, you also have some additional information about the data and its properties, how the data was collected etc. This information might be very useful to narrow your choice to several probability distributions. – Example. If you are analyzing the sales data of a company, it should be clear that this kind of data cannot contain negative values (unless the company sells at a loss), and thus it wouldn't make much sense to fit distributions which can take on negative values (such as the Normal distribution) to your data. In addition, some particular distributions are recommended for use in several specific industries. An obvious example of such an industry is reliability engineering which makes great use of the Weibull distribution and several additional models (Exponential, Lognormal, Gamma) to perform the analysis of failure data. These distributions are widely used in many other industries, but in reliability engineering they are considered "standard". Lecture 3: PDFs 53 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Additional Distributions Lecture 3: PDFs 54 0.8 0.7 0.6 0.5 Geometric distribution 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Probability that an event occurs on the kth trial. – Let p = probability of event occurring in a single trial. – f(k)= probability of event occurring on kth attempt.= (1-p)k-1p the is called the probability mass function or pmf. Example: Find probability a capacitor shorts due to puncturing of dielectric after there have been k voltage spikes. If the probability of a puncture on the jth voltage spike is pj, then the probability of puncture on the kth spike is – f(k)=(1-p1)(1-p2)…(1-pk-1)pk If all the p values are the same then f(k)=(1-p)k-1p which is the geometric distribution probability mass function or pmf. Lecture 3: PDFs 55 0.8 0.7 0.6 0.5 Hypergeometric distribution 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Hypergeometric distribution (use in sampling w/o replacement) M N M M mean np , p x P(X=x,n n x | N,M) = N f ( x) P( X x) N npq N n N n 2 n np (1 p ) ( N 1) N 1 N population size M=# bad units in population n = sample size drawn from N w/o replacement x = # failed units in sample Ex. N=600 missiles in inventory, M=80 missiles were worked on by Operator #323 and they are believed to have been defectively assembled by this operator. If we take n=20 missiles randomly from the inventory, then f(x) = probability of x defective missiles in the sample. f(0)=0.054 Lecture 3: PDFs 56 0.8 0.7 0.6 0.5 Poisson Distribution 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Used extensively for situations involving counting of events – e.g. probability of x false alarms during time period T=60 seconds when the average alarm rate is (false alarms/ sec). – The pmf is f(k) = (T)k*exp(-T) / k!= (60)k*exp(-60) / k! E[k]=1/T,1/T1/2 Also used in predicting number of failed units over time period TW, given the mean time to failure MTTF =1/. – f(k units failing in time TW)= (TW)k*exp(-TW)/k!. If want a 90% probability of x or fewer failures in time span TW , then find – F(x)=0.90= exp(-TW) + (TW)1*exp(-TW)/1! + (TW)2*exp(-TW)/2! + … + (TW)x*exp(-TW)/x! (solve numerically for TW the warranty period.) – If you were considering the case of x=0 failures during the warranty period, then you could set the warranty period = TW =-ln(0.9)/ ~ 0.105/ years. If your warranty period TW = 1 year and the MTTF=10 years then you should expect only 1- exp(-.1*1) ~ 9.5% of units produced to have one or more failures during that time. Lecture 3: PDFs 57 0.8 0.7 0.6 0.5 Uniform Discrete Distribution 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Assume b > a integers f(x|a,b) = 1/(b-a+1) for a≤ x ≤ b. n uniformly distributed pts. = 0, otherwise a a+1 k b F(x)= 0, x≤a = (x-a+1)/(b-a+1), a≤ x ≤ b =1, x≥b There are n=b-a+1 discrete uniformly distributed data points – Mean = (a+b)/2 – Var(X) = [(b-a+1)2-1]/12 Lecture 3: PDFs 58 0.8 0.7 0.6 0.5 Continuous distributions 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17.6 17 0 Logistics distribtion f (x) 0.6 1 xa b e b 1 e 0.5 0.4 0.3 0.2 0.1 0 -2 -1 0 1 2 3 4 F (x) xa b 2 1 xa b e b 1 e xa b 2 m e a n = a = m e d ia n = m o d e 1 1 e xa b v a ria n c e p 3 4 2 b 2 3 0 4 .2 Gamma distribution (Erlang for = integer) x 1e x / ,x 0 f ( x) P( X x) 0 ,x0 mean 2 2 variance 2 3 1 distribution of time up to having exactly =k events 3 2 occur. Used in queueing theory, reliability, etc.. 4 2 Lecture 3: PDFs 59 0.8 0.7 0.6 0.5 Beta & Weibull distributions 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 = + Beta distribution x 1 1 x f ( x) B( , ) 0 B( , ) 1 ,0 x 1 2 ,otherwise ( )( ) beta function ( ) xmode 2 1 + 2 + 1 1 + 2 0 0.5 Weibull distribution f (t ) t 1 e 0 t , , 0, t 0 ,otherwise F (t ) P (T t ) 1 e t 1 1 2 1 2 2 1 2 1 Exponential distribution e t , 0, t 0 f (t ) 0 ,otherwise F (t ) P(T t ) 1 e t 1 2 1 2 3 1 2 4 2 9 Lecture 3: PDFs 60 0.8 0.7 0.6 0.5 Rayleigh & Triangle distributions 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Rayleigh distribution x f ( x) 2 1 x 2 2 , 0, x 0 e 0 ,otherwise F ( x) P( X x) 1 e 1 x 2 2 p 2 2 2 2 1 p 3 1 4 3 p 2 p2 3/ 2 p 2 0.63 4 2 32 3p 2 /(4 p ) 2 3.245 Triangle ditribution 2 xa b a m a , a X m f ( x) 2 b x , b X m b a b m a= lower limit b = upper limit m = most probable or mode a m b a m b / 3 2 a 2 m 2 b 2 ab am bm /18 3 b a / 3 3 10 (m a) (b m) (m a) 1 1 1 2 (b a) (b a) b a 4 324 135 2.4 Lecture 3: PDFs 61 0.8 0.7 0.6 0.5 Lognormal and Normal distributions 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Lognormal distribution 2 ln x ln 2ln 1 ln =mean e MTTF 2 ln 1 2 e ,0 x ln 2 2 f ( x) =variance = e 1 x ln 2p , otherwise m =median e ln 0 2 Normal distribution (Gaussian distribution) 1 f ( x) e 2p 1 x 2 2 mean 2 variance std dev 3 0 1 coef. of skewness 4 3 2 coef, of kurtosis Many statistical process control calculations assume normal distributions and incorrectly so! Standard Normal distribution Any normal distribution can be scaled to the unit normal distribution. Useful when using tables to look up values Used to model mean time to repair for availability calculations (Unit Normal distribution) mean=0 1 12 z2 f ( z) e z 2 variance=1 2p std dev=1 define z = (x - ) / 3 0 1 coef. of skewness 4 3 2 coef, of kurtosis Lecture 3: PDFs 62 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 23 22.4 21.8 21.2 20 20.6 19.4 18.2 18.8 17 17.6 0 Mixed distribution. Used to model bimodal processes. f(x)=pg1(x) + (1-p)g2(x) p1 (1 p) 2 2 2 p( 12 12 ) (1 p)( 22 22 ) 3 S p( S1 13 3 121 13 ) (1 p)(S 2 23 3 22 2 23 ) 4 K p( K1 14 4S1 131 6 1212 14 ) (1 p)(K1 24 4S1 23 2 6 22 22 24 ) Definitions. xf ( x)dx, j xg j ( x)dx, r x r f ( x)dx, rj x j r g j ( x)dx, j j , 2 1 2 1 12 21 x 1 2 g1 ( x)dx, 12 22 x 2 2 g 2 ( x)dx, S j 3 j x j g j ( x)dx, K j 4 j x j g j ( x)dx, 3 4 Lecture 3: PDFs 63