A short script to fit a W-L-F model to experimental data. This is flaky, depending on the input guesses and convergence is poor. We simply stop the convergence at the point we have an acceptable fit. This is not robust!

This R script is useful for creating average curves of experimental material property data. it simply breaks the multiple curve data into discrete chunks and averages the values within each chunk. The size of the window is customizable. Additionally, it also generates a lowess fit for the resultant dataset, applying additional smoothing if desired.

# R Script to compute an average curve from material test data
options(digits = 5)

# Read raw data (data should be in 2 colums) and sort by x-value.
# Data should be offset to the origin and all but a single 0,0 point should be included)
Data <- read.csv("Z:/Christopher/15/141531/Materials/Planar_23.csv", header=FALSE)
Data.sort <- Data[order(Data$V1),]

# Define the number of samples to generate
points <- 19
samples <- floor((length(Data.sort$V1)-1)/points)

# Initialize the matrix (we assume that the )
Result <- matrix(nrow = points, ncol = 2)
Result[1,] <- c(0,0)

# Loop over the data computing the average of groups of data points
# defined by the width of the sample window
i <- 1
x <- 1
for(x in 2:points){
V1 <- V2 <- 0.0
for(y in 1:samples){
V1 <- V1 + Data.sort[y+i,1]
V2 <- V2 + Data.sort[y+i,2]
}
i<-i+samples
Result[x,] <- c(V1/samples,V2/samples)
}

# Plot the raw data and the averaged curve in red
plot(Data.sort, pch = 1)
lines(Result, col = "red", lwd = 3)

# Create a lowess smoothed curve and plot in blue.
# This is helpful to remove noise in the average especially with a
# large number of sampled points
Smooth <- (lowess(Result, f = .3))
lines(Smooth, col = "blue", lwd = 3)

# Perform some data manipulation and export as tab delimited text files
Smooth<-(do.call(rbind, lapply(Smooth, unlist)))
Smooth[,1] <- c(0,0)
write(t(Result), file = "Planar_23_avg.txt", ncolumns = 2, sep = "\t")
write(Smooth, file = "Planar_23_smth.txt", ncolumns = 2, sep = "\t")

In this post we will look at the procedure for determining the Mooney-Rivlin constants from simple tensile test data of an elastomeric solid. The definition and derivation of the material model is left to others. For our purposes, all we need to know is that the material model yields a predicted engineering stress under simple tension of

where C_{1} and C_{2} are the Constants that need to determine and is the stretch ratio, defined as the ratio of the stretched length to the initial length of the sample. This can be defined in terms of the measured strain in a simple tensile test

We’ll start with a set pf published stress strain data for a 40 Shore A material from GLS corporation. The stress strain curve from the literature is shown below.

Using g3data we can extract the points and tabulate them. For simplicity, we also convert the strain to stretch ratio and the stress to SI units. The data file can be downloaded here.

3. The "attach" command allows us to access the variables directly without having to reference the original structure.

> attach(SS_Data)

4. A quick plot of the imported data can then be generated.

> plot(Strain, Stress)

5. Since the MR model uses the stretch ratio, not the strain, we convert the strains and then plot the stress vs. stretch ratio.

> Stretch = Strain + 1
> plot (Stretch, Stress)

6. Now for the curve fitting itself. We use the "nls" function which stands for "Nonlinear Least Squares". We provide the model as given in the equation at the beginning of the post where C1 & C2 are the two constants we wish to determine, guesses for the initial values of those constants, and request that the trace is provided. Then we ask for a summary of the results. This prints the fitted values for our two constants along with some other helpful data.

7. Lastly, we would like to see how our curve fit matches the data. First we extract the coefficients into a vector "C" With the plot shown above still open, we add the curve and clean it up with a title. Notice the method for calling the coefficients.

> C <- coef(MR.fit)
> curve((2*C[1]+2*C[2]/x)*(x-1/(x^2)), from=1.0, to = 4.0, add=TRUE)
> title(main="Mooney-Rivlin Fit to Simple Tensile Data")

Caveat lector — All work and ideas presented here may not be accurate and should be verified before application.

In this post we discussed the use of the Taylor Series to evaluate the propagated uncertainties in measurements to a calculated result. We can also use a similar approach to determine the sensitivity of a transfer function to know perturbations in the independent variables.

In this case we will look at the volume of a right conical frustum or a truncated cone. The volume can be calculated from the three variables: , and shown in the figure. The volume is given as

The first step is to calculate the partial derivatives of with respect to the independent variables.

The total uncertainty in the volume, , is given in terms of the uncertainties in the independent variables, , and .

Each term of the form is the contribution to the total perturbation of the function by the perturbation of the independent variable, . Therefore we can calculate the sensitivities, , of the volume to small changes in each of the three variables , and , by calculating the percentage contribution of each term to the total perturbation. Typically, we wish to also define the direction that the pertubations in the independent variables will affect the total, so we remove the absolute values and evaluate the partial derivatives while maintaining their signs.

It is also evident that .

Example:

Suppose we have a design for a conical frustum with nominal dimensions of 0.500 in, 0.375 in and 1.250 in. Our manufacturing process can hold to ±.002 in, to ± .007 in and to ± .010 in. We then have the following inputs to our formulae.

From this we can calculate the nominal volume of our manufactured part.

Next we can calculate the partial derivatives.

And having the partial derivatives, we can calculate the total propagated uncertainty in the nominal volume.

From this we can obtain the predicted volume of the manufactured part subject to the manufacturing tolerances: 0.757 ± .021 in^{3}. We can then calculate the relative sensitivity of the total volume to the manufacturing tolerances specified.

It is important to note that these results are only valid for the initial conditions specified. If the nominal values for the dimensions or the tolerances change, both the uncertainties and sensitivities will change.

Caveat lector — All work and ideas presented here may not be accurate and should be verified before application.

This is a basic introductory look at using R for generating descriptive statistics of a univariate data set. Here, we will use the historical dataset of Michelson’s experiment to determine the speed of light in air provided as a an ASCII file with header content and the observed speed of light for 100 trials.

We need to first read the data into R. Since the data is in a properly formatted ASCII file, we only need to tell R to ignore the first 60 lines, which is header information. R will then import the data into a list of class data.frame.

>C <- read.table("Michelso.dat",skip=60)

We can take a look at the dataset by simply typing the dataset name at the prompt. Here you can see that R automatically assigned the variable V1 to the data.

> C
V1
1 299.85
2 299.74
3 299.90
4 300.07
...

The summary() command in R provides the summary statistics: MIn, 1st Q, Median, Mean, 3rd Q and Max. We call this function with the argument 'C$V1' which tells R to act on the named variable, V1, in the data.frame C. (The options commands set the output number formatting to something realistic.)

> options(scipen=100)
> options(digits=10)
> summary(C$V1)
Min. 1st Qu. Median Mean 3rd Qu. Max.
299.6200 299.8075 299.8500 299.8524 299.8925 300.0700

Standard deviation, trimmed mean and number of data points can be obtained individually.

To determine confidence intervals on the mean, we can use the one sample t-test. We can ignore the mean value to test against since in our case it is not known (or relevant for confidence interval estimation)

> t.test(C$V1, conf.level=0.99)
One Sample t-test
data: C$V1
t = 37950.9329, df = 99, p-value < 0.00000000000000022
alternative hypothesis: true mean is not equal to 0
99 percent confidence interval:
299.8316486 299.8731514
sample estimates:
mean of x
299.8524

Another method for obtaining much of this information in a single step can be found in the stat.desc() function from the pastecs package.

> install.packages("pastecs")
> library(pastecs)
...
> options(scipen=100)
> options(digits=4)
> stat.desc(C)
V1
nbr.val 100.0000000
nbr.null 0.0000000
nbr.na 0.0000000
min 299.6200000
max 300.0700000
range 0.4500000
sum 29985.2400000
median 299.8500000
mean 299.8524000
SE.mean 0.0079011
CI.mean.0.95 0.0156774
var 0.0062427
std.dev 0.0790105
coef.var 0.0002635

We'll look at the generation of some standard statistical plots for exploratory data analysis in a future post.

Caveat lector — All work and ideas presented here may not be accurate and should be verified before application.

In courses on experimentation, propagated errors are typically treated through the use of a Taylor series expansion to evaluate the total contribution of individual measurement uncertainties to a final calculated result. As an example, suppose we wish to experimentally determine the acceleration of a body due to gravity. We could take an object and drop it a measured distance while recording the elapsed time. From basic physics we know that the distance traveled is proportional to the time squared and that the proportionality constant is or,

Solving for the acceleration, we obtain,

which is a function of two measured variables, the distance travelled and the elapsed time. Both of these measurements, no matter how carefully obtained, will have some uncertainty. Suppose we measure the distance travelled with a ruler that has graduations every 1 inches and the time with a stopwatch with a resolution of 0.1 seconds. With both of these instruments it is evident that we cannot measure the quantity to a higher resolution than the instrument provides, therefore it is typical to take the total uncertainty in the measurement as the least significant digit in the scale, centered on the measurement value. This would equate to uncertainties in the measurements of ± .5 inches and ± .05 seconds.

Let’s say we dropped the object from a height of 36 feet and measured the elapsed time as 1.5 seconds. From the above equation we would find the acceleration to be 32 . But how carefully did we measure? Was the distance exactly 36 feet (432 inches), or was it 432.3 inches? Was the time 1.48 seconds? As long as these uncertainties are within the predefined ranges established above, we can calculate the total uncertainty in the measurement of the acceleration.

In this post we discussed the approximation of any function by a Taylor series expansion about a specific point. We can apply that technique to determine how the value of the acceleration may vary with perturbations in the input values of time and distance about the measured point. For the simplest implementation, we restrict ourselves to the first order terms of the expansion^{[1]}.

Recal that the Taylor series of a function about the point is given as

Expanding this to the first order terms yields

rewriting as

We can now see that the left side of the equation evaluates to the change in the function corresponding to a perturbation of and by a small amount and . Examining the partial derivative terms, we can see that we are multiplying the rate of change of the function in a single variable to a change in that variable from the interested point . Since we are interested in small perturbations of and about the point , we will denote these changes and . The change of the function under these perturbations we will denote .

Substituting we obtain

Lastly, since there should be no preference for the uncertainty to be in the positive or negative direction, we take the absolute value of the derivative terms and require that our perturbations be defined as positive,

This can be generalized to a function in any number of variables as

Returning to our example, to find the total uncertainty in the calculated acceleration, we simply need to determine the partial derivatives of the function in the independent variables,

and insert them into our formula

Evaluating with our collected data

or

Therefore our calculated acceleration should be given as .

It should be noted that the calculated uncertainty is the worst case situation that is possible under the individual assumptions in the treatment. Alternate treatments based on statistical treatments may be more realistic. Also, another useful application of these methods is to define the sensitivity of the dependant variable to inputs to the function. These will be address in subsequent posts.

Caveat lector — All work and ideas presented here may not be accurate and should be verified before application.

^{[1]}
2nd order and higher terms include the square and higher powers of the perturbation amount. Assuming that the perturbation is small, the square of this small perturbation is much smaller still, and higher order terms become negligible. If the relative size of the perturbation to the curvature of the function is in doubt, the magnitude of the 2nd and higher order terms should be checked.

Under certain conditions we may approximate an analytic function about a specified point on the function by an infinite series. The most useful series for our purposes is the Taylor series.

The Taylor series of a function about the point is given as

The equation above allows us to approximate the value of the function as an infinite series for any point sufficiently close to while only knowing the value of the function and its’ derivatives at . As an example, consider the equation. This is an inverted “W”-shaped function with roots at 0,. We are interested in a region centered about , so we may begin by evaluating the above equation with and for increasing values of .

For ,

For ,

For ,

etc.

For our 4th order function, any values of will result in the derivative being equal to . Therefore, we have a finite number of terms in the full Taylor series expansion up to a maximum of . The following graph shows the original function and the Taylor series approximations for . In this case, the series obtained when is algebraically equivalent to our original function.

The Taylor series may also be defined for multivariate functions. This allows us to expand the usefulness of the series to functions of multiple variables, and as we shall see later, allow us to predict the function values for small deviations about a nominal point, as well as ascribe sensitivity of the function to the independent variables.

The Taylor series of a function about the point is given as

We’ll use this later as we begin to discuss error propagation.

When speaking of attribute data in this case we are concerning ourselves with the measurement of of a certain quantity which can take on one of two values: TRUE or FALSE; 0, 1; Pass, Fail; Heads, Tails; etc. — a binary output. Probability theory gives us a simple tool for the analysis of such a system: The Binomial Distribution.

Note: The binomial distribution is valid for sampling with replacement. It is a good approximation for sampling without replacement when the parent population is large. For small populations, the hypergeometric distribution should be used.

The binomial distribution is defined by the PMF

which gives the discrete probability of n outcomes in N trials where N is the number of independent experiments yielding a binary result, n is the number of occurrences of a specified binary outcomes in N trials and p is the explicit probability that the specified outcome will occur in a single trial.

When performing experimentation to determine the probability of a product failure rate or to verify that a product will meet a specified reliability rating, we are more interested in the probabilities that the specified number of outcomes or less will occur. For this we use the cumulative probability function for a binomial distribution, defined in our case as the probability that n or less outcomes will occur. This is simply the sum of the probabilities that x outcomes will occur for

With this background, lets layout the problem. A newly designed product must meet a certain reliability rating which is defined as a maximum percentage of unit failures during operation. Our goal is to determine, through testing, and to a certain confidence level, whether the product meets this criteria. For this, we need to approach the binomial distribution from the inside-out. What we are actually specifying when we define a reliability criteria is the probability that the device will fail, or p in our equations above. Much like with a six-sided die that has a probability of presenting a given number on each roll and will, on average, present that number one out of six times over a large number of trials, we wish our product to have a specified probability (typically low) of failure. The question then becomes: How do we test for this?

Having defined the desired probability of failure of the parent population, p, and realizing that we will accept probabilities that are lower, but not higher, , we can see that the cumulative probability will give the probability of n or less failures occurring in N trials of a product that has a probability of failure of p. Thus the cumulative probability yields the likelihood that we will see n or less failures of the product purely by chance. Therefore is the probability that n or less failures would present in N trials where the probability of failure for each trial is p. We call this value our confidence.

Now we can specify our confidence that the probability of failure of each individual device id p. The question then becomes: How many trials, N, must be run and how many failures, n, are allowable in those trials? By specifying either N or n, the other can be found directly from the formula. For example, if we specify that we want to determine to a 99% confidence that 5% or less of the devices in the parent population will fail during operation () and we are willing to allow one failure during our testing (), then the planned number of test samples needs to be 130.

Unfortunately, the total number of trials cannot necessarily be specified in advance, because the laws of chance may dictate that a failure occurs in the first few samples. Therefore it is prudent to plan for at least one failure during testing. We can then compute the number of trials required for occurrences of failure of both zero and one device and may halt the testing if no failures have been recorded after the lower number of trials. In the example above, that would equate to 90 test samples.

It is useful to create a table similar to that shown below to better understand the testing requirements of a given situation. This table is set up in a slightly different format to allow the problem to be stated in a slightly different fashion. Here we look at the confidence, , in the probability that the failure will not occur, , if we see a number of observed failures, n, in the total opportunities, N.

Caveat lector — All work and ideas presented here may not be accurate and should be verified before application.