Release 11.0 Documentation for ANSYS
www.kxcad.net Home > CAE Index > ANSYS Index > Release 11.0 Documentation for ANSYS
This section presents useful guidelines for defining your probabilistic design variables.
Here are a few tips you can use to determine which random input variable in your finite element model follows which distribution function and what the parameters of the distribution function are.
First, you should know to
Specify a reasonable range of values for each random input variable.
Set reasonable limits on the variability for each RV.
The values and hints given below are simply guidelines; none are absolute. Always verify this information with an expert in your organization and adapt it as needed to suit your analysis.
The number of simulation loops that are required for a Monte Carlo Simulation does not depend on the number of random input variables. The required number of simulation loops only depends on the amount of the scatter of the output parameters and the type of results you expect from the analysis. In a Monte Carlo Simulation, it is a good practice to include all of the random input variables you can think of even if you are not sure about their influence on the random output parameters. Exclude only those random input variables where you are very certain that they have no influence. The probabilistic design system will then automatically tell you which random input variables have turned out to be significant and which one are not. The number of simulations that are necessary in a Monte Carlo analysis to provide that kind of information is usually about 50 to 200. However, the more simulation loops you perform, the more accurate the results will be.
The number of simulation loops that are required for a Response Surface analysis depends on the number of random input variables. Therefore, you want to select the most important input variable(s), the ones you know have a significant impact on the random output parameters. If you are unsure which random input variables are important, it is usually a good idea to include all of the random variables you can think of and then perform a Monte Carlo Simulation. After you learn which random input variables are important and therefore should be included in your Response Surface Analysis, you can eliminate those that are unnecessary.
The type and source of the data you have determines which distribution functions can be used or are best suited to your needs.
If you have measured data then you first have to know how reliable that data is. Data scatter is not just an inherent physical effect, but also includes inaccuracy in the measurement itself. You must consider that the person taking the measurement might have applied a "tuning" to the data. For example, if the data measured represents a load, the person measuring the load may have rounded the measurement values; this means that the data you receive are not truly the measured values. Depending on the amount of this "tuning," this could provide a deterministic bias in the data that you need to address separately. If possible, you should discuss any bias that might have been built into the data with the person who provided that data to you.
If you are confident about the quality of the data, then how to proceed depends on how much data you have. In a single production field, the amount of data is typically sparse. If you have only few data then it is reasonable to use it only to evaluate a rough figure for the mean value and the standard deviation. In these cases, you could model the random input variable as a Gaussian distribution if the physical effect you model has no lower and upper limit, or use the data and estimate the minimum and maximum limit for a uniform distribution. In a mass production field, you probably have a lot of data, in which case you could use a commercial statistical package that will allow you to actually fit a statistical distribution function that best describes the scatter of the data.
The mean value and the standard deviation are most commonly used to describe the scatter of data. Frequently, information about a physical quantity is given in the form that its value is; for example, "100±5.5". Often, but not always, this form means that the value "100" is the mean value and "5.5" is the standard deviation. To specify data in this form implies a Gaussian distribution, but you must verify this (a mean value and standard deviation can be provided for any collection of data regardless of the true distribution type). If you have more information (for example, you know that the data must be lognormal distributed), then the PDS allows you to use the mean value and standard deviation for a definition of a lognormal distribution.
Sometimes the scatter of data is also specified by a mean value and an exceedence confidence limit. The yield strength of a material is sometimes given in this way; for example, a 99% exceedence limit based on a 95% confidence level is provided. This means that derived from the measured data we can be sure by 95% that in 99% of all cases the property values will exceed the specified limit and only in 1% of all cases they will drop below the specified limit. The supplier of this information is using mean value, the standard deviation, and the number of samples of the measured data to derive this kind of information. If the scatter of the data is provided in this way, the best way to pursue this further is to ask for more details from the data supplier. Because the given exceedence limit is based on the measured data and its statistical assessment, the supplier might be able to provide you with the details that were used.
If the data supplier does not give you any further information, then you could consider assuming that the number of measured samples was large. If the given exceedence limit is denoted with x1 - α/2 and the given mean value is denoted with xμ then the standard deviation can be derived from the equation:
where the values for the coefficient C are:
| Exceedence Probability | C |
|---|---|
| 99.5% | 2.5758 |
| 99.0% | 2.3263 |
| 97.5% | 1.9600 |
| 95.0% | 1.6449 |
| 90.0% | 1.2816 |
In situations where no information is available, there is never just one right answer. Below are hints about which physical quantities are usually described in terms of which distribution functions. This might help you with the particular physical quantity you have in mind. Also below is a list of which distribution functions are usually used for which kind of phenomena. Keep in mind that you might need to choose from multiple options.
If you are designing a prototype, you could assume that the actual dimensions of the manufactured parts would be somewhere within the manufacturing tolerances. In this case it is reasonable to use a uniform distribution, where the tolerance bounds provide the lower and upper limits of the distribution function.
Sometimes the manufacturing process generates a skewed distribution; for example, one half of the tolerance band is more likely to be hit than the other half. This is often the case if missing half of the tolerance band means that rework is necessary, while falling outside the tolerance band on the other side would lead to the part being scrapped. In this case a Beta distribution is more appropriate.
Often a Gaussian distribution is used. The fact that the normal distribution has no bounds (it spans minus infinity to infinity), is theoretically a severe violation of the fact that geometrical extensions are described by finite positive numbers only. However, in practice this is irrelevant if the standard deviation is very small compared to the value of the geometric extension, as is typically true for geometric tolerances.
Very often the scatter of material data is described by a Gaussian distribution.
In some cases the material strength of a part is governed by the "weakest-link-theory". The "weakest-link-theory" assumes that the entire part would fail whenever its weakest spot would fail. for material properties where the "weakest-link" assumptions are valid, then the Weibull distribution might be applicable.
For some cases, it is acceptable to use the scatter information from a similar material type. Let's assume that you know that a material type very similar to the one you are using has a certain material property with a Gaussian distribution and a standard deviation of ±5% around the measured mean value; then let's assume that for the material type you are using, you only know its mean value. In this case, you could consider using a Gaussian distribution with a standard deviation of ±5% around the given mean value.
For temperature-dependent materials it is prudent to describe the randomness by separating the temperature dependency from the scatter effect. In this case you need the mean values of your material property as a function of temperature in the same way that you need this information to perform a deterministic analysis. If M(T) denotes an arbitrary temperature dependent material property then the following approaches are commonly used:
Multiplication equation:
M(T)rand = Crand
(T)
Additive equation:
M(T)rand =
Linear equation:
M(T)rand = Crand
Here, (T) denotes the mean value
of the material property as a function of temperature. In the "multiplication
equation" the mean value function is scaled with a coefficient Crand and
this coefficient is a random variable describing the scatter of the material
property. In the "additive equation" a random variable ΔMrand is
added on top of the mean value function (T).
The "linear equation" combines both approaches and here both Crand and ΔMrand are
random variables. However, you should take into account that in general for
the "linear equation" approach Crand and ΔMrand are,
correlated.
Deciding which of these approaches is most suitable to describing the scatter of the temperature dependent material property requires that you have some raw data about this material property. Only by reviewing the raw data and plotting it versus temperature you can tell which approach is the better one.
For loads, you usually only have a nominal or average value. You could ask the person who provided the nominal value the following questions: If we have 1000 components that are operated under real life conditions, what would the lowest load value be that only one of these 1000 components is subjected to and all others have a higher load? What would the most likely load value be, i.e. the value that most of these 1000 components have (or are very close to)? What would the highest load value be that only one of the 1000 components is subjected to and all others have a lower load? To be safe you should ask these questions not only of the person who provided the nominal value, but also to one or more experts who are familiar with how your products are operated under real-life conditions. From all the answers you get, you can then consolidate what the minimum, the most likely, and the maximum value probably is. As verification you can compare this picture with the nominal value that you would use for a deterministic analysis. If the nominal value does not have a conservative bias to it then it should be close to the most likely value. If the nominal value includes a conservative assumption (is biased), then its value is probably close to the maximum value. Finally, you can use a triangular distribution using the minimum, most likely, and maximum values obtained.
If the load parameter is generated by a computer program then the more accurate procedure is to consider a probabilistic analysis using this computer program as the solver mechanism. Use a probabilistic design technique on that computer program to assess what the scatter of the output parameters are, and apply that data as input to a subsequent analysis. In other words, first run a probabilistic analysis to generate an output range, and then use that output range as input for a subsequent probabilistic analysis.
Here, you have to distinguish if the program that generates the loads is ANSYS itself or your own in-house program. If you have used ANSYS to generate the loads (for example, FLOTRAN analysis calculating fluid loads on a structure or a thermal analysis calculating the thermal loads of a structure) then we highly recommend that you include these load calculation steps in the analysis file (and therefore in the probabilistic analysis). In this case you also need to model the input parameters of these load calculation steps as random input variables. If you have used your own in-house program to generate the loads, you can still integrate the load calculation program in the analysis file (see the /SYS command for details), but you must have an interface between that program and ANSYS that allows the programs to communicate with each other and thus automatically transfer data.
You also have to distinguish if the load values are random fields or single random variables. If the load is different from node to node (element to element) then it is most appropriate to include the program calculating the load in the analysis file. If the load is described by one or very few constant values then you can also consider performing a probabilistic analysis with the program calculating these load values. Again you need to provide an interface to transfer input data to this program and get output data (the loads) back to ANSYS. If there is more than just one single load value generated by the program then you should also check for potential correlations.
The Beta distribution is very useful for random variables that are bounded at both sides. If linear operations are performed on random variables that are all subjected to a uniform distribution then the results can usually be described by a Beta distribution. An example is if you are dealing with tolerances and assemblies, where the components are assembled and the individual tolerances of the components follow a uniform distribution. In this case the overall tolerances of the assembly are a function of adding or subtracting the geometrical extension of the individual components (a linear operation). Hence, the overall tolerances of the assembly can be described by a Beta distribution. Also, as previously mentioned, the Beta distribution can be useful for describing the scatter of individual geometrical extensions of components as well. The uniform distribution is a special case of the Beta distribution.
The exponential distribution is useful in cases where there is a physical reason that the probability density function is strictly decreasing as the random input variable value increases. The distribution is mostly used to describe time-related effects; for example, it describes the time between independent events occurring at a constant rate. It is therefore very popular in the area of systems reliability and lifetime-related systems reliability, and it can be used for the life distribution of non-redundant systems. Typically, it is used if the lifetime is not subjected to wear-out and the failure rate is constant with time. Wear-out is usually a dominant life-limiting factor for mechanical components, which would preclude the use of the exponential distribution for mechanical parts. However in cases where preventive maintenance exchanges parts before wear-out can occur, then the exponential distribution is still useful to describe the distribution of the time until exchanging the part is necessary.
The Gamma distribution is again a more time-related distribution function. For example it describes the distribution of the time required for exactly k events to occur under the assumption that the events take place at a constant rate. It is also used to describe the time to failure for a system with standby components.
The Gaussian or normal distribution is a very fundamental and commonly used distribution for statistical matters. It is typically used to describe the scatter of the measurement data of many physical phenomena. Strictly speaking, every random variable follows a normal distribution if it is generated by a linear combination of a very large number of other random effects, regardless which distribution these random effects originally follow. The Gaussian distribution is also valid if the random variable is a linear combination of two or more other effects if those effects also follow a Gaussian distribution.
The lognormal distribution is a basic and commonly used distribution. It is typically used to describe the scatter of the measurement data of physical phenomena, where the logarithm of the data would follow a normal distribution. The lognormal distribution is very suitable for phenomena that arise from the multiplication of a large number of error effects. It is also correct to use the lognormal distribution for a random variable that is the result of multiplying two or more random effects (if the effects that get multiplied are also lognormally distributed). If is often used for lifetime distributions; for example, the scatter of the strain amplitude of a cyclic loading that a material can endure until low-cycle-fatigue occurs is very often described by a lognormal distribution.
The uniform distribution is a very fundamental distribution for cases where no other information apart from a lower and an upper limit exists. If is very useful to describe geometric tolerances. It can also be used in cases where there is no evidence that any value of the random variable is more likely than any other within a certain interval. In this sense it can be used for cases where "lack of engineering knowledge" plays a role.
The triangular distribution is most helpful to model a random variable when actual data is not available. It is very often used to cast the results of expert-opinion into a mathematical form, and is often used to describe the scatter of load parameters. However, regardless of the physical nature of the random variable you want to model, you can always ask some experts questions like "What is the one-in-a-thousand minimum and maximum case for this random variable? and other similar questions. You should also include an estimate for the random variable value derived from a computer program, as described earlier. This is also described in more detail above for load parameters in Choosing a Distribution for a Random Variable.
The truncated Gaussian distribution typically appears where the physical phenomenon follows a Gaussian distribution, but the extreme ends are cut off or are eliminated from the sample population by quality control measures. As such, it is useful to describe the material properties or geometric tolerances.
In engineering, the Weibull distribution is most often used for strength or strength-related lifetime parameters, and it is the standard distribution for material strength and lifetime parameters for very brittle materials (for these very brittle material the "weakest-link-theory" is applicable). For more details see Choosing a Distribution for a Random Variable.
Output parameters are usually parameters such as length, thickness, diameter, or model coordinates.
The ANSYS PDS does not restrict you with regard the number of random output parameters, provided that the total number of probabilistic design variables (that is random input variables and random output parameters together) does not exceed 5000.
ANSYS recommends that you include all output parameters that you can think of and that might be useful to you. The additional computing time required to handle more random output parameters is marginal when compared to the time required to solve the problem. It is better to define random output parameters that you might not consider important before you start the analysis. If you forgot to include a random output parameter that later turns out to be important, you must redo the entire analysis.