Correlation of Large Structures 1

[BATman2]

I have a large structure (60 ft by 10ft by 10ft) made of formed sections, plates, sheets and numerous welded joints (spot welds and arc welds). There are many large openings reinforced accordingly. It is designed to carry direct loads at variety of locations, in various directions. The design was analyzed via FEA using plates, shells and beam elements, medium mesh, point and distributed loads. This structure is then tested using hundreds of single and rosette strain gages,numerous deflection gages and load cells.

Question: Does a point by point correlation really make sense for a given load case given the complexity? By that I mean, if one section of the model checks out within, say 10%, clearly that can not mean the rest of the model is ok. How is this situation handled by others?

Also, what is a common criteria used by others? 10%, 15%?

Last, How many gauges (percentage) need to fall within the criteria before the model is considered "correlated".

I have guidelines for these but would like hear about others.

Thanks, Batman2

 

[EdR]

Batman2:

You are correct....Just because the results match in one location does not mean they match everywhere...You may have gages in areas of stress concentration which will cause deviation from model results (depending on how the structure is modeled)...Since the structure appears to be steel you probably know the material properties within 5% or so so I think you are in the correct ball park using 10-15% as a match criteria.....If you find differences exceeding this and cannot explain them by stress concentrations I would be looking for the reason why....

As to how many gages make it good....don't think you can quantify the results in this way....

Just my 2 cents worth

Ed.R.

 

[ProbaSci]

I suggest you consider this as a sampling comparison rather than a correlation. Demonstrate that the difference between FEA sample and Gage sample is not significant (you will need proof of high statistical power to do this) or that the observed difference is not appreciable in the real world application (ie 10- 15 %). Assuming your population (not sample) data is normal you can use a paired t-test, or if you have at least 30 samples, Central limit states it will be close enough to normal. You could get fancy a do an ANOVA and investigate differences in different parts of the sturcture, different loading scenarios, different FEA analysist

If you are concerned about too many data points, then randomly select from your population.

Correlation, to me, implies that you are demonstrating the two systems (FEA and Gage) demonstrate simmilar _response_ to a variety of inputs. This maybe more informative to what you want to do, but you will need to have measurements for multiple loading scenarios to create the correlations. Then you must be carefull, because two systems can be highly correlated and have a bias offset which will not show up in the standard correlation coefficient calculation. Correlating both systems to one input simply reduces to the comparison problem stated above.

Man do I miss getting to work on this kind of stuff. Not only will management's eyes glaze over, but so will most engineer's...


Probasci - implantable FEA

 

[BATman2]

EdR,

Thanks very much. I fully agree with your points. I will certainly pay close attention to sections of the model rather than a global correlation.

Thanks,
Batman2

 

[BATman2]

ProbSci,

While our approach is not nearly as rigorous as the stat. methods you suggest, we are looking to ensure our models are responding in the same way as the analysis. To get a true correlation coeff., I agree that a good stat. method should be used with multiple load scenarios. I would even argue that EACH load case should have its own coeff. Why? In fact the models are slightly different as the boundary conditions (reaction points and load inputs) change.

Your approach is very interesting and new to the particular field I'm in. Typically, we have 300 to 800 strain gages to work with and 6 to 8 loading scenarios, although each scenario may not induce strain at every gage. Many of the load scenarios have multiple steps up to the "full" load. In other words, plenty of data is available. Any specific references on your suggested approach, other than stat. books ?

Thanks very much,
Batman2

 

[ProbaSci]

Batman2,
Sorry, I can't recommend any other refs. Infact, I wouldn't even recommend a stats book. If there is someone in your company that understands stats, I'd approach them and try to learn as much as you can.

For me:
stats in a class = very hard and incomprehnsiable, but
stats on a project that I am working on = somehting that I can see/feel/touch and understand.

One other suggestion, you may want to implement a threshold value, below which you don't consider the numbers (ie, a minscule error in regions that are not stressed may bias your data). Not to say that predicting zero when the answer is zero isn't useful....

http://www.probasci.com

-
Implantable FEA for medical device manufacturers

 

[chyriso]

Batman2,

I'm actually doing some FEA wing analysis and I did apply some stat. on the strain gauge data that we have collected.

What you probably need for your studies is Hypothesis Two-sample t-test. If you're just after some quick reference, this one is very helpful.

http://mnemosyne.csl.psyc.memphis.edu/home/hiteh/Stats S04/stats L10 two sample t.ppt

cheers,

Yon