I am measuring the surface roughness of 3D printed tensile specimens, that have different raster angles and raster widths.
Calibration and proper fixture of the specimens are set up to conduct the experimentation. However, upon measuring and re-measuring the surface roughness of all 30 specimen, R squared values yield 0.91, adjusted, 0.85, but predicted, 0.47.
I measure 4 raw values across the width of the tensile specimen before utilising the average for generating the ANOVA r squared values. I have even tried using one set (instead of the average of all 4) of raw values but to not much success on the predictive front. So it's made me curious.
What are some useful techniques for instance when going about identifying potential specimen outliers, when it's difficult to know which surface roughness value is 'correct' from the raw values obtained?
I've attached a snapshot of my Fit Statistics from the ANOVA using Design Expert 13 software. I know I can reduce the model using techniques like backward elimination but, I'm trying to get high R squared values natively without doing that.
Any advice will be most appreciated!
AM surfaces are typically very rough and varied and therefore it would be expected any correlations to have a lower confidence than for conventionally manufactured samples.
What "roughness" are you measuring and how? There is an ISO standard (4287?) for profile measurement of surfaces that suggests measuring >10 times, and taking the maximum value from these measurements. You also need to be careful about the roughness parameter(s) you are using, they show different things and therefore you may want to look at alternative parameters.
Depending on the type of AM your measurement setup may need to be changed. L-PBF would need a very high fidelity measurement to capture the individual laser passes correctly (e.g. for hatch spacing) while FDM a much larger measurement range will be needed.
Be wary of chasing correlations where there may not be any. Laser melting of metals is a somewhat chaotic process with many confounding factors that may be more significant than those you are changing / measuring.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python