Hi,
I have a customer from the food product segment trying to do an accelerated life test. There are some questions coming up during our last conversation:
* What is the advantage of using an DOE with accelerated life time design over an custom design. Especially if they are not just interested in the temperature acceleration but light change acceleration (don't think there is a model to be chosen from, right?
* How can I include Duration as a factor in an accelerated design?
There are two goals they want to achieve:
* Light test, where they test different packaging and light influences as well different settings for fat and protein for a cooled joghurt product with a response on light tasting
* non-cooled light measurement at room temperature to test the recepy stability. Factors are time, light, pressure. Response is Stability, aroma, color, duration
They take each day a test and compare it to standard restrictions (cooled).
How would you set up such two experiments to accelerate the design runs?
Hi Martin,
The accelerated life test DOE uses a Baysian approach and may give a very different DOE plan depending on the strength of prior knowledge about the model parameters. If you check the Torque Prior example used in the documentation for ALT Design then the number of units tested out of a total of 100 units is 72 units at a torque of 50, 8 units at a torque of 75 and 20 units at a torque of 100. This uneven distribution of experimental resource will result in most information for the accelerated life test model based on prior distribution. Contrast this with a custom design where we have no prior information about the life distribution model parameters, we would allocate equal number of units to be tested to each level of the factor (torque in above example), perhaps 33 units tested at torque 50, 34 untis at torque 75 and 33 units at torque 100. Custom design would be somewhat equivalent to an uninformative prior. So the choice of DOE type would depend on what they currently know about the model parameters. From your description, my guess would be your customer knows very little about the prior in which case they could keep it simple and go with custom design.
With regard to model choice. If they have no suggestion as to the non-linear model to use for modelling the effect of acceleration factors on life time, then I would try parametric model personality of fit model.
Best,
Malcolm
