As mentioned by Krypa, we need to create a platform for data analysis. The goal is to bring this into the Excel template. I have two suggestions on what we can do to evaluate a period.

1.) Evaluate regression curve BPFSL

In the Excel spreadsheet attached above, the calculation of the regression curve is already included. Once we have collected some data sets, we can try to find similarities. Excel calculates the coefficient of determination R^2 and the corresponding formula.

For logarithmic regression, the formula looks like this: y = c + a* ln(x).

I could imagine that the following equation could fit y = BPFSL_Start + a * ln(x)

With multiple logs, there is now the possibility to determine the quantity “a”. So maybe a simple prediction about the course of a period would be possible.

2.) Evaluation of existing data or DoE (Design of Experiment)

I must say right at the beginning that I am not an absolute professional for DoE, I have worked with it a few times.

If you want to optimize different input variables for a perfect result, you theoretically need a full factorial experiment. This means that every combination is tested and the result is evaluated.

To do this, you need to know how many values an input variable can take. If we consider the US frequency as an example in our case, there are 2 values circulating here in the forum (1 MHz and 3 MHz).

For example, if you have 5 input variables, each of which can take 2 values, you would have to run 2^5 (=32) trials to fully factorize all 5 input variables.

If these 5 input variables can take 3 values, then it would already be 3^5 trials (=243) trials.

And now DoE comes into play! DoE means that an experimental design is performed statically. A special software proposes an experimental program that tests only certain combinations. The other missing combinations are calculated with statistical methods. The test effort is drastically reduced. This is interesting for companies to reduce test efforts.

But what is our benefit?

Well, there are two ways to deal with this.

The first would be a statistical design of experiments and everyone would get an “order” to perform their period in a certain way. I don’t think we can make that work here.

The second, and probably more feasible, would be to analyze existing data using exactly this method.

To make this clear, I have created a FAKE model with fictitious numbers. ==> see FAKE_model_collected_data.png

This model is calculated using regression where the number of cycles, frequency, power, strain duration, US duration are input variables and BPFSL is the output variable ==> FAKE_model_Regression.png

The model allows us to use the predicted response graph to calculate the expected BPFSL after the period has been finished. The result in this FAKE model for the input variables 15 cycles, 27 min strain duration, 20 min US duration, at 1 MHz and 1.6 cm^2 is 0.97 cm +/- 0.89 cm (FAKE_model_Prediction_01.png).

Now we can adjust the input variables and determine the new expected output value based on the regression. The output in this FAKE model at 18 cycles, 33 min strain duration, 25 min US duration, 1 MHz and 2.5 W/cm^2 is now 1.86 cm +/- 0.81 cm.

The more measured values we have, the smaller the confidence interval becomes. I want to emphasize that these are no real numbers, it is FAKE data to explain the approach.

Then we should convert the whole thing into reality. This is not a short-term task, but a long way until we have reliable data. The whole thing only works if we have variances between periods (duration, power, frequencies.). It is important that the training does not change during a period, otherwise the evaluation of the period is useless and distorts the data. If someone of you documents several periods, he can do each period differently, no problem, but each period must be gone through without change.

I would like to encourage discussion here in this round about what all we should consider. Just to give an example: This FAKE model does not take into account the sweet spot, it is assumed that everyone knows it and is in the range. If we think that needs to go in, then we should discuss it here. However, we must not forget that the need for data increases with each new input variable. Therefore, I would like to keep the model compact.

Furthermore, we should consider whether we want to optimize the input variables for BPFSL or for Strain. So there are still many unresolved issues.

Let the discussions begin *:)*

/.Slashdot