For the purposes of that graph on GitHub, I did something quick and dirty – I looked for the point where the Weibull survival function crossed 50%. This is not exactly correct, but, as I said in the last post, it’s reasonable, and it gave me what I wanted – a quick and dirty way to plot my results against actual values and prove that the model learned something. If you’re looking for me to lay the math out for you, the survivor function is exp(-(t/a)^b). Set it equal to .5 and solve for t and you get a(-ln(.5))^(1/b). Plug in a and b and you get your estimate.

As it turns out (see pages 3-4 of the document linked below), the correct way to estimate expected future life is to integrate the survival function from 0 to infinity. Please, please, though… understand that neither of these are a prediction that the subject will experience a failure on that particular day, it’s just our best estimate for how long it will survive.

]]>it gave me alpha, beta values but what would be the TTE?

can I feed that to Weibull function to predict the Remaining Useful Life (RLU)?

]]>The outcome of mode is 4 fields

1. Actual RUL

2. Event Indicator

3. Alpha

4. Beta

it might be a silly question but please explain me how did you evaluate the Predicted RUL from the above outcomes.

]]>Its a wonderful work with a beautiful dataset to understand the Time to a failure event, I want to know more about some of the parameters in the test_results which it generates, as per my understanding.

Test_results -> [ TTE, Event Indicator, Alpha, Beta]

where TTE -> actual value

Event Indicator -> yes or no [1/0]

Alpha -> some value

Beta -> some value

to calculate the predicted Time to failure do we need to feed it to the formula inverse CDF

1. The inverse cumulative distribution function is I(p) = alpha (-ln(P))^(1/beta)

or

2. Probability density function (PDF)

3. Cumulative density function (CDF)?

the formula for Weibull distribution function f(x), PDF, CDF, ICDF(p) is provided in http://www.real-statistics.com/other-key-distributions/weibull-distribution/

Please correct me if my assumption is wrong.

]]>As for why you might want to send it to Google Analytics and export from there… Google Analytics does a *ton* of value-adds on the data that flows in there. They geo-locate IPs for you. They tell you what ISP somebody is on. They filter out bots and spam. They aggregate referral information into channels. I could go on. Allowing Google to do all that stuff for you gives you a far more analyst-friendly data set to work with when you export it.