Predicting Life of the Product

[Samir]

Somewhere I have read that for the reliability of the product 50th percentile is a good assumption. What is the basis of this assumption. Doesn’t it mean that 50% of the parts have life less than the 50th percentile?

Normally we test the product at 80 degC for 168 hrs in Laboratory. I want to reduce the testing time by accelerating the test. So I tested product survival at higher temperatures.

I got 98% reliability at 168 hrs at 80degC. I calculated how long I should run a trial at 200 deg C to get the same reliablity and it turned out to be 113 hrs.

Based on this am I right to conclude that if the product survives for 113 hrs at 200 deg C then the same would have survived 168hrs at 80deg C?
Or do I have to take 50th percentile of time and forget about 98% survival? What happens to Confidence Intervals mainly lowerbound? Should not the prediction based on lowerbound Confidence limit rather than the Fitted value?

[Pantelis]

Per Q1: Somewhere I have read that …
Yes, the 50th percentile means that 50% are below and the other 50% above. Now, what do you mean by a good assumption? It is not assumption, it’s a definition.

Now with regards to the acceleration mentioned, how exactly are you computing the values mentioned?

[Samir]

Pantelis,
Thanks for responding .

Well we use Minitab software. I selected the distribution (Lognormal) with lowest Anderson Darling-Adj value (35.13) with co-relation co-efficient of 0.979. On the probability graph it looks OK.

The relationship that I selected is Arrhenius because temperature is the most significant factor which affects the bonding strength of the lamination and over the period of time the part delaminates.

This data is just an arbitirary data that I have selected with right, left and interval censors. I have selected 5 different temperature levels in my experiment with 4 to 10 samples in each level.

Here is the data:

Temp StartHrs EndHrs Freq Censor
80 168 * 20 C
80 216 * 3 C
80 96 168 1 F
110 168 * 4 C
110 216 * 3 C
140 144 168 1 F
140 192 * 3 C
140 168 192 2 F
170 168 192 1 F
170 168 * 3 C
170 156 156 2 C
200 120 144 1 F
200 168 * 3 C


I am in the process of collecting actual data and I wanted to chalk out before-hand how I shall be doing the analysis. While now I have got some actual data I tend to believe that the failure rate would be significantly higher with increase in temperature than what I had assumed in my above data-set.

Is this the information that you required? Pl. let me know.

[Pantelis]

Without getting into how you did the analysis, then what you are asking is:

“…I got 98% reliability at 168 hrs at 80degC. I calculated how long I should run a trial at 200 deg C to get the same reliablity and it turned out to be 113 hrs. Based on this am I right to conclude that if the product survives for 113 hrs at 200 deg C then the same would have survived 168hrs at 80deg C?”

The answer is Yes – a more precise statement would be that it would survive with a 98% probability (or 98% of the products would survive).

If you use the median (50th percentile) then your answer would be that it would survive with a 50% probability (or 50% of the products would survive).

As for confidence bounds you can compute these from your data set -- and yes they should be included in your result. In general the lower bound is what is of most interest.

P.S. I am not familiar with Minitab on this. I am familiar with ALTA (http://www.reliasoft.com/alta6/index.htm), and it does compute all items of interest.

[Cong Nguyen]

Help,

Please show me the best way to calculate MTBF, Mean life for a product when you have a set of failure data as follows:

Unit 1: failed after completing 1000 hours, was repaired and currently running OK at 500 hours.

Unit 2: currently running OK at 1500 hours. No failure

Unit 3: failed 1st time after completing 400 hours, was repaired and failed again after completing an additional 600 hours (total=1000 hours). This unit is, again, repaired and now running OK at 200 hours (total=1200 hours)

Which method should I treat the data, 'Complete', or 'Right Censored'.

I appreciate your help very much.

Sincerely,

Cong Nguyen
Email: anhvana@yahoo.com

[Pantelis]

What is the assumption for the repairs -- as-good as-new, as-good-as-old, something else...

If as good-as-new then you can use life data analysis concepts (fit a distribution) with the times-to-failure (complete) data for the fail events and suspend (right censored) the times of operation without failure (i.e. for unit 1 you have a failure at 1000 and a susp. after 500)

If as good as old you can use RG principles (NHPP model) see http://www.weibull.com/RelGrowthWeb/n_h_h_p_.htm and also http://www.reliasoft.com/RG/examples/rgex4/index.htm

Now if it is something in between ... it gets harder. You have to get into a generalized weibull process.

[AshishSeth]

I am measuring reliability of different Paint systems and recording the following data:
- Time since last painted
- Percentage (%) of the surface failed

If the threshold for Failure is 10% then is it better to use:
- degradation analysis with a critical degradation of 10%, OR
- Weibull analysis with everything over 10% classified as Failures and less than 10% as Suspensions

What's the difference between the two analysis methods ?

[Pantelis]

Degradation analysis allows you to “extrapolate” a failure time from multiple measurements. In other words you have time to 1%, 3%, 5% etc. degradation and want to find out the time to 10%. You then use “Weibull analysis” on the extrapolated data. Now if you already have sufficient data points to 10% degradation then you are better off using standard Weibull analysis.