Hi to all.

Need expert advice on how to go about implementing SPC effectively for Tape & Reel Process, the process of packaging ICs by placing units on carrier tape then sealed by cover tape. Currently we monitor process by doing peel-back force testing (PBFT) on a 15cm sealed sample once every 12hrs shift. Sample is loaded on a PBF Tester w/c measures the sealing force every mm as the cover is pulled from the carrier tape. Therefore for a 15cm sample, we obtain 150 individual measurements. We tried plotting data in Xbar-S chart with s/s=150 using SPC software. The control limits generated were so tight resulting to almost all datapoints out-of-control on both xbar & s charts. It's impossible to implement the controls else we always end up w/ machine downtime. Something must be wrong. Kindly enlighten us. Thx.

First, wellcome to the cove ajm723.

once every 12hrs shift... we obtain 150 individual measurements ... The control limits generated were so tight resulting to almost all datapoints out-of-control on both xbar & s charts.

IMHO Tight control limits almost allways means wrong charting, try "individuals with moving range chart" (I-MR), this may be your ticket, this kind of chart is recommended for data with too much autocorrelation (altho autocorrelation factor is recommended for such case) and continuos processes, if that don't solve your problem add autocorrelation factor to individuals and moving range chart.:cool:

Are you able to share the data so that we may take a look ourselves and offer some ideas?

Dear all,

Really appreciate your support & concern.

We do thought of X-Mr but then realized we could be missing the w/in subgroup/sample variations. We think that this is critical for this type of process (tape sealing), that every segment (or mm) of the tape should be checked for sealing strength.

Pls. see attached file for one of our machines, contains the raw data, plus the Xbar-S chart generated (1) using tables, then (2) using std. dev. for calculating control limits. You'll notice the tight control limits generated using either method for both xbar and s charts. I also included the histogram.

Thx again.

I haven't had time to look closely at the data, but I see two things.


The limits on the S-bar chart are quite different for the two methods. In particular, the mean values are not the same. I'm not sure what method you are using for the two graphs, but I would expect the mean to be the same no matter how it is done.
The process probably really is out of control. In other words, some of the variation is not simply due to random luck, but some samples really do have a different mean and/or standard deviation than the others.If no one else lools at the data soon, I may be able to look a little more over the weekend....


Tim F

When in doubt, always plot the data in time sequence. See the attached file.

Your data raises many issues:

Rational subgrouping; your subgroups consistently include trends and outliers. Your subgroups should be small enough that no trends or outliers can occur on a regular basis.
Reducing the subgroup size to 2 or 3 provides more realistic limits but shows an out of control process in both charts
This makes the short term trends very obvious, as well as other modes at ~ 35 and 45.
Have you performed an MSA on this measurement device? If not, perform an MSA to rule out the effect of the gage.

A subgroup size of 150 units is too large. Any control with that size subgroup will have very narrow limits. This is because the A2 value is a function of sample size, so basically you are expecting every lot to have almost the same mean. You can try to take the mean of every 10 readings doing a control chart on the 15 means instead of the 150 individuals. This will give a little wider limit while still being rational and capturing the within and between variability.

The trends on the data are clear, too much autocorrelation, cyclical process (2 of them, one a 150 points, the other each 16 points), the better aproach may be to get rid of trend (if part of the process- then in a mathematical way), then to use wichever tool is wanted. I-MR could do nice control limits, but as Xbar-S, without the trend consideration no usefull info could be achived.:nopity:

ajm,
As already pointed out, aside from being autocorrelated,your sub -grouping is very large.
Using the data in your x-bar-s_chart_using_Stdev worksheet, notice the ff:

Subgrouping 150
3-sigma value in your chart 1.9344
sigma value in your chart 0.6448
estimated sigma by your chart 7.897154931
Actual sigma of your data 7.8972


Note that sigma value used by most softwares in Control charts is calculated by:

Stdev/sqrt of sub-group size

using a subgroup size of 150 gives you 0.6448 as sigma and CLimits 56.2088 and 52.3400. Very small approximations. As already suggested you can use I-MR. Or plot series in using as max and min scales the natural 3-sigma limits as shown in your histogram. With any of these the autocorrelation should be evident.

Dear all,

In trying to control-chart the tape & reel process, we are actually attempting to hit 2 birds w/ 1 stone -- (1) capture the flaws on every tape we sampled, & (2) the usual monitoring of process variation over time. In fact, the semicon industry std. for peel back force testing is supposed to be 10 measurements/mm w/c translates to 1500 measurements for a 15cm sealed tape sample. We tried to compromise (actually, sacrificing better chances of capturing flaw) by reducing to just 1/mm hence we came up w/ s/s=150.
Are we taking the right strategy?

B.t.w, just want to note that process specs. is 20 to 120 g. Low seal strength may result to delamination, too high.. you end up w/ tearing. 20 to 120 is too much that we actually want to limit to w/in 40 to 80. Ironically, the xbar-s charts are giving us around 52 to 57.

Hope this gives good lead to your brilliant minds and help us come up w/ the appropriate solution/strategy to our problem.

Best rgrds.

There is another method of setting control limits called Fixed Capability. The VP for quality at my previous employer wrote the ASQ paper on that method. I will try to find it.

Dear all,

anyone else care to help???

Dear all,

anyone else care to help???

Hi Ajm723,
Perhaps, you could consider this.
I have a similar situation where the peel tester will capture 500 points of individual data for 1 strip of tape. What we did is that, we plot the average and standard deviation of the 500 points as one single point on each control chart. You will have to determine the frequency of sampling. Additionally, we also monitor the min and max instantaneous value and ensure it to be within our internal spec.
thks.
jeffrey:thanks:

Hi Jeffrey,

That's actually what we did... considered each tape sample as 1 datapoint or subgroup in the Xbar-s control chart, with each datapoint (or subgroup) consisting of 150 individual values as measured by the Peel Back Force Tester on every mm of the 15cm tape sample. The problem is, when you've collected enough datapoints and begin calculating for the control limits, the control limits generated are so tight that it's impossible to implement else you end up with out-of-control point every sampling. I attached in page 1 a file of one of our machines containing the raw data, the xbar-s chart & histogram. Note the tight control limits computed for both xbar & s charts. Thx.

Hi Ajm,
For our case, we are using the 3 sigma limits as calculated per your histogram without considering the correction factor C4 and we are also monitoring min and max instantaneous value.
thks.
jeffrey.

Hi Jeffrey,

You are a great help! ....owe you one. Just a couple more, hope you don't mind.
> Any references/bases on the use of xbar +/- 3s as control limits.
> What do you use to monitor the instantaneous min & max.

Thx. again. :agree1:

Hi Jeffrey,

That's actually what we did... considered each tape sample as 1 datapoint or subgroup in the Xbar-s control chart, with each datapoint (or subgroup) consisting of 150 individual values as measured by the Peel Back Force Tester on every mm of the 15cm tape sample. The problem is, when you've collected enough datapoints and begin calculating for the control limits, the control limits generated are so tight that it's impossible to implement else you end up with out-of-control point every sampling. I attached in page 1 a file of one of our machines containing the raw data, the xbar-s chart & histogram. Note the tight control limits computed for both xbar & s charts. Thx.

I like Jeffrey's approach. It sounds like you are interested in two basic bits of information
Does the mean for different subgroups remain consistent compared to the variations seen in the means of the other subgroups?
For this, an I-MR chart of the means of the subgroups (i.e plot the mean of the 150 points as a single point on the I-MR chart) would look for unusual events in the means, independent of the actual amount of variation within each subgroup
Does the standard deviation within one subgroup remain consistent compared to the variations seen in the standard deviations within the other subgroups?
For this, an S-chart by itself (without the X-bar chart) would be effective. An R chart would also work, but perhaps not as effectivel.An X-bar chart will tell you if the mean for different subgroups remain consistent compared to the variations seen within each subgroup. This is a subtle distinction from case 1 above, but if you expect long-term variations, then the X-bar chart will almost certianly show many OOC points.

I also like Jeffrey's idea of monitoring the max & min values - perhaps with I-MR charts for these numbers.:agree1: This will give you additional info about the most troublesome data so you can watch specifically for changes in these values.


Tim F

First, I want to thank all who reponded to our query. And of course to the managers & contributors of this forum.

Reynald, Jeffrey, Tim... you're one in your suggestion, Xbar-S is the way to go. So for Xbar chart we will use the straightforward xbar +/- 3sigma (similar to histogram) for the control limits. But how about for the S chart? How do we compute for the control limits?:confused:

But how about for the S chart? How do we compute for the control limits?:confused:

Can't remember the exact details but this will do:
1.) First create a histogram of your sigma to estimate it's distribution.
2.) Determine which distribution that would best fit your histogram (if i remember it correct a chi-square would do)
3. ) find the value that corresponds to the 0.001 and 99.999 percentile and use it as your sigma limits (note: this is for .001 limit. You may choose arbitrarily. Normally 0.00135 and 100-0.00135 percentiles are used.)

Basically the concept is to determine the point in your distribution where the probability of fonding a value greater than the upper value and less than the loower value is less than 0.001.

20 to 120 is too much that we actually want to limit to w/in 40 to 80. Ironically, the xbar-s charts are giving us around 52 to 57.
My sincere advise is to stop looking for statistical sense in calculation of these CL's, because the process's control may not benefit.
If your experience suggests that for your application 40 to 80 are the optimal limits, then set the 40-80 as the fixed X_bar control limits, same for all tape-and-reels (I guess nobody is going to set control limits for each machine individually?!:lmao: ).
Set CL's for R in the same way, basing on recently-achievable performance.
I'd suggest to use X_bar and R charts as easiest to understand by the technicians; using R chart or MIN/MAX (instead of S chart) is highly justified for this application, because any single low or high point may potentially result in either delaminating and lost devices or in tear of the tape; then it may make a better sense to look at actual ranges (or min/max values) instead of std. dev.
When R-chart is going out off control: have the technicians cleaning the taper's "irons" with some alcohol-moisturized cloth (or polishing or replacing the units); when X_bar goes out off control: have the guys playing with pressure and temperature (and with the speed, as a temporary measure, if nothing else helps).
The recipe has been proven at Vishay-Siliconics (small mos-fet's); the charts have provided great input for the technicians: the guys really got the point and the charts amazingly improved after very few days…. our statistician was not very happy though...:bigwave:
Of course: the initially set CL's were just a first iteration; the CL's have been tightened few times, as the technicians have gained the experience with setting the tapers to gather the charts around the center lines.... :cool:

Hello Q Friends,
is there a global industry standard for PBFT?
What is an acceptable specs ?

:2cents: : do not forget to include to the SPC routine the "manual" taper, which you probably have (it is usually used for repaires of tape un-sealed to replace e.g. devices with damaged leads).
When we've found the device out off control, that has been almost too late!: //

I might not be able to shed any clarity to the problem, but I can add more concerns as anyone can. Having performed adhesive testing in R&D in the past, I know that the method - not necessarily the measurement instrument capability - can provide a variety of error. Mounting of the sample to eliminate error from carrier movement is key. Is it a 90 deg or 180 deg pull test? Inspector operation variation of the testing can create variation on that type of testing. Adhesive testing is never as easy as it looks. The other question I have is what kind of adhesive system are you using? If you are using a rubber based hotmelt system applied via hose and nozzle, then you will readily get variation from pot life and from addition of fresh adhesive. Also variation of bead height and width would confound control. No mechanical control is going to cover that problem. SPC might prove you have a problem, but you have an organic mess that requires more than just an "adjustment" that it can predict for you. I know, not much help... :frust:

Hi Jeffrey,

You are a great help! ....owe you one. Just a couple more, hope you don't mind.
> Any references/bases on the use of xbar +/- 3s as control limits.
> What do you use to monitor the instantaneous min & max.

Thx. again. :agree1:

Hi ajm723,
Sorry for not getting back to you earlier as I have been down with chicken pox and was under quarantined for 2 weeks. I believed most of the wonderful folks here have already answered your queries.:applause:
thks.
jeffrey.:thanx:

Hello Q Friends,
is there a global industry standard for PBFT?
What is an acceptable specs ?

For 8mm wide carrier tapes, it shall be 10 to 100 grams.
For 12mm to 56mm carrier tapes, it is 10 to 130 grams.
You check EIA-481 for details.

First, I want to thank all who responded to our query. And of course to the managers & contributors of this forum. The help Moderators and the 'Regulars' here give is VERY much appreciated! Thanks for thanking them!

The Xbar & R chart that we use on our tape & reel process is a little different. We test 3 taped samples consecutively every machine set-up change or material lot change, record the mean values of each sample on a subgroup and then the average of these mean values is plotted on the Xbar chart. Surprisingly, there is still some variation seen on the mean values. But before we proceed, the max and min readings of the samples must be within the internal spec limits.
When we were recalculating the control limits (not including the out-of-control conditions), we noticed that it gets tighter each time to the point that it is no longer achievable. So what we did was we fixed the 3 sigma control limits to a certain value (based from previous readings).
Is there any deviation when we did this?

So what we did was we fixed the 3 sigma control limits to a certain value (based from previous readings).
Is there any deviation when we did this?

In "advanced topics on SPC" by Donald Wheeler I saw a similar procedure (to fix control limits), of course not recommended for continuous improvement.

But, did you tried autocorrelated control limits?, It may be the ticket out. The autocorrelated control limits (also explained on the same book) inflate the variation estimate by an autocorrelation factor, if I remember it: 1/(1-r2)^0.5

An autocorrelated variable looks like a bunch of trends, changing up and down (the trends of course).