Dirk Dusharme/Adobe
Social Sharing block
With the publication of the new VDA AIAG SPC Manual in July 2026, one of the most significant changes to statistical process control (SPC) in decades has been introduced. While much attention has been given to the new process capability indices, an equally important change is the introduction of a new approach for calculating control limits.
|
ADVERTISEMENT |
For nearly a century, Walter Shewhart’s three-sigma control limits have formed the foundation of SPC. They have proven to be remarkably effective and are still widely used in manufacturing industries today. So why does the new manual introduce an alternative approach?
The manual recommends allowing organizations to select different confidence levels when calculating control limits. Although the manual explains how these limits can be calculated, it provides relatively little guidance on why this flexibility has been introduced or how it should be applied in practice.
…
Comments
VDA Manual
Who are the people that have asked for the computation of limits to be altered to apparently alter the confidence for the detection of a signal? I wonder if they've ever used process behaviour charts to actually improve processes?
A triumph of computation over common sense.
This is defined by the…
This is defined by the comittee. Maybe people from the committee can explain where this comes from. In case you want to minimize the risk to produce products out of spec and you are not doing 100% control it might make sense to use tighter limits than 3 sigma and the OC curve gives you an idea of the risk in relation to the Ppk but you are right for normal process improvement and an acceptable Ppk of 2 or higher it would not make much sense.
Focusing on the wrong things
False alarms are of concern only when a process is operated predictably. When a process is operated unpredictably false alarms are of no concern since the signals of change will be much more abundant than the false alarms.
Likewise, the objective is to detect changes that are large enough to be of economic consequence rather than detecting any and every small shift. The speed with which we detect a shift is not so important as how we react to the shift. If you think process behavior charts are there to tell you when to adjust the process you will not get the full benefit of the charts. If you look for the assignable cause of the process change, then you will gain new insight that will allow you to operate with greater capability, less inspection, and higher quality.
The purpose of a process behavior chart is to characterize the process behavior as being either predictable or unpredictable. This classification has nothing to do with conformance of the product, although the AIAG VDA manual is clearly focused on using their version of SPC as an alternative to inspection (Process monitoring rather than process improvement). This classification into these two broad categories is relatively easy to do. Ultra precise limits are not needed.
So while the comments about the trade-offs for the different width limits are technically correct, they are irrelevant for those who wish to simply operate their processes up to their full potential and thereby increase quality, increase productivity, and increase their market share.
Process behavior charts as created by Shewhart are a simple, statistical axe. They get the job done. An axe works by brute force. A process behavior chart simply separates the probable noise from the potential signals. It does not do to try to put too fine an edge on an axe. Yet the AIAG VDA manual is all about trying to shave with an axe.
I totally agree..
I totally agree. The problem with high amounts of data nowadays is that we easily have 60000 control charts in place at the shop floor. With a semiconductor customer 60000 characteristics are monitored with control chart continuously, so we need to make sure when we address an OOC, operator and engineering time is spent as effective as possible. And almost all 60000 charts are in control. In aerospace we have products where with one CMM measurement 7000 characteristics are measured...Real time SPC is required in more and more industries and with new technology we follow more product and process characteristics with a higher frequency.
I have seen too many installations where all data is collected, charted and given to the operators to apply SPC. That will always fail so therefore I think it is mandatory to provide solutions to use the techniques in the appropriate way and not get lost in thousands of signals (Correct or false alarms)
You can argue that this is not where behavior charts are intended for and I agree but it is the most powerful tool I know to identify if all processes are (still) in control and which process should get our attention but we need to assist the operator how to prioritize and not drop an impossible task on his plate. But like always thanks for your valid comments
It's not a problem with high…
It's not a problem with high amounts of data, it's a problem of thinking about what, in the opinion of a knowledgable person, are the key characteristics that need to be demonstrated to enable the process to be efficient.
Just because you provide intelligence to enable this to happen with 1000's of points doesn't mean that it's adding value to controlling and minimising the variability of the part or assembly.
There will never be a substitute for knowledge, this is increasingly getting lost over time.
AIAG always does a good job, and this is easy to implement
I see AIAG as the best place to go for authoritative, well thought out references for quality. The AIAG/VDA FMEA manual, for example, does away with a lot of the problems (risk priority numbers--Don Wheeler's article on RPNs is worth reading--and probability-based occurrence ratings that cannot be calculated in practice) that make people not want to do FMEAs. IATF 16949:2016 offers a lot of material that is not in ISO 9001:2015, and most is relevant to all manufacturing and not just automotive.
It's always been possible, and relatively easy, to set probability-based control limits, including for non-normal distributions. The previous AIAG SPC manual showed how to do this. You just calculate the 99.73 and 0.27 percentiles of the actual distribution, and StatGraphics will do this. There are Excel functions that will do it as well. While it is true that traditional 3 sigma limits work for sample averages due to the central limit theorem, you have to fit the non-normal distribution anyway to get the Ppk. If you use the normal distribution, the estimate of the nonconforming fraction can be off by orders of magnitude.
The chi square distribution (available in Excel) can set probability based limits for the s chart. I meanwhile created a Visual Basic integration routine to set them for the R chart, and I think this can even be used for non-normal distributions.
If your Ppk is really good, and you don't want to respond to every out of control signal that comes along, you can set a center band rather than a center line that shows where the process mean is allowed, i.e. small shifts don't require action. The control limits are 3 standard deviations from the upper and lower parts of the band.
Meanwhile in real life…
Here we go again. The people who focus on the false alarm rate are concerned about the absolute wrong thing. Most probably never managed an effective control chart program. Theoretical statistics do not represent real life. In theory they do, but in practice they don’t. The real concern is the ‘probability’ of detecting a real change when it actually occurs. This condition occurs far more often than the probabilistic ‘false alarm’ as most processes rarely stay in a stable condition. But let’s address that last. First let’s address the two common theoretical myths stated above.
Control charts are not a theoretical exercise. All of the work on precise probabilities ignores the difference between a theoretically perfect distribution and a real world messy process stream. I’ve led control chart programs for decades. In my last position, I led a data analysis and problem solving team that managed tens of thousands of control charts (yes we counted them) from incoming inspection and manufacturing, to field failures and instrument function.
The first myth is that operators respond to out of control conditions. It is not the volume of data that is our problem, it is that we do not live in the simple and rather manual manufacturing and product environment of early Quality practitioners. Having worked in various industries I can say that in my experience with today’s high tech manufacturing and products, operators rarely have the ability to make changes let alone determine causes. This responsibility typically lies with the engineers. After solving and managing the solution of hundreds of ‘assignable’ causes, the complexity makes ‘assignable’ causes more difficult than ‘common’ causes to determine and fix. Operators can and do respond to systemic patterns such as tool wear; a known cause and pattern with a known fix. But these are relatively rare and ‘false alarms’ only occur when the process engineers have failed to provide process control guidance and the operators are left on their own to ‘tamper’ with the process.
The second myth is that there are a bunch of false alarms. False alarms are not the result of today’s large volume of data. In my experience, true ‘false alarms’ were so rare I only remember a handful. One was due to an isolated distribution and use of a single lot of a seasonal biological test device in a geographic location where the disease was endemic. This was not a false probabilistic alarm, but a true assignable cause not related to the quality of the product. Two others were due to incorrect control limit settings by groups that didn’t understand control charts. I had more real OOC conditions that various managers tried to wish were false alarms, delaying response. If you are having a lot of false alarms the first thing is to investigate why they are happening - probably due to sampling frequency, irrational subgrouping, incorrect limits (2 sigma), etc.
The major reason that false alarms are rarer than the probability calculations is that the driving cause for OOC conditions is actual OOC conditions. Remember that false alarms only occur when the process is stable. But assignable causes are rampant. Processes simply don’t stay stable for long. The only time I’ve really seen a multitude of false alarms is when the control limits or subgrouping are incorrectly set. 2-sigma limits, piece to piece subgrouping in the presence in other dominant components of variation such as lot to lot, etc. The fault my friend, lies not in Shewhart’s calculations but in our own lack of understanding.
Faulty idea
The number of faulty ideas about what Dr. Shewhart achieved is growing exponentially since his time. We have 100 years worth. This is a great example.
The entire argument rests on this sentence:
“Traditional three-sigma control charts have a confidence level of approximately 99.73%…” This is false. Shewhart’s limits are not confidence intervals. They are empirical process behavior limits. The moment you treat control limits as confidence levels, you’ve left SPC and entered enumerative statistics — the exact thing Shewhart warned against. Everything downstream of this mistake is built on sand. Listen to Dr. Wheeler on this one.
Allen
Add new comment