Sep 26 2020
More about Toyota and SPC
The post on Does Toyota Use SPC? elicited many comments on LinkedIn. Some suggested that it was scoping SPC too narrowly when contrasting it with Toyota’s approach. In fact, SPC as referenced in the post is the body of knowledge described in the American literature on quality and taught in professional courses.
As to why Toyota is not using SPC, the answer is simple: SPC is about process capability and the quality problems Toyota addresses in 2020 are not due to lack of process capability. In industries that lack process capability, modern data science outguns the old SPC toolkit but that is a different discussion. The most vital question raised in the comments was why we have been not learning Toyota’s approach to quality. In the past 30 years, American industry has learned “Lean Six Sigma” instead.
The comments also enriched the public sources of information cited in the post with corroboration by current and former employees of Toyota.
Contents
Scope of SPC
Inside an organization, members can use any label that suits their purpose to describe their activities. In an open forum, on the other hand, participants must agree on a common meaning for any discussion to be possible. Otherwise, it’s the tower of Babel and they all talk past each other.
In the Quality Literature
If you search Google Images for “SPC,” you see Control Charts. In the literature, Control Charts make up the bulk of SPC:
- In Douglas Montgomery’s Introduction to Statistical Quality Control, Part II, about SPC, covers only Control Charts, in 350 pages. DOE is a separate topic and the subject of Part III.
- In Juran’s Quality Control Handbook, 5th Edition, Section 45 about SPC covers exclusively Control Charts.
- Tom Pyzdek, in The Handbook for Quality Management, covers this ground in Chapter 9, Quantifying Process Variation using 55 pages but does not call it SPC.
A process capability study is Step 1 in setting up Control Charts, and the C_{p} and C_{p_k} are ratios based on control limits and tolerances. It’s all about the same model of process behavior. I see other items under the same heading but you are welcome to view them differently.
If a manufacturer doesn’t use this specific body of knowledge, for the purpose of this discussion, it’s not using SPC. It may well be using data science to develop and stabilize processes. At a sufficiently high level of abstraction, an abacus and a computer are both digital calculation devices but you don’t see abacus training offered as part of Industry 4.0.
SPC as Taught
SPC is taught as a specific set of tools. Students pay to get trained to use them and receive a belt as certification from various organizations. Sign up for a class called “SPC” and you are not going to learn in general about eliminating variation. You are going to be indoctrinated into specific and ancient statistical techniques, including Shewhart charts, Western Electric rules, Cusum charts, and XmR charts. The syllabus may even include scatterplots and histograms, at the level taught in Middle School.
You will not learn any of the tools and techniques a company like Google requires from data scientists, some of which are relevant to process control.
The American Society for Quality (ASQ) has an extensive course catalog that includes five entries with “SPC” in the title. SPC Basics is all about Control Charts. Advanced SPC adds the Normal Distribution, and Design Of Experiments (DOE).
The Union of Japanese Scientists and Engineers (JUSE), known locally as Nikkagiren (日科技連) is the organization that hosted Deming and Juran in the 1950s and awards the Deming Prize. In Japanese, they offer many courses on quality, but none focused on SPC. They have separate courses on TQM, QC-circles, Problem-Solving and Statistical Quality Control (SQC). Under SQC, SPC appears part of their course on the 7 tools of QC. Their other SQC courses cover testing and estimation, correlation and regression, design of experiments, multivariate analysis, and Taguchi methods.
Quality versus SPC
One participant in the discussion, Tony Burns, cites a post from his own blog that opens with “The control chart is at the heart of the very definition of quality.” With all due respect, the control chart is just a tool. Manufacturers have pursued quality by other means since before Shewhart invented it and today as well.
We can summarize Juran’s perspective by saying that quality is the agreement of reality with expectancy. This koan-like statement can be unpacked in different ways, as expectancy can be interpreted both as customer expectations and as expected value from probability theory. Juran sets a goal but does not mandate any particular tool to reach it.
If not SPC, then what?
If you lack process capability, you need to get it. You may achieve it just based on your engineers’ experience and knowledge of physics and chemistry. This is the way it was done pre-SPC, and many companies still do it, including successful ones. The promise of data science is that you will do it faster if you know how to collect and analyze data, and communicate about the results, even if you take into consideration the learning required.
Changes since 1950
In 1950, SPC would have provided the best available tools for this. Today, data science does, encompassing data acquisition from sensors, data storage and retrieval through a variety of databases, analysis with a variety of tools, and communication via infographics and other methods…
Shewhart et al. developed SPC around the limitations of the information technology of the 1920s, which no longer apply. The boosters of old-school SPC like to point out failures of machine control systems to eliminate the need to take measurements and tally attributes on workpieces after they complete each operation.
Of course, you can shoot yourself in the foot with fancy toys. What matters is what you can do when using them wisely. Unfortunately, all the time you spend learning old methods is no longer available to become effective with new ones.
Compared with the pre-WWW-II era technology that SPC was developed for, a few things have changed. Sensors, with data acquisition systems give you more and better data than manual collection, including photographs and videos. Cyber-physical systems, for example in additive manufacturing of metal parts, run a process simulation along with the actual process and respond when the two diverge. In semiconductors, Run2Run controls tweak the next operation based on the present one.
SPC versus TWI
It’s not just a matter of vintage. Other concepts from the same era as SPC, like TWI, have better stood the passing of time. TWI is about human behavior and cognition, which have not changed. You can use a smartphone app in lieu of 3×5 cards, and instructional videos but the concept remains the same.
Quality Strategy and Why Toyota doesn’t need SPC
Shingo’s rejection of all statistical methods, together with Mikel Harry’s embrace of them got me thinking about why they held such diametrically opposed views, which they both expressed as context-free absolutes. In reality, they worked in different contexts, Shingo’s in mechanical industries and Harry’s in electronics. Once you clarify this, they can both be right.
Toyota doesn’t need these techniques because the automotive industry is so mature that its quality problems are not due to process capability but to discrete events like machine failures and to human errors, like wrong part picks. One-piece flow lines let you promptly detect and respond to the discrete events and poka-yoke/mistake-proofing addresses human errors. This is explained in the first article in the “For the Curious” section of the blog post.
Honda is similar. As backed up by initial quality surveys, they have the highest quality engines in the industry. Doing some training at their engine plant in Ohio, I was surprised to find out that they had not a single engineer trained in DOE or any other kind of statistical method. Honda relied exclusively on process knowledge.
Sources of Information
I have known about this for decades, from personal communications I can’t cite. In the post, I am citing publicly available sources so that it is clear I am not asking anyone to take my word. Since the post came out, current and former Toyota members have corroborated its main point through their comments.
Public Sources
About the origins of this information that you find amusing, the internet and YouTube are just channels. Through these channels, I am citing the following sources:
- A Toyota web page about the history of the company.
- Art Smalley, an ex-Toyota employee turned consultant whom I have met and respect
- A video from a Toyota forklift plant in the UK. What you see in the video involves no sampling and no statistics.
Toyota People
Multiple current employees and alumni of Toyota confirmed the main point:
- David Fleming “My experience is that Toyota does not use SPC in their processes. They have a long and successful history of understanding product design and equipment capability in research and development.”
- Art Smalley “By the time I worked for Toyota in the latter part of the 1980’s these manual charts were essentially gone.”
- David Martin. “TOYOTA does NOT specifically use SPC to assure quality…”
- Don Kersey.
Contradiction
Michael Bremer reported that a friend of his, a plant manager at a Toyota subsidiary, used SPC as of 5 years ago. It would be great to have this friend explain what they were doing. From Michael Bremer’s account, it’s hard to tell whether they were (1) plotting parameters over time, or (2) taking samples from batches, computing stats from samples drawn from the batches, and checking these stats against control limits set in a process capability study.
Plotting time series of critical parameters is generally useful and (almost) everyone does it. If they did (1), they responded to what they saw in these plots, and it may have served them well. Unless they did (2), however, it wasn’t “SPC” as in this discussion, which is what you learn when you take a class from an organization like the ASQ.
Learning about Quality from Toyota
Robert Nugent asked: “If Toyota does not use SPC and yet still meets or exceeds its competitors quality why aren’t we learning from them?”
Good question! I have been trying my best to learn from them and to pass on what I learned. As a quality improvement strategy, it is obvious to me that you should use statistics/data science to make processes capable, then move on to using one-piece flow for rapid problem detection and response, and, finally, to Poka-Yoke to eliminate human error. The data science gets you from 30% defectives to 3%, one-piece flow to 0.3%, and Poka-Yoke to 1 ppm.
I presented this concept at the SME Lean Management Solutions Conference in 2001 and taught it with Kevin Hop in our Lean Quality course but I have never seen it in the literature of the quality profession. The presentation was on September 11, a day we remember for other events. Since then, “Lean Six Sigma” has drowned out our message.
##toyota, #spc, #tps, #asq, #juse
September 26, 2020 @ 10:47 am
Very instructive! Learned a number of new ideas! Thanks!
September 26, 2020 @ 1:11 pm
When I’ve visited the Toyota plant in Texas, they tell visitors that they teach all employees the “7 basic quality tools” which includes control charts. It doesn’t seem like you can say definitely they Toyota doesn’t use SPC.
I’ve had former Toyota people tell me about how they use control charts for the classic manufacturing purposes of processes where the products have measurable dimensions and specifications. Are there a ton of control charts visible everywhere when you’re driven through the factory for the tour? No. I didn’t see metrics in a “Process Behavior Chart” format (one typically sees line charts / run charts without calculated control limits).
But to say that Toyota is so advanced that SPC is no longer needed doesn’t seem completely true.
September 26, 2020 @ 2:17 pm
I learned SPC in the early 1980s. What has puzzled me since is never seeing it used to control processes as specified in the literature. The closest I have seen was a factory in Germany where they were actually tweaking machines based on control charts. They were plotting Xbar-R charts on samples of 5 units picked from batches of 5,000. Where they didn’t quite get it right was that their idea of “random sampling” was for the operator to pick “here and there.”
Otherwise, I have seen many control charts on walls, placed there exclusively for the benefit of outside auditors. You can tell by their locations and by the fact that they are pristine, because any chart that is actually used is next to its object and manually annotated.
September 27, 2020 @ 1:42 am
My sensei Shunji Yagyu taught: statistical methods are only useful when you have monocausal symptoms which occure repeatedly. The better you master a process or machine, problems become something very rare and multicausal.
That means the same phenomenon on the surface could have been caused by a combination of different events. Or the same deep, deep root cause can result in different phenomenons.
When you are on this stage of maturity in „mastering your stuff“, statistics, which only tell you about what happened in the past, contain too little information about the actual interaction of events, because you can interprete and analyze them only with your bias, which you’ve earned on the base of past experiences.
If you want to solve the problems of this type you should be able to go to the Gemba immediately, and try to grasp the problem „in flagranti“ to solve them = you have to find out, why you couldn’t secure the ryohinjoken (conditions for good work).
Because in this stage, most of the events are caused by „unthinkable events“. And the only people who have a key are the well trained associates, who can tell you more about the case as any statistics can. They are the best witnesses and problem solver.
September 27, 2020 @ 6:16 am
Shunji Yagyu’s articles in Kojo Kanri about chaku-chaku lines were my best source of information on this topic and I acknowledged his contribution in Chapter 8 of Working with Machines. While I have much respect for him, I don’t quite follow the argument you outline.
I see the reason statistical methods lose their effectiveness when processes mature in the ratio of true to false alarms. A system that produces 100 true alarms for every false alarm is effective. If, on the other hand, you have a perfect process, SPC will still occasionally produce alarms, and they will all be false. If you are conscientious, the system will do nothing for you most of the time and send you on a wild goose chase two to three times a year.
In machine shops, you often see parts dumped by machines into a heap in a wire basket, which scrambles the process sequence. Sometimes, this scrambling is done by an operation in the process itself, like tumble-deburr. This also scrambles the message about process conditions that is in the sequence about, for example, the effect to tool wear. Once you have done that, you really don’t have much of an alternative to drawing a sample of parts from the heap as a belated source of data on the process.
It is the context in which control charts were introduced. If you can directly observe the sequence of parts as they come out of the machine or, even better, when they are still in it, you have information that is richer and more timely.
Machining processes are visible. You can see the tool touching the workpiece, perhaps in a haze of coolant droplets but you can see it, hear it, and smell it. You will get the best information on the shop floor, and insights from the machinists in charge.
There are, however, deviations from normal process conditions that are too small to see — like a 5-sec automatic cycle that suddenly takes 5.1 sec — and there are fabrication processes that you just can’t see because they occur in opaque chambers and change the internal structure of workpieces in ways not visible on its surface — like ion implantation into a silicon wafer.
In those cases, you have no choice but to rely on data from instruments and sensors about each workpiece. Making sense of a high volume of multidimensional data as it arrives is beyond the scope of traditional SPC.
September 27, 2020 @ 4:23 am
Hi Michel – You’re right. Who needs SPC when you have one-piece flow, 100% go/no-go gages for critical dimensions, and poka-yoke? These are the basic elements of “built in quality” that eludes almost everyone, and which consequently reveals how little shop floor experience most people have. Lean has become much more social than technical in its orientation, and knowledge accumulation (vs. hands-on experience) is the measure of one’s (highly valued) social status.
September 27, 2020 @ 8:58 pm
As an alumni of Toyota, I agree with the statement that Toyota does not use SPC. It is simply because of assurance of process capable individual machines, single piece flow and poka yoke( both human based and machine based).
The associates do follow Jidoka religiously and stop the line on finding any abnormality. Such empowerment of associates is the key for Toyota’s success in quality.
The plant where I worked as Quality Head (later on moved to head total manufacturing) we used to have more than 12000 quality check points in the car before it gets ready for shipping. All these points are assured by people starting from press shop right upto despatch. It is all about strict implementation of Jidoka that is the key for world class quality.
Understanding Jidoka and implementation of the same is a matter of culture. That is why no company other than Toyota could succeed eventhough volumes have been written about Toyota way and it is an open book. When i used to give closing remarks and answer Questions from the visitors to iur plant , one question was usually aasked to me ‘ You show everything so openly; are you not afraid of copying?’.
I used to say that we at Toyota have no fear as it is not what the material flow or the way we make cars ithat matters but that is the culture which each and every employee breathes every moment that matters ; that is really very hard to copy. So we at Toyota do not bother about visitors from our own competitors.
September 28, 2020 @ 12:10 pm
Thank you for this valuable information. I am digressing here from the topic of this thread, but I was struck in your comment when you said that you moved from head of Quality to head of Total Manufacturing. I don’t believe this transition is common in US manufacturing plants, where people who join the Quality Department stay there, and possibly move on to the same department in another company.
That Toyota makes this kind of move is a statement of their values.
September 28, 2020 @ 5:54 am
Thanks, Michel, for raising the discussion. To be clear, I have never been in a Toyota plant. (I’d like to nonetheless.)
First, I have concern with your sentence “SPC is taught as a specific set of tools.”. This may well be how it routinely is – e.g. the mistaken notion that the data need to be normally distributed for a control chart to work – but it touches a fundamental problem (i.e. why SPC is unsuccessful or doesn’t last long in a workplace). Without the way of thinking, and a suitable organisational framework, the tools can’t really achieve anything, certainly not sustainably. The tools instead risk becoming some enumerative procedure in the hands of one or two specialists or some graph that nobody pays attention to. A common problem, and I’m sure you’ve seen this, is that rather than learning from points that cross a control limit people more than likely react only to points that cross spec limits…
I work for a non-auto manufacturer and I see SPC as still highly relevant in this “digital age”. (I know your scope is SPC and Toyota.) In this context, and especially now with us drowning in data everywhere, I find more and more that control charts can often be MORE complicated than what is needed. We seem to have forgotten the basics in our rush to learn digital things. Time series plots can be so useful especially when they reflect the context of the data (e.g. line breaks or different colours to reflect different days of production, different shifts, different process orders, different production lines, different suppliers/batches of a key raw material …). Time series plots can facilitate process insight which is a first step to improvement. These graphs are most useful (or only useful?) if the way of thinking behind SPC is in place: As Shewhart said in 1931 his “control theory” was a way of getting the most out of a process (exact quote below*).
Might Toyota have limited use of SPC in assembly? Perhaps. I’d be disappointed if Toyota used SPC tools when they didn’t create some value. (You have respondents to your posts suggesting that your original comment “I stated as a fact that SPC was not part of the Toyota Production System (TPS)” is unjust.)
What would Shewhart have advocated if he were alive today? I think to get the most out of processes in the most efficient/effective way possible (so any change versus 100 years ago?). Would he have leveraged modern capabilities if he were alive today? Surely since he was a pioneer. But, the way of thinking will never die. Nor will the fact that special cause variation represents an opportunity to discover, learn and improve.
Lastly, if I use a time series plot because it is simpler** than a control chart, does that mean I am not using SPC? (Ironically, here I go backwards in time, so pre-1920s, to enable discovery and learning as a precursor to improvement.)
I’m sure you’ve already guessed I’m a big fan of SPC. For the me the way of thinking is primary, the tools secondary.
page 25: Economic Control of Quality of Manufactured Product: This state of control appears to be, in general, a kind of limit to which we may expect to go economically in finding and removing causes of variability without changing a major portion of the manufacturing process as, for example, would be involved in the substitution of new materials or designs
** One example: With high-frequency data, e.g. values every second or two as we see so often these days, a time series plot removes the need to debate and define a sampling frequency which is fundamental to effective use of a control chart. This point is not trivial.
September 28, 2020 @ 8:39 am
I just wanted to be clear about what I meant by SPC as a body of knowledge. What Shewhart and his cohorts designed in actually too subtle for today’s manufacturing professionals, and intended to work around the limitations of information technology in the 1920s.
The role of the “normal” distribution is a case in point. All the parameters used to set control limits are based on it but many SPC gurus insist that it doesn’t matter.
When points cross control limits, operators should react; when they cross spec limits, they must react. You can ignore an early warning but not an out-of-spec condition.
Everybody plots time series, even USA Today. Every patient chart in every hospital contains time series plots of multiple vital signs. They are effective at revealing major changes, like a sales surge in toys before Christmas, fever breaking out in a patient, or a cutting tool breaking. Statistical models are intended to detect the less obvious. John Tukey touted power spectra as pointing out a 1mm change in the height of ocean waves. I don’t consider plotting time series to be doing SPC. It’s what everybody has been doing since before SPC.
I used to be a fan of SPC, and even gave a talk at the 1985 Quality Congress in Baltimore. But I was surprised to hear young engineers describe how they were introducing it at their companies and struggling with the culture. At that time, it had been around for 60 years already and I told myself that, if it had failed to take hold by then, it probably never would.
We can’t put words in the mouths of the dead but I read Shewhart as determined to apply the most advanced data analysis techniques available in his day and I have to assume he would do the same if he were working today.
His words on probability and statistics are difficult to understand today because the way the subject has been taught for the past 50 years is based on a theory formulated after Shewhart did his work and using a different vocabulary. What we now call “control theory” is also largely a post-WWII development.
September 29, 2020 @ 6:25 am
I rather handled Product Design and Development, Supplier Development too as Head Of these functions in Toyota.
It was an amazing period in my life. Toyota is a University in Manufacturing and Leadership. Development. A manufacturing guy’s life is incomplete without a stint in Toyota.
October 11, 2020 @ 2:01 pm
Interesting thoughts Michel. I do agree that data science has great promise and results already. I am certain that over the years I have seen several photos of control charts with Japanese writing on them and being attributed to Toyota.
Maybe I misread, but I have been under the assumption Toyota used SPC as a tool over the years. What is missing to me, and it may be due to my laziness that I have not seen is a clear description of a methodology for using data science concepts as there are for SPC.
SPC is relatively easy to explain. I have not seen the same for data science. It would be interested in how one would go about ensuring process capability, which if not there, would produce one of the forms of waste in defects, without using statistics. Anyways, thanks for the article. It was thought provoking.
October 12, 2020 @ 6:10 am
Data science for manufacturing is an emerging topic, open and unfinished.
Data science is an extension of statistics, which is why university statistics departments are changing their names to it.
Statistics takes you from data to answers. Data science starts with data acquisition and ends with the presentation/visualization of results. Along the way, it encompasses data storage and retrieval, preparation, and analysis.
More about Toyota and SPC – Continuous Improvement
April 14, 2021 @ 4:42 pm
[…] post More about Toyota and SPC appeared first on Michel Baudin’s […]
Process Control and Gaussians
April 2, 2024 @ 11:32 pm
[…] More about Toyota and SPC (2020) […]