No-Equations Six-Sigma for Lean Manufacturing in process plants

This tutorial highlights the use of Geometric Process Control (as opposed to Statistical Process Control), to quickly establish and fix the causes of quality variation without the need for statistics or mathematical modelling.

Can it be possible to apply Six-Sigma to your most difficult-to-understand and highly multi-variate process without seeing an equation or control chart?

In this tutorial we will gather more insights into the example process than could be achieved with a week of traditional statistical analysis. First, a very quick Six-Sigma recap. Six-Sigma is focused on reducing quality variation and improving process yield by a methodical and systematic application of statistical process tools in order to gain knowledge that leads to improvements. The Six-Sigma methodology has proven over the last twenty years or so that it is possible to achieve dramatic improvements in the cost-of-production, quality and throughput by focusing on process performance.

Measurement

In order to characterise the process, the various input variables, such as pressure, temperature etc., that affect the process need to be measured. This tutorial assumes that process data is available from a plant data acquisition system. To be able to characterise the process we are also going to need to have measures-of-quality of the product being manufactured. These will typically be the criteria on which the product specifications are based, such as weight, viscosity, strength etc. These may be collected from the process Data Acquisition system, or may be determined from testing in the lab some time after the product has been manufactured. An example of input process variables and output quality variables is shown in the context of a simplified tablet manufacturing process here.

A pre-requisite for Six-Sigma in manufacturing processes is the availability of these kinds of data for the process being improved. Having plenty of data is not usually a problem in process/manufacturing plants. Usually the problem is that we have so much data and so many variables that we don't know how to get the important information out. That is the challenge of the next stage.

Analysis: In this stage we basically want to discover why defects are generated by identifying and prioritizing the key variables that are most likely to create process variation. In the real-world of our process plants with their non-linear and multi-variate processes, finding such cause and effect relationships is usually done with highly statistical techniques such as principal components analysis, mathematical modelling from first principles, and usually lots and lots of XY and/or XYZ charts.

Improvement: Once the reasons for our out-of-spec process operation have been identified we can recommend changes to the process operation that should improve the performance of the process. These recommendations now need to be implemented and proven.

Control: Once we have proven that the changes to process operation are valid we need to continuously monitor and control the process to the new guidelines. This is traditionally done with many control-charts. We also need to be able to cope with abnormal situations. Hopefully the analysis phase has identified what an abnormal situation is.

Six-Sigma Using Geometric Process Control

With the new kinds of graph and dynamic models that GPC provides continuous process improvement is hard to be avoided. The images here show Visual Explorer (CVE), the process analysis tool of GPC, making valuable discoveries in minutes.

Initial analysis: In this example, P1-P14 are process variables and q4-q8 are quality variables; the data used is process history data captured on the 14th day of the month between 08:00 and 17:00. All of the past process operation is laid onto the parallel-co-ordinate plot in black. Yellow has then been used to highlight the operation which meets the quality specifications shown by the red ranges. Now we have identified in-spec operation i.e. "good product" we can immediately make many valuable observations.

Firstly, good product is only made between approximately 10am and 2pm; see green arrows. We know that this process was operated in shifts with changes at 6am and 2pm. Each set of Process Operators had their own unconfirmed ideas on how to operate to achieve good product. We can deduce that the time banding indicates poor quality for the 4 hours after a shift change; after the process operation had been 'tweaked'. Armed with this knowledge and the ability to prove it, the process engineer will find it much easier to get shift management improved.

Secondly, the good quality product does not fill the full magnitude of every quality range. The blue arrows show 'empty' quality ranges. This provides an opportunity for the production department to work with the marketing department and tighten up the quality specs and thus have ‘better’ quality without changing the process (if the quality is 'better', you may be able to charge more for the product!) Alternatively there may be areas that the ranges can be expanded to allow more product to be classed as good - again without changing the process. Has your plant ever looked at the sensibility of their product specifications?

Thirdly, the process is only making 12% good product shown at the base of the display. That means from all of the operation (underlying black areas) the 'good product' (yellow highlighted) accounts for only 12%. This was a shock to everyone from the MD downwards in this company, as they firmly believed they were making 50% good product. How does this happen? We've seen it in several plants. It happens because there hasn't been, until now, a way to plot one graph showing simultaneous achievement against quality specifications over a period of time. Instead Production Reports have contained perhaps three or four graphs each showing 50% achievement against one or two specifications ...and everyone from the MD downwards has wanted to believe it was the same 50% on each chart. It was probably closer to 50x50x50 or 12.5%.

Identifying Bad Operation: Now we will concentrate on the black areas of the process that are at either end of the operating ranges. Remember, the process has been operated in these areas although it has not made 'good product' whilst doing so. We will focus on one type of bad operation - black extremities - although other types exist such as black holes. Black extremities are regions of black that fall entirely at the edge of a range. Highlighted in this case by red outlines. Black extremities are one of the most obvious and easy to correct problems by simply changing operating instructions or altering process control limits.

When we tighten the ranges of the process variables down they no longer operate to produce black extremities but still meet all of the quality specifications. The benefits are immediate and the cost of implementation is virtually nil; A great ROI.

Improvements: By following the steps above, this process and the operator's process knowledge has already been improved no end. We have found out where to operate the process to get the best results. The new operating limits are shown here, highlighting the ranges in which the process should operate (red triangles). These are automatically generated using a query in CVE. These ranges encompass all of the 'good product' (yellow) but also include some 'bad product' (blue). (For those that are interested in multi-dimensional physics this is the lowest dimensionality box that fully encloses every yellow data point!). These colours are laid on top of each other from black to blue then yellow.

The question you are probably asking now is: "How much will my yield improve by with these new operating limits?"

The two percentages at the bottom tell us that the number of blue points are 39% of the dataset and the number of yellow points are 12%. So now we can operate only within these new limits outlined by blue our new yield has jumped to 12/39 = 30%. 30% from 12%; A yield improvement of 250%. That's got to be worth a black belt! You may also have noticed that the variable positions (ie the vertical axes) have changed order in the last picture. This is because the last query we did also worked out for us which variables affect the yield the most. The most important variables are now ordered from left to right.

This is incredibly valuable information as it tells us where to focus our limited resources, for the fastest business benefit. In this case the biggest contributor to bad product is the time of day. So we really need to sort out those shifts' behaviour! The next most critical parameter is P10, so we need to get our over-worked control engineers to focus on getting that particular control loop within its new limits as a priority. And even better, we can show them WHY we are asking them to work on that loop by showing them the diagram above.

Summary

To sum up this tutorial: from one display we have gained more knowledge than previously possible in hundreds of pages of analysis. On top of that we have performed the analysis very quickly. In fact it took about as long to perform the analysis as it did for you to read this explanation of it. Going back to our original definition of six sigma in manufacturing. You can see from our explanation that we have covered the Measurement and Analysis phases pretty well. The Improvement phase comes from implementing the improvement opportunities identified by our analysis. This brand new method of analysis can show us much, much more than we have described here. For example it also identifies 'Black Holes', 'Best Operating Zones', Clusters, Contours and Modes Of Operation.

Once we've done this 'static' analysis we can put the data-set into a unique Modelling Package - Process Modeller which auto-generates an interactive model that encapsulates every interaction between every variable. The resulting model allows us to drag variables up and down between their operational ranges and for the first time see the instantaneous effect this has on every other variable. It sounds corny - but it has to be seen to be believed. We can go direct from the data-set to a fully interactive model in minutes.

This model can then be used to explore the dynamic interactions between every variable, giving us more detailed insights into quality trade-offs and optimisation possibilities. The model can also be used to advise operators how to always keep the process in the 'Best Operating Zone' (the yellow zone) thus fulfilling the last stage of the Six Sigma methodology.