Those are great URL’s for any beginner to study.
One might use some “rules of thumb” based on practical experience, as well as the more rigorous statistical methods.
For example, if the DATA is to be from a MECHANICAL process for making discrete parts, then one should first try sampling the FAMILIES of possible variation, using sample size 2 for each family, per Shainon’s recommendations. 2 sites on each of 2 parts, repeated every hour for 2 shifts perhaps, then graphed. Once it is clear which family of variation is the main problem, then SPC sampling (subgroups measured over time) can be used IF the problem is temporal. But if the problem is variation WITHIN the parts, then perhaps closer stratification of data is needed, or measuring more sites per part, or conparing that variation for all similar machines, or looking at tool wear trends of this “within-part” spread over time. Means and ranges both can drift with tool wear. Sampling for mean data involves the famous sample size of 30 (or 15) for OOC determinations. Sampling for changes in variance require much larger samples. So a wandering mean is not the same as a wandering variance. Think 1000 parts.
RULE OF THUMB for SPC chart startup is given as 15 to 30 but if the process is non-stationary (drifts, has wandering mean and unstable variance, for example) then other methods are needed. Box and Luceno’s book (Amazon.com) talks at great lenght about modern issues with process monitoring nad adjustment methods. Assessing non-subgrouped data is another issue.
RULE OF THUMB: Individual Charts are less sensitive, less powerful, (they give more false alarms for each rule added, for example) than X-bar charts.
RULE OF THUMB: Subgroup size of 2 to 4 is common and usually adequate. But for diagnostic reasons, many engineers use subgroups of 10 to 100 sites per part or parts per subgroup. ASQ had a paper recently on effect of large subgroup sizes. In general, it invalides the method used by most people to calculate control limits, as many of the subgroup sites or parts are CORRELATED and so the control limits would be wrong. What is good for diagnostics is often not good for control, given various control models. With automated gages, more data is cheap. But how you use it depends on whether the data is really INDEPENDENT and RANDOMLY SAMPLED and IDENTICALLY DISTRIBUTED. Most important is that INDEPENDENCE. And if the data is AUTO_CORRELATED, its also messy (wandering mean, showing predictability instead of randomness).
Then there is the data that comes from CHEMICAL processes, such as continuous refining. Read Svante Wold or Dr. John McGregor’s books on PCA/PLS multivariate methods for sensor-based data, which is HUGE stream of data. Rule of Thumb: Get help.
Central Limit Theorem: Only for subgrouped data!!!
Shewhart Charts: Only for stationary processes where samples are independent!!! (I am not a statistician, and those guys are still arguing about these issues. See Journal of Quality Technology, Woodall’s papers, for example.
Don’t forget; NIST online handbook. http://www.itl.nist.gov/div898/handbook/
.
.