Control Chart Pattern Detection

[jhwatts]

I am a web developer that builds applications centered around business analytics and statistics. I have worked as a engineer and scientist for about the last 15 years in various laboratories across the country. I am very familiar with laboratory measurements, associated uncertainties, and instrument process control. I have basically developed a pattern analyzer algorithm that detects particular patterns forming as a stock price changes over time. Its fairly sophisticated. I am wanting to expand it to other applications. I have been then thinking that I can apply this algorithm to SPC. The algorithm can detect patterns that are associated with particular QC failures. Control chart software currently looks at single point or even a group of points outside of control bands. My software not only would detect a point outside the control limit and groups of points outside the control limits but how the points are orientated and if the are related to a particular pattern. This will allow particular patterns to be associated with certain process that affect SPC. I've never seen it and was wondering if anybody here knew any software out there can detect specific patterns in control charts. Also can anybody give me any thoughts about if something like this would be use full or be in demand.

 

[Miner]

Not sure what type of patterns you mean, but most SPC software does have pattern check rules, such as Nelson's rules, or Western Electric rules, as well as many other less known rules.

What is different about your proposal, and how would it benefit?

 

[Steve Prevette]

There are several variations on the rules out there, even something called zone control charts. But once you pick a set of rules it is pretty easy to program. Generally though I prefer that a human look at the chart. Even though they might miss a rule now and then, at least it guarantees someone has actually looked at the chart.

Just to prove it could be done, I programmed up the rules in Excel myself, and it does work okay.

 

[jhwatts]

It is fairly easy to write a conditional statement in excel to say if points a, b, and c are above a ucl then do something, however this algorithm will detect shapes.
Let?s say a variable measured over time begins to form an s pattern with high and low points slowly expanding over time above or below a control limit, it will detect it. What if this pattern can be linked to a particular process failure, and be detected ahead of time? It?s also adaptive and begin to learn what patterns show up the most over time.