Sunday, March 9, 2014


I recently asked my friends in the social media to pull any pictures that they "tagged" with my name.  I am sure that most of them think I am paranoid.  I choose to use this post to explain my reasons.

Electronic data acquisition systems (EDAS)

For a significant period of my working life I worked in the area of dimensional integrity for an expensive, consumer product.  At one time my firm was not competitive with our over-seas competitors and I was tasked with bring our plant to world-class standards.

Before I received that job assignment, my firm installed "vision systems" to electronically inspect approximately 300 dimensional attributes on every product.  Then they hired a very promising quality engineer to come in and "bring us to Jesus."  That person tried to make traditional Statistical Process Control rules work with the new vision systems.  It was a major fail.

Statistical Process Control is based on the premise that good information is rare and expensive.  However, one can make inferences about the larger population based on the measurements of a small number of wisely chosen samples.  A common scheme would be to measure the first 5 samples of a production run, then, if they met certain criteria to run 100 parts and measure number 106 and 107.  If those met certain criteria, to keep make another hundred parts and measure number 208, make another hundred.....

People who reload ammunition do this all the time.  We perform our set-up.  We might throw 5 or 10 powder charges that we toss back into the hopper.  The we weigh the charge thrown by our powder measures and make adjustments as necessary.  We make our first few rounds, measure COAL.  We might go out and fire a few over our Chrony.  Then we will make one or two hundred and then check another one or two powder charges and COAL.

One consequence of measuring only 1% of the parts was that "Out of Control" triggers were chosen to be exquisitely sensitive.   A gentleman named Walter Shewart proposed the following 8 rules:

1.) The most recent point plots outside one of the 3-sigma control limits.
2.) Two of the three most recent points plot outside and on the same side as  one of the 2-sigma control limits. 
3.) Four of the five most recent points plot outside and on the same side as one of the 1-sigma control limits.

4.) Eight out of the last eight points plot on the same side of the center line, or target value.  
5.) Six points in a row increasing or decreasing. 
6.) Fifteen points in a row within one sigma. 
7.) Fourteen points in a row alternating direction. 
8.) Eight points in a row outside one sigma.

These rules worked pretty well when one or two percent of the population was being monitored for relatively few attributes.  The purpose of the rules was to send up a flare that the process might have drifted.  Humans could then more closely monitor the process.  Keep in mind that you might be making 1500 parts if you are only checking one in every hundred (rule number 6).  Prudence demanded that the system be fused "tightly" if it were to get attention before it made 1500 non-compliant parts.

These rules failed miserably when 100% of the population was measured for 300 attributes and the system was hardwired to shut down every time a rule was violated.  Too many rules with triggers that were way too sensitive.  The system locked up.  Every time.

Just how good are vision systems?

One vision system was tasked with keeping track of the presence and location of a stud (threaded fastener) that was welded to the assembly.  The system kept tripping out the process.  We had a disabled person who would inspect the part, write down the unique assembly sequence number and "by-pass" the alarm to ship the part if the stud was present and in the correct location.  The process stopped about 40% of the time.  The disabled person never found a bad one.

After a couple of weeks, a programming hot-shot was freed up to check out our problem.  They determined that the camera was checking the threads on the fastener.  The original program used all of the precision that was available to them.  The fasteners that were twisted so the root of the thread was scanned were bounced for being out of position.  The hot-shot unfocused the camera and retaught the go/nogo and life was good after that.

So, what does this have to do with paranoia?

Our law enforcement system and our legal system have a culture (like the SPC rules) that evolved when information was rare and expensive.  The laws and checks-and-balances in our system are tuned to the historical availability of information.  There might be 4 reliable informants in a town of 4000.  One saves them for "big fish" because they are compromised if they must testify. 

Let's be realistic.  The police in my small town know who is dealing drugs.  They know but they lack the level of proof required to successfully prosecute unless the dealers do something really stupid.

The question at hand is, does the harm the drug dealers cause to society warrant burning untold manhours (SWAG, 500 manhours) and compromising an informant?  Or does it make sense to keep tabs on the devil you know and keep the informants in reserve in case something big decides to move into your corner of the universe?

And even when the system attempts to mete out justice to criminals, eye witnesses are notoriously unreliable.  A typical ploy of a defense attorney  is to ask the star eye witness, "What color was the car that the defendant was driving?"  A typical eye witness will correctly guess the answer about 30% of the time.  If they guess incorrectly, the attorney will hammer on the reliability of the eye witness.

That finely tuned system goes gravely out of balance if electronic data acquisition systems go on-line.  I think it is inevitable that they do.  I don't want my picture with multiple profiles out there to make it any easier for them.  Our laws are not ready.

No comments:

Post a Comment