We recently teamed up with the data science team at Panasonic in Sparks Nevada on a quality demo that included ways control charting autocorrelated data.
Panasonic has measurements in their continuous manufacturing lines that are subject to autocorrelation (large lag 1 correlations). The patterning of these types of data are inconsistent with what is expected in standard process behavior charts and the control limits tend to be too narrow leading to many false alarms. This problem with Panasonic data has given the six sigma team a black eye -Dan Lund, Data Science Manager.
In the demo I showed charting the residuals from a time series model (residual chart) and using confidence bands on time series predictions to see evidence of special cause variation. I did not show but mentioned using Wheeler's recommendation of using sample correlation coefficient to scale the confidence limits Rbar/d2(1-r^2) (Advanced Topics in Statistical Process Control , Chapter 12).
My near term proposal:
Include an option to compute Wheeler's adjusted confidence limits in control chart builder.
Long term proposal:
Develop a platform that will generate residual run, CUSUM or EWMA charts directly.
If it helps the business case, I am confident we can get Dan Lund back on the phone to explain in detail the benefit that correctly charting autocorrelated process data brings to Panasonic.
Thanks!!
