Manufacturing lots ranking based on their impurity profile using JMPA production plant makes 1000s of lots and analytical lab analyzes samples and generates data, concentration of impurities and other test results. Impurities need to be minimized but there are certain attributes that need to be maximized. Question is, what is the best way to rank and visualize such lots based on overall data and also by adding weight to certain variables? It's like analyzing a DOE data having all the responses but no factors. Thank you very much for your suggestions. Regards.
Re: Manufacturing lots ranking based on their impurity profile using JMPThis case is about multivariate analysis. There are many methods that you might use.
A parallel plot displays all the variables across lots. See Graph Builder information in Help > JMP Documentation Library > Essential Graphing guide.
The next level might be to look for correlations. The Multivariate platform displays a matrix of linear correlation coefficients and a matrix of scatter plots for all variables two at a time. See Multivariate information in Help > JMP Documentation Library > Multivariate Methods guide.
Principle components analysis is a way to use correlation between two or more variables and reduce the dimensions of the data. For example, you might have 100 variables but through correlations among them you might find that 5 components (linear combinations of the original variables) account for practically all the information. That represents a 20-fold reduction in complexity.
Multivariate control charts use the principle components to look for anomalies in the correlations instead of the individual variables.
There is more that can be done but see how this start might help.
Re: Manufacturing lots ranking based on their impurity profile using JMPThe only thoughts I would add to Mark's suggestions are:
1. Question the measurement systems. Any way to determine the measurement errors in the data sets? Do you have any repeated measurements?
2. Look at the data using simple time series. Since you have not provided any hypotheses as to why the impurities vary (so you would have no subgrouping strategy), you might just use a Moving Range and Individual chart to look at the data.
Re: Manufacturing lots ranking based on their impurity profile using JMPMark, Thank you. Multivariate analysis was what I was doing too. Glad to know I'm on right path. I appreciate it.
Re: Manufacturing lots ranking based on their impurity profile using JMP@ftadesse_genoma,
@markbailey and @statman made very good suggestions. You did not mention if each "lot" is like a vat of material or made up of smaller components like wafers in semiconductors, or boards in a solder line. To expand @statman's suggestion, how the sample(s) from the lots are taken and the sample-to-sample variation within a lot also should be reviewed as a time or sequence. I suggest a variability chart using Lot id and Sample id within lot for grouping on the X-Axis vs. the measured values, creating a plot for each metric of interest.
A mutivariate control chart might highlight not only an unusual result for one metric, but also a change in the relationship if there is correlation.
It is not that you have no factors, no controlled factors, in fact, it is likely you have many uncontrolled factors: equipment used, batches of incoming material, time of day, temperatures, gauge readouts, etc. all dependent upon the manufactiring line you described. That is why the control chart methods, univariate or multivariate, are suggested. They can detect anomalies, unusual results, and then collect and analyze the process history, which is similar to a method popular today's current events, "contact tracing." The goal is to answer what might have caused the anomalies. If you have JMP PRO there are several tools to help with that analysis as well. While collecting process history data, I suggest bringing-in people knowledgeable of the manufacturing line. Even in sophisticated, automated manufacturing, there always seems to be something that is not captured in the data alone.
Good luck.
