Has anyone have experience in anomaly detection in manufacturing process data?

Question

Has anyone have experience in anomaly detection in manufacturing process data?
My customer in chemical company has a sensor data of manufacturing process (ex. temperature, pressure data in every 10 sec), and he want to identify abnormal operation.
His data set doesn't have label(Y), so supervised learning (like logistic regression or boosting tree) is not applicable.
I think multivariate control chart or Mahalanobis distances is good in this situation.
They want to get graph like in page 35.
https://www.yammer.com/sas.com/#/files/162858856 
Is there any idea?

Phil_Kay · Accepted Answer

I would think that multivariate statistics like Mahalanobis distance should be useful for this. You could also consider PCA scores.

However, if this is a batch process, Functional Data Explorer could be a better way to detect anomalous batch profiles. You could use the FPC scores as the statistic in this case.

This is an old Discovery presentation from a customer in chemicals that is quite relevant. Although it does talk mainly about supervised learning methods.

MikeD_Anderson · Accepted Answer

Adding a second to Phil's comments about FDE.  I've been tinkering with some data from my grad school days and FDE give's me some interesting results (see image).  The data in this analysis are spectra.  (You could use traces or other types of data like what is shown in the presentation Phil references.)  What I found striking was that the score plot correctly discriminated between the different phases of my reaction.  So, it might be a fruitful avenue of inquiry.