Conformity of 1 batch vs other

[statistiques1992]

We suspect that one of our batch is impacted by a drying problem in the production line. We would like to verify if the residual humidity of this batch is similar to others batches (not significantly higher than their humidity level). I am thinking about 2 options:

• The first option is to use the actual control chart established by historical batches for continuous process validation. We will conclude that our batch is not impacted if the its value falls within the control limits. In that case, only 5 units will be measured for the batch (each point of the actual CC is the average value of 5 random units of a batch).
• The second option is to build a new control chart from this batch by using 30 random units for example. Based on the control limits of this chart, I can check if the Out-of-spec risk is low or not.

From your point of view, which option is better? Do you have any recommendations? Feel free to ask if you need more details.
Thanks.
S

 

[Jen Kirley]

We suspect that one of our batch is impacted by a drying problem in the production line. We would like to verify if the residual humidity of this batch is similar to others batches (not significantly higher than their humidity level). I am thinking about 2 options:

• The first option is to use the actual control chart established by historical batches for continuous process validation. We will conclude that our batch is not impacted if the its value falls within the control limits. In that case, only 5 units will be measured for the batch (each point of the actual CC is the average value of 5 random units of a batch).
• The second option is to build a new control chart from this batch by using 30 random units for example. Based on the control limits of this chart, I can check if the Out-of-spec risk is low or not.

From your point of view, which option is better? Do you have any recommendations? Feel free to ask if you need more details.
Thanks.
S

Hello statistiques1992,

I think SPC charts would be a good approach but one should be done for each line or the data points will not represent the same process: equipment and perhaps people are different, even if everything else is the same.

If you suspect a loss of control, I would not support random sampling because sampling should only be used after a process is in control. In fact, industry experts don't recommend the usual SPC methods for chemical manufacturing.

• Members of the University of Newcastle Department of Chemical and Process Engineering published the article Multivariate Statistical Process Control in Batch Process Monitoring in Science Direct. This article is long and looks thorough.
• A University of Manchester thesis paper titled Use of Multivariate Statistical Methods for Control of Chemical Batch Processes is also detailed and thorough. See page 42 for the section titled Overview of Classical Methods for Batch Process Control.
• The World Batch Forum's paper Statistical Process Monitoring of Industrial Batch Processes is shorter and easier for non-PhD candidates (like me ) to read.

I hope this helps!

 

[John Predmore]

There are two ways (I can think of) a drying problem might manifest itself. All parts that go thru the dryer are dried to a higher or lower value of residual humidity, and you would look for this to appear as a shift in the average humidity of the batch. The second problem could be some parts in the batch are dried more than other parts in the batch. The change from usual results might not be noticed in average humidity, but might cause subgroups out of limits in a control chart. If you are evaluating whether location is a factor in the dryness of parts in the dryer (assuming the parts have not yet been tumbled and randomized!) you would not choose random samples, but subgroups from top, bottom, inlet, corners, etc of the dryer vessel. Obviously, I don't know which problem is more likely to happen, but if you collect data from stratified samples, you can always combine data to calculate an overall average.

 

[statistiques1992]

Hello statistiques1992,

I think SPC charts would be a good approach but one should be done for each line or the data points will not represent the same process: equipment and perhaps people are different, even if everything else is the same.

If you suspect a loss of control, I would not support random sampling because sampling should only be used after a process is in control. In fact, industry experts don't recommend the usual SPC methods for chemical manufacturing.

I hope this helps!

Thanks for sharing the articles, I do agree with you that random sampling is not a good choice, since we suspect a loss of control in our process.
Also, I understand that it would be a better approach using a multivariate SPC, but in my case I have only 1 variable which is residual humidity.

 

[statistiques1992]

There are two ways (I can think of) a drying problem might manifest itself. All parts that go thru the dryer are dried to a higher or lower value of residual humidity, and you would look for this to appear as a shift in the average humidity of the batch. The second problem could be some parts in the batch are dried more than other parts in the batch. The change from usual results might not be noticed in average humidity, but might cause subgroups out of limits in a control chart. If you are evaluating whether location is a factor in the dryness of parts in the dryer (assuming the parts have not yet been tumbled and randomized!) you would not choose random samples, but subgroups from top, bottom, inlet, corners, etc of the dryer vessel. Obviously, I don't know which problem is more likely to happen, but if you collect data from stratified samples, you can always combine data to calculate an overall average.

Thanks John, yes I think it would be a nice idea to verify the results from different locations. In that case, how can I determine the number of units to be tested for each subgroup? Suppose that I have 3 locations in the dryer, should I do 3 control charts with n=30 or there are some specific control charts which allow to combine all the subgroups ?

 

[John Predmore]

The reason for 30 subgroups in a x-bar/s chart is to establish a stable baseline for SPC calculation of control limits. The question you asked, is this one batch within the interval we expect based on our baseline from the same process, you don't need 30 samples from this one batch to answer that question. You can answer that question with one subgroup, as long as the parts in the one sample are representative of the question you are trying to answer.

If your SPC chart is based on randomly sampled parts, or all samples taken off the top, be aware what you are doing if you compare subgroups using a radically different sampling strategy. That might be apples and oranges. Any difference you see should be confirmed comparing apples and apples.

The second question I posed, the question of variation between different locations within the dryer, I was actually thinking of a Multivari chart as the preferred multi-factor graphical tool for problem-solving. Multivari highlights patterns and contrasts of various samples, and is very flexible and intuitive, but is not as statistically rigorous as SPC. Multivari does not give a yes-no answer, but it can uncover factors and relationships that were previously unrecognized (like, perhaps, non-uniform dryness in your dryer).