Hi

I am new to this site in fact this is my first post. Anyhow I am working with SPC. It is a new position and I have only been with the company for seven months.

In the production(at the shopfloor) we weigh pieces of meat every 30 min to determine if they are within specification. The data is collected using SPC-software.
My problem is that the data is autocorrelated, but only when I choose more than the last 75 observations as the sample size, that is when the sample size consist of more than the last 75 observations the observations are autocorrelated otherwise they are fine.
How do I interpretate this result. Is it fair to conclude that the autocorrelation in question is not relevant since we are only reacting to observations within the same day, and a day is far less than 75 observations. Or should I conclude that 75 observations is to small a sample to say anything significant about autocorrelation?

Also it seems that the larger amount of data I choose to analyze the greater the autocorrelation - for instance the last 150 observation may have autocorrelation at lag 1, while say the last 200 observations have autocorrelation at lag 1 and 2 and so on. What explains this?

In case of the latter, how do i cope with this. I would rather not lower the sampling frequency, and also I would rather not use a EWMA charge - as I prefer to keep things simple.

Yor help and input is appreciated.

Thanks

/troelsi

Hi

I am new to this site in fact this is my first post. Anyhow I am working with SPC. It is a new position and I have only been with the company for seven months.

In the production(at the shopfloor) we weigh pieces of meat every 30 min to determine if they are within specification. The data is collected using SPC-software.
My problem is that the data is autocorrelated, but only when I choose more than the last 75 observations as the sample size, that is when the sample size consist of more than the last 75 observations the observations are autocorrelated otherwise they are fine.
How do I interpretate this result. Is it fair to conclude that the autocorrelation in question is not relevant since we are only reacting to observations within the same day, and a day is far less than 75 observations. Or should I conclude that 75 observations is to small a sample to say anything significant about autocorrelation?

Also it seems that the larger amount of data I choose to analyze the greater the autocorrelation - for instance the last 150 observation may have autocorrelation at lag 1, while say the last 200 observations have autocorrelation at lag 1 and 2 and so on. What explains this?

In case of the latter, how do i cope with this. I would rather not lower the sampling frequency, and also I would rather not use a EWMA charge - as I prefer to keep things simple.

Yor help and input is appreciated.

Thanks

/troelsi

SPC by its very nature has autocorrelation. If you are using an X/movR chart on the individual weights, the range between successive readings should capture the autocorrelation in the range estimate. It is not perfect, but unless you have natural subgroups (such as days or shifts), your best estimate are the individuals.

It would be helpful if you could:

post some of your data - your Chart(s) and the raw data in Excell format so we can play with it
Expand on the description of your sampling plan with relationship to the population data: how many pieces are weighed every time period and how many pieces are produced in that tiem period, etc.

I have dealt with SPC on an autocorrelated process in the past (rubber extrusion).

First, use an I-MR chart, not an Xbar-R chart. Control limits from an Xbar-R chart are 100% measurement error.

Determine the interval for the autocorrelation, then set the sampling frequency to greater than that frequency in order to include actual process variation. This will take a little experimentation to arrive at the right frequency.

I have dealt with SPC on an autocorrelated process in the past (rubber extrusion).

First, use an I-MR chart, not an Xbar-R chart. Control limits from an Xbar-R chart are 100% measurement error.

Great point!!! :agree1:

I have dealt with SPC on an autocorrelated process in the past (rubber extrusion).

First, use an I-MR chart, not an Xbar-R chart. Control limits from an Xbar-R chart are 100% measurement error.

Determine the interval for the autocorrelation, then set the sampling frequency to greater than that frequency in order to include actual process variation. This will take a little experimentation to arrive at the right frequency.

I am currently using an individual X-chart with fixed LCL & UCL. Could you please explain in detail why I should use an I-MR chart? Would that reduce the autocorrelation?

Thanks

You cannot "reduce" autocorrelation, you can only understand that it is present and for how long it is in effect.

The Individuals (X) Moving Range (MR) chart solves this:


An Xbar-R chart's control limits are based on within subgroup variation. Since all units within the subgroup are autocorrelated, this variation is 100% measurement error. This usually results in one of two issues depending on how frequently you collect subgroups:



Points hugging the central line, or
Almost all of the points out of the control limits



An X-MR chart's control limits are based on the variation from one measurement to the next. If the measurements are spread out long enough (beyond the period of autocorrelation) actual process variation will be included in this giving better control limits.

You cannot "reduce" autocorrelation, you can only understand that it is present and for how long it is in effect.

The Individuals (X) Moving Range (MR) chart solves this:


An Xbar-R chart's control limits are based on within subgroup variation. Since all units within the subgroup are autocorrelated, this variation is 100% measurement error. This usually results in one of two issues depending on how frequently you collect subgroups:



Points hugging the central line, or
Almost all of the points out of the control limits



An X-MR chart's control limits are based on the variation from one measurement to the next. If the measurements are spread out long enough (beyond the period of autocorrelation) actual process variation will be included in this giving better control limits.


That's the chart I am currently using - an X-MR chart. The real problem is then that the measurements are not spread out long enough. According to my studies, where I have removed observations from a dataset until there is no autocorrelation left, the measurements would have to be spread out pretty much - leaving only every sixth measurement left. I would prefer not having those measurements spread out that much.
What to do...
I have been considering a moving average range chart, but I am not sure if that chart has the same problem as the X-bar chart.

Your input is appreciated.

Have you had a look at the 2 files in our post attachments list (http://elsmar.com/Forums/fileslist.php?mode=allfiles&sortby=filename&pageamt=2&criteria=autocorrelation)?

According to my studies, where I have removed observations from a dataset until there is no autocorrelation left, the measurements would have to be spread out pretty much - leaving only every sixth measurement left. I would prefer not having those measurements spread out that much.
What to do...


Again if you post your data with a more detailed explanation of the process and your current sampling plan we can be MUCH more helpful.

It will aslo be helpful if you can explain what you want to accopmlish with an SPC chart...

Also, remember that inspection plans to determine acceptance are NOT the same as those one would use for SPC. As the intent is different so must the approach be different.

Again if you post your data with a more detailed explanation of the process and your current sampling plan we can be MUCH more helpful.

It will aslo be helpful if you can explain what you want to accopmlish with an SPC chart...

Also, remember that inspection plans to determine acceptance are NOT the same as those one would use for SPC. As the intent is different so must the approach be different.

A dataset has been attached. A machine is producing meat pieces and with every stroke the machine creates 18 pieces of meat. Every 30 min a single piece of meat is weighed and at the same time all 18 pieces of meat are weighed as a total. As of now the total of the 18 pieces are used on the control chart to control the process. The goal is to keep the process in control.

The data in the excel file is raw data!

I am not sure what you mean by inspection plans to determine acceptance...

Any suggestions?

During the data set provided, was there any adjustment made at all? Is there any adjustment that can be made? Since you chart the 18 pc composite data, is there a reason why you measure the individual "cavities"? Do you expect that they can become a problem individually? Is there any change of density of the meat, such as a lot to lot change? Is there any change in weight depending on the amount of meat in the feed area, such as pulsing?

Just trying to get a handle on the components of the total variation that contributing to what you are plotting. Autocorrelation may or may not be problem - even if it exists.

For those following along, I was massaging the data to see how it was behaving. Attached is a I-MR data, and a X-Y plot of the 18 individual "cavities" over the run. May not mean much right now, except for visualization of the data.

Yes adjustments have been made, but the reason for them and the corrective action have not been recorded. it's not expected that the single pieces will become a problem individually, however, the individual pieces are different from each other. There might be a change in the density of the meat between batches. The amount of meat in the feed area -does not have an effect on the weight of the pieces.

I have made a small experiment where I have taken 30 samples with very little time interval (seconds) between each observation. The 30 samples shows no sign of autocorrelation. So it's over a longer period of time that autocorrelation becomes evident.
This suprises me a bit, since I thought the more frequent sampling the stronger autocorrelation.

Also the last 50 observations with a 30 min time interval shows no sign of autocorrelation either, but if I instead analyze the last 100 observations with a 30 min time interval then autocorrelation is evident. What can be concluded bases on this?

When modelling with the data I have found that if I only use every sixth observation for the analysis then there is no autocorrelation. Does this mean that If a sample is only taken every third hour then there will be no autocorrelation. Or is it possible that the autocorrelation also will be eliminated if the sampling frequency is lowered to say one sample per hour.

I have made a small experiment where I have taken 30 samples with very little time interval (seconds) between each observation. The 30 samples shows no sign of autocorrelation. So it's over a longer period of time that autocorrelation becomes evident.
This surprises me a bit, since I thought the more frequent sampling the stronger autocorrelation.

Also the last 50 observations with a 30 min time interval shows no sign of autocorrelation either, but if I instead analyze the last 100 observations with a 30 min time interval then autocorrelation is evident. What can be concluded bases on this?

When modeling with the data I have found that if I only use every sixth observation for the analysis then there is no autocorrelation. Does this mean that If a sample is only taken every third hour then there will be no autocorrelation. Or is it possible that the autocorrelation also will be eliminated if the sampling frequency is lowered to say one sample per hour.

I am not sure. From the earlier data, I would see one day that showed autocorrelation, and the next day did not. But, I could not tell if the reason for not having autocorrelation was because there was adjustment - which was not recorded.

Usually, autocorrelation can be seen in the data as a trend. If the trend makes sense - as in you would expect it from the process - then often it is better ignored and the data taken as is. One thing I found from the original data that was interesting, there seems to be an upward trend throughout the time period (see attached trend chart.) The trend, however, is dwarfed by the within-day variation. Also, there is some significant variation from day to day. I am not sure if what you are seeing is process variation, operator variation or "raw material" variation. Hard to say from a distance. But understanding the origin of the variation - which would require tracking operators, raw material lots (maybe even raw material variation, such as density) and adjustments to understand the variation - would be the first order of business in sorting out what is going on.

After sorting out and understanding your process variation, the next question is, does autocorrelation matter? Does it affect the ability of the charting methodology to provide the signals you need to make decisions about the process. If not, I would focus on other issues.

I need to ask: why are you concerned about possible autocorrelation? how do you think it affects you?

Bob is correct: autocorrelation typically appears as a predictable pattern: that's a good thing. it tells you something about how your process is behaving. Is the pattern one that creates out of tolerance parts? if so you now have a great clue as to how to fix it. If it's just creating an "out of control" condition but it is predictable, acceptable product and explainable - you can just ignore that particular OOC condition.

remember the goal of control charts or any other analysis is not ot create some statistical condition that is theoreticallly 'perfect' but rather it is to understand adn improve our processes...Physics precedes statistics

Thanks for the input.
I am currently investigating whether the temperature on the meat-lots primarily are causing the variation. If that is the case autocorrelation won't matter as it is explainable and part of the process.