I am looking for literature (books, articles, etc) on PV filtering (noise removal) for common control loops (flow, temperature, pressure).A few articles that I was able to find give some rules-of-thumb or the recommendation is to get a loop performance assessment package (Matrikon, Expertune, Protuner) and let the software does the analysis for you. I am looking for something that would help me to create a standard for my company that would give details on how should someone go about identifying noise, possible solutions to eliminate it and the consequences for the control loop. The point here is not to have a table that would say: this loop, filter should be between such and such but also I don't want a document that says only: use this software. Thanks!

 

I just use filters, I don't design them but a "name" comes to mind is Savitsky-Golay, the authors of an algorithm for smoothing filters. Lots of web hits on the name (try with and without the dash)

Here's a guy who's already done some research and posting:
http://lorien.ncl.ac.uk/ming/filter/filter.htm

Bud

 

you need surge supressor.

 

First, buy the book "Noise Reduction Techniques in Electronic Circuits" by Henry Ott.

http://www.amazon.com/exec/obidos/search-handle-form/ref=s_sf_b_as/103-1513309-5468652

Second, the first rule for filtering the "Present Value" feedback signal in a control loop is to not filter at a frequency below one half the controller sampling rate. For example, if the sampling rate is once per second, the feedback filter should not be set below 0.5 hertz. Single pole filters should usually be used as they minimize the phase shift.

 

From a control viewpoint, you want to filter to enough to remove the measurement noise, but not enough to add an appreciable lag to the perceived process response (which would result in worse control). Technically, this can be done with time series analysis (see the book by Box and Jenkins). There was also some good work also done by MacGregor (Box's academic son) and Harris (Box's academic grandson). Basically, they fit models that have 2 parts - the first part is the process response, and the second is the noise model. You get the filter from the noise model.

Technically, the Matrikon, Expertune, etc. are not time series analysis products, so they might be of limited help here. Matlab probably has something; there's also the Controller Model ID product on the Control Arts website www.ControlArtsInc.com that closely follows the Box formulation (disclaimer: I work for Control Arts), and is fairly easy to use.

Finally, you might want to question why you need to filter at all. A filter will only slow things down. It might be argued that an unfiltered PV results in an noisy controller output, but most processess act as first order filters, and these high frequency controller outputs get attenuated in the process anyway. Note this is true only if there is **no** derivative action. If you want to smooth things so they look nicer on a plot, then probably the best thing to do is to drag the data into a spreadsheet and filter it from there - it's aesthetics at this point, not signal processing.

 

Hugo et. al.,

Thanks for the replies and advice. This is definitely not a topic that you can get much information, I've bought several control books, downloaded several internet articles and talked to some experienced professionals and there is no consensus. I agree that no filtering is best but some times we have interactions within a plant control loops that cannot be easily isolated (a common combustion air header suppling several burners) or processes that are noisy per se (positive displacement pumps). What bothers me is the lack of a methodology on how to go about deciding the filter type and time constant size, specially for an area that has so much research done (chemical process control). One article that I found out to be very interesting is the one below (see url). It treats noise using a statistical approach rather than the regular frequency response ones. Still, changes to the overall process dynamics due to the filter cannot be neglected and depending on the filter size constant, loop stability can be impacted. Now, talking about existing software packages, it seems that Expertune does a "what-if" analysis of a given control loop and propose the best filter that would not deteriorate the overall loop performance. I wonder how the software does that or if it requires an engineer to use best judgement....

http://www.che.okstate.edu/faculty/R3 Folder/JPAC STF Manuscript I-104F.pdf