Gerrit Pathuis

Important changes to forums and questions

All forums and questions are now archived. To start a new conversation or read the latest updates go to forums.mbed.com.

10 years, 7 months ago.

Loess smoothing (loess local regression) for measured data.

Hi there,

I have a ADXL213 sensor measuring vibrations of a unbalanced electric motor running at 800 rpm. Unfortenately the measured data is not very smooth.

Yes i tried different capacitors at the sensor, all according to the manual this ,did not make a lot of difference. This sensor is great for measuring measuring tilt not so great for fast vibrations.

Now i am planning to use Loess regression to smooth the results.

The question is did anyone try this and is there code available?

Thanks,

Gerrit

I cannot help you on LOESS. I had to look at Wikipedia what that is./

But I lookedd at your sensor specs. It can measure up to 60Hz. At 800rpm you have 13Hz, so that is ok. But it has a full scale range of +/- 1.2g. According to my quick calculation this allows a maximum unbalance (amplitude) of only 0.17mm at main frequency of 13Hz. If your sensor feels more than this 0.17mm amplitude... that may ruin your results in my opinion.

posted by Geert Hospers 25 Jan 2014

I think I made a factor 10 mistake in my quick calculation: maximum amplitude can be 1.7mm.

posted by Geert Hospers 26 Jan 2014

2 Answers

Erik -
10 years, 7 months ago.

I also had to look up what LOESS is, but I don't think it applies to your situation. If I understand it correctly it is an alternative to for example least square fitting. But that is for when you have to fit a function to your measurement data, I don't think that is what you want here. Even if it is, something like least square method might be easier and faster to implement.

But for smoothing measured data you generally just use a filter, which can be digital. I would start by saving a bunch of measurements, and if you have something like matlab available to put it there. Then start by looking at in matlab, try some filters, see what the FFT looks like, and when you have a good solution try to get it in your mbed.

Accepted Answer
Gerrit Pathuis
poster
10 years, 7 months ago.

The ADXL213 did proof not to be the problem, in fact the numbers i see now are very smooth. The digital damping formula i use now is something like Result_new = result_old x 0.95 + last_sample x 0,05. The problem was in the software that presents the results.