0
$\begingroup$

I'm working with degradation data and are trying do use change detection methods to detect repairs. Since I'm looking for repairs I'm only interested in positive changes. Between the repairs the data has a negative trend, differing between the periods.

I'm currently applying lowess as a first step since the data is rather noisy. The detection is done by computing the one step gradient and detecting points with a gradient higher than a set limit.

This approach is working relatively well byt I would like to compare with some other approaches. Due to the negative trend between repairs I'm not sure which methods are appropriate. Any sugestions would be much appreciated :)

An additional note is that repairs have a minimum interval which means after detection there shouldn't be any other detections within a set period. If multiple detections are made some sort of ranking would be suitable. My current idea is the largest gradient but I have a feeling that the performance won't be what I'm looking for.

Thank you in advance. Cheers!

Improve this question
$\endgroup$
1

1 Answer 1

Reset to default
0
$\begingroup$

Since you expect a trend in you data, Holt-Winters could be a good candidate. There also is an extension that deals with cyclic data (seasonality), triple exponential smoothing.

These models are generalizations of simple, weighted and exponential moving average. They help to filter noisy data and to forecast. For more information see: http://en.wikipedia.org/wiki/Exponential_smoothing

Change detection can be added by checking if the forecast is positive. Time-out periods I would somehow fit in manual.

Improve this answer
$\endgroup$
1
  • $\begingroup$ I implemented a EWMA instead of the LOWESS and the detected changes are slightly closer to the truth. I still have some false positives and false negataives, which will probably be inevitable. To clearify what I'm trying to do, the detection of change points is just a step taken to be able to estimate the degradation trends in the periods between repairs. The trends will later be used to predict the date of coming change points. $\endgroup$
    – GustafG
    May 13, 2015 at 7:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.