Jump to content

Recommended Posts

Hi All,

 

Suppose I have created a capsule with a trend which looks like this.

image.thumb.png.a6dbf09bfc7ee5d6754f6cfc60870cdd.png

 

a) I would like to select a capsule in these capsules and make it my "Standard Trend".

b) Also I would like to find out find out which trends are NOT at least 90% of my reference trend. Basically trying to find out outlier amongst the capsules.

c) Using that outlier result in an table format as (% duration out of bound/ Number of capsules which are not atleast 90% of my standard trend) etc would be my next step.

Any help or direction for this path would be appreciated.

Link to comment
Share on other sites

  • Seeq Team

This is an interesting challenge. I have some questions as I ponder some approaches.

Are those capsules all the same duration? Or is there some expected time warping of the selected capsules in the result?

What is the criteria for 90%? For example in the profile search tool, there is a "similarity" value that is L2 norm (aka euclidean norm) that has been normalized by the sample count. That would be a great value to compare against.

  • Like 1
Link to comment
Share on other sites

Ben.

 

Thanks for responding.

Are those capsules all the same duration? Or is there some expected time warping of the selected capsules in the result?

Idea Situation, YES the capsules will be of same duration, but the profile search to find anomalies should be able to find out the capsules which are not +- 10% within range of the "Good Trend" (Good trend=One random capsule which is chosen as normal or standard operation).

Anomaly I am trying to detect is both in terms of difference in capsule duration, as well as, trend profile.

 

What is the criteria for 90%? For example in the profile search tool, there is a "similarity" value that is L2 norm (aka euclidean norm) that has been normalized by the sample count. That would be a great value to compare against.

I just chose this randomly, If all capsules show up as 90% similar. Then I may adjust it to 95% or 98% to find the "Bad" trend or capsule. 
I agree that "Similarity" feature is something I have been trying to use.

 

So in summary steps I am thinking:

a) Select a good capsule.

b) In profile search specify this "Good" capsule as my standard and find capsules ~90% similar profile.

c) Find out number capsules which do NOT match (~90%) profile.

d) [Sum (Capsule time of Bad capsules)/ Sum (All capsule times)]* 100 , will be % duration Bad/Not normal.

 

Again any help or guidance will be appreciated.

 

Link to comment
Share on other sites

  • Seeq Team

I'll describe a similar solution, and maybe it's close enough to meet your needs.

  1. Let's call the all capsules $on
  2. Make a selection on screen that matches the ideal capsule.
  3. create a profile search against signal using a 90% similarity (or whatever value you find works well).  (The tool in the UI doesn't allow a high fidelity match - you could do a better job by declaring the start/end of the profile search manually in formula). We'll call this $good
  4. use the composite condition tool to create a new condition $bad that is $on outside $good
  5. create a new periodic condition $periods for the intervals you want to get the percentage over. For example weeks()
  6. create a signal from condition for each $bad and $on, aggregating the total duration per week. Using a discrete interpolation is easiest because the samples from each signal will align according to $periods
  7. Create a formula that is 
         ($bad/$on).convertUnits('%')

image.png

Caveats

  • This is really depending on the L2 norm to match the good periods. It's not doing any time warping, so it's not going to match +/- 10% on the x-axis very well. 
  • It's really not going to do well with varying durations of the on capsules. The profile search is going to compute the similarity for the entire width of the good capsule, regardless of the duration the $on capsule was.
  • In fact, the profile is totally independent of the $on condition. $good could contain capsules that have nothing to do with $on.
  • You may need to convert some of the tool items to raw formula so that you can adjust their parameters more.

A completely different approach would be to use the reference profile tool to create boundaries of signal. The bad capsules could be identified based on the percentage of time the signal is outside the boundary. That would mitigate the effects of varying on durations

  • Like 1
Link to comment
Share on other sites

  • 3 months later...

@Ben Johnson: Sorry for replying little late. So This approach has been working great so far. 

However I have additional question.

If this analysis needs to be done in a different format: eg. I would like Upper and Lower Bound for the trend in Capsule to be generated NOT from the capsule I select as being good, but from the Average of the capsules, lets say last 3/4 capsules? Or capsules in past 12 hrs. 

So instead of having a static upper and lower bound, it will be a rolling average limits based on previous 3 capsules.

Any assistance in this direction would be appreciated.

Edited by SeeqKing
Link to comment
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now
×
×
  • Create New...