Methods of feature comparison for time series data

Hi all,

I'm interested to learn about methodologies for scoring similarity between times series data specifically for feature recognition.

In the waveforms shown the black data series with markets is used for comparison against the other series.

The difference I'm interested in occurs between 0.5 - 0.6s and also in the peak amplitude and settling after the peak.

I thought to use Euclidean or Mahalanobis Distance (or Dynamic Time Warping), to score the breakaway from the horizontal before the peak occurs but also looked at polyval comparisons.

Also considered piecewise Aggregate Approximation segmenting the region of interest and then weighted that higher than the difference in peak value (as peak value can change independently of the breakaway before the peak) to come to an overall score that biases the correlation to the features of interest.

I have lots of ideas but wanted to know if there's a best practice for how knowledgeable folks go about this type of classification problem?

All suggestions kindly appreciated.

[Arvind Kumar Shekar]

Hi, there are several complex ways but I request you to do a small expleriment by taking the gradient of each signal and using them as features. This will be used as features for classification and the sudden spike can be capture by the gradients. 2. Did you try NARX models? Kindly use ntstool box of MATLAB for a quick prototype.

[Gregory Deemer]

I am also interested in a nearly identical process of classifying best-fit, and remarkably, for a similarly shaped time series.

I'm currently investigating SAX methods for doing so. Have there been any developments in which method has been chosen?

Thank you,

Greg