I am building an application which compares time series at a very high volume and velocity. My approach of using Dynamic Time Warping (DTW) in combination with scaled time series (range between 0 and 1) seems to work alright. I want to detect occuring anomalies in those time series. For that i am calculating a score (range from 0% to 100%) to easily visualize the "health status/similarity score" of a time series. For calculating that "health status/similarity score" i am calculating:
where
Source: Normalized measure from dynamic time warping
In Theory that is really good, but in practice I am facing the issue of a lot of 90%+ similarities even i know that there are anomalies. Now i am trying to find a better solution for calculation MX f.E in combination with the standard deviation or something, but i wasnt able to find any solution yet. Just decreasing the size of MX to a smaller fixed value feeled not good.
