Methods for Pattern Matching with Multivariate Time series?

General Forum
1

Hey Everyone,

I’m trying to figure out if a pattern in the way my car moves is similar to patterns in other vehicles. If I have a piece of data that is 5 seconds long and shows drifting, for example. In what way could I look at the data of a full drive cycle of a trip to see if this drifting (or something similar) happens on this trip?

I already have a few ideas for how to do this, but I’m thinking if there’s something I should look into so I don’t have to start from scratch!

Thanks for any ideas or help!Hey Everyone,

I’m trying to figure out if a pattern in the way my car moves is similar to patterns in other vehicles. If I have a piece of data that is 5 seconds long and shows drifting, for example. In what way could I look at the data of a full drive cycle of a trip to see if this drifting (or something similar) happens on this trip?

I already have a few ideas for how to do this, but I’m thinking if there’s something I should look into so I don’t have to start from scratch!

Thanks for any ideas or help!

7 Likes
2

It sounds like “causality” by Granger might be interesting to you.

In short, it’s a statistical test that checks whether adding another time series to an existing model can make it better.

6 Likes
3

Also, I have a lot of sequences—around 1000 of them—is there a way to check to see if any of them are in a trip?

I am also thinking of a way to look at these trips like pictures, where we could use the known sequences as seeds and search through the trip for matches… That might not be the exact way it works, but is there something close that people know of?

5 Likes
4

In a transformation space that already brings the curves as close together as possible, like rotation and offset, or maybe differentiation or integration, that’s pretty much the right idea. Just do that for every possible sequence and see if any of them match to within a certain threshold. Some people talk about dynamic time warping, which is similar but takes into account how the patterns are time-aligned, which might or might not be what you want.

4 Likes
5

As other people have said, you could try dynamic time bending. Could you tell me what the program is? What you said sounds interesting to me because I work in cars.

3 Likes
6

Have you thought about dynamic time warping or shapelets?

1 Like
7

Aoife Keogh

He’s the best at time series, haha. He used to write on this sub a lot more often, and I miss it.

I wish he would write a long blog post like this:

general idea

standard ways of doing statistics

M.L. methods (regressions with different base models and how to build them)

ML tools for deep learning (mlps, gcns, message passing)

what really works (he did this already, but the poll is from a few years ago. I don’t know if it has changed.

sets of data

Add-ons, such as multivariate, graphs, and so on

1 Like