Identifying patterns in time series data

Question

I'm currently aiming to build out a system that identifies patterns in time series data.

The end goal is: Given N sections of at rest time series data, classify the section with the largest similarity between the N signals. The subpatterns are all identical or pseudo identical, but time shifted at various intervals.

I've looked at a variety of things including Dynamic Time Warping , Bag of Patterns searches , Kalman filters, and a couple other signal processing techniques that I can remember from college (Convolution, Fourier, Laplace).

However, all of them seem to fall short as the pattern isn't predefined so I can't use something like a match filter. I'm assuming I have to dip into something like an RNN or LSTM to identify the pattern between the signals.

I'm wondering if there are resources on this topic, or a proposed optimal solution before I switch over to modeling the RNN.

Answer 1

Your problem is slightly ill-defined. However, I am 99% confidenct that the answer is the matrix profile [a][b] If you want more help, give me a more rigorous problem definition.

[a] https://www.cs.ucr.edu/~eamonn/PID4481997_extend_Matrix%20Profile_I.pdf [b] https://www.cs.ucr.edu/~eamonn/Matrix_Profile_Tutorial_Part1.pdf