How to use machine learning to find pattern of similar regions in signals

Question

How to use machine learning to find pattern of similar regions in signals

Linminxiang

2021年7月1日 15:42

I have a long time series signal. This signal is usually very stable, but it will change when the sensor is stimulated, and this change is usually very short. I know this can be trained using the labeled method(like neural network ,CNN, etc), but it takes a lot of time to label, this is because my change time is very short(about 4 seconds), and the change time is not much. So, I want to generate a number of signals similar to patterns using random numbers, and then use an autoencoder(or feature extraction method) to learn features before performing detection.

What I want to ask is if there are any errors in my ideas or can anyone provide some ideas or opinions. Many thank!

The pattern will not be obvious, but like this:

My thoughts:

Topic sequential-pattern-mining time-series machine-learning

Category Data Science

David Marx · Accepted Answer · 2021年7月1日 15:42

1

David Marx answered at 2021年7月1日 15:42

Dynamic time warping (DTW) is a great way to find related chunks in different signals.

https://dynamictimewarping.github.io/

Eulenfuchswiesel · Accepted Answer · 2021年7月1日 12:45

A good package for finding 'similar' signals is the matrix profile for Python.

If I understand your question correctly, then it would spare you from creating random signals to train a classifier. Also, in the package you are able to give a window length, so assuming your signal is evenly sampled, then you can give the maximal duration and it should work. It is specifically designed for long signals and to my understanding is based on autocorrelation.

https://matrixprofile.docs.matrixprofile.org/

https://matrixprofile.docs.matrixprofile.org/examples/ECG_Annotation_Vectors.html

Nicolas Martin · Accepted Answer · 2021年7月1日 08:42

If your signal follow simple square patterns like you've displayed, why not using a simpler solution? Some smoothing to reducte noise (ex: kalman filter) + a derivative function (ex: diff function in numpy) to detect ups and downs should be enough to detect the signal patterns, including their durations.

How to use machine learning to find pattern of similar regions in signals

About