Learning signal patterns

Question

I have 3 signals, A, B and C let's say:

A = [1.5, 2.6, 3.7, ... ]
B = [4.5, 2.3, 1.7, ... ]
C = [6.5, 1.7, 5.9, ... ]

What are the methods used to detect patterns in signals like that ? An easy take would be using a number of samples like 128 samples and then train a classifier like SVM for instance, my problem is that one pattern could be done faster or more slow; imagine these A, B, and C signals as axis of an accelerometer, the user can make a movement fast and the same movement slow, I want to detect this movement as the same movement, so you can have the same movement done in 128 samples or in 64 let's say. Is there any method to convert this signals in something invariant ? I know they often use fast fourier but I don't think it's useful in this case.

Accepted Answer

Detecting user activities from accelerometers is not a trivial task and you have some loops to jump through. Luckily there are some solutions.

From the Signal Analysis point of view, it is always wise to sample at twice the frequency of the Natural Frequency you are working with, that way you do not have aliasing. When I did human activity recognition with accelerometers, I tried to set windows of time that were ate least twice the size of the longest any one will take to complete any activity. You do need exhaustive and multiple oversampling for that, but it is usually a must in these kind of systems.

To account for invariances there are various methods, I've been trying to look particularly at Neural Networks with tangent propagation, so you can train for these invariances in the model itself. (like a shift in time). In the last chapter of Bishop's book there is a good intro in methods for training neural nets for invariances.

Learning signal patterns

Subscription:

Related questions