I’m building an application for Android devices that requires it to recognize, by accelerometer data, the difference between walking noise and double tapping it. I’m trying to solve this problem using Neural Networks.
At the start it went pretty well, teaching it to recognize the taps from noise such as standing up/ sitting down and walking around at a slower pace. But when it came to normal walking it never seemed to learn even though I fed it with a large proportion of noise data.
My question: Are there any serious flaws in my approach? Is the problem based on lack of data?
The network
I’ve choosen a 25 input 1 output multi-layer perceptron, which I am training with backpropagation. The input is the changes in acceleration every 20ms and output ranges from -1 (for no-tap) to 1 (for tap). I’ve tried pretty much every constallation of hidden inputs there are, but had most luck with 3 – 10.
I’m using Neuroph’s easyNeurons for the training and exporting to Java.
The data
My total training data is about 50 pieces double taps and about 3k noise. But I’ve also tried to train it with proportional amounts of noise to double taps.
The data looks like this (ranges from +10 to -10):
Sitting double taps: 
Fast walking:
So to reiterate my questions: Are there any serious flaws in my approach here? Do I need more data for it to recognize the difference between walking and double tapping? Any other tips?
Update
Ok so after much adjusting we’ve boiled the essential problem down to being able to recognize double taps while taking a brisk walk. Sitting and regular (in-house) walking we can solve pretty good.
Brisk walk
So this is some test data of me first walking then stopping, standing still, then walking and doing 5 double taps while I’m walking.
If anyone is interested in the raw data, I linked it for the latest (brisk walk) data here
Have you considered that the “fast walking” and “fast walking + double tapping” signals might be too similar to differentiate using only accelerometer data? It may simply not be possible to achieve accuracy above a certain amount.
Otherwise, neural networks are probably a good choice for your data, and it still may be possible to get better performance out of them.
This very-useful paper (http://yann.lecun.com/exdb/publis/pdf/lecun-98b.pdf) recommends that you whiten your dataset so that it has a mean of zero and unit covariance.
Also, since your problem is a classification problem, you should make sure that you are training your network using a cross-entropy criteria (http://arxiv.org/pdf/1103.0398v1.pdf ) rather than RMSE. (I have no idea whether Neuroph supports cross-entropy or not.)
Another relatively simple thing you could try, as other posters suggested, is transforming your data. Using an FFT or DCT to transform your data to the frequency domain is relatively standard for time-series classification.
You could also try training networks on different sized windows and averaging the results.
If you want to try some more difficult NN architectures, you could look at the Time-Delay-Neural-Network (just google this for the paper), which takes multiple windows into account in its structure. It should be relatively straightforward to use one of the Torch libraries (http://www.torch.ch/) to implement this, but it might be hard to export the network to an Android environment.
Finally, another method of getting better classification performance in time-series data is to consider the relationships between adjacent labels. Conditional Neural Fields (http://code.google.com/p/cnf/ – note:I have never used this code) do this by integrating neural networks into conditional random fields, and, depending on the patterns of behavior in your actual data, may do a better job.