Feature extraction of accelerometer data for machine learning


Currently, I am doing a project with the aim of classifying potholes through machine learning. The data collected is from an accelerometer in which the z-axis measures the "vertical" acceleration of the car, when a pothole is struck.

I have tried to deconstruct the signals and create features using two methods:

  1. PACF along with a moving average to combat the noise

  2. Calculated a periodogram for spectral analysis

Is there any other ways to differentiate a pothole from "rough" road surfaces, as I am unsure which approach to take?


Posted 2017-08-03T10:49:30.910

Reputation: 93

Perform SSA to preprocess your signal.

– Emre – 2017-10-19T22:34:36.507

did you find any solution to processing the accelerometer signal? – Manas Gupte – 2017-10-19T20:16:26.517

hey so i see that this was 2 years ago, have you completed it? can i have some references on it....THANKS – Nikhil mahajan – 2019-10-04T05:59:27.793



Actually you could treat your acceloremeter signal like normal audio signals. There are endless possibilities for processing audio data (e.g. chroma features).

Another way would be to directly process your raw signal with the help of a neural network (1D-convolution or LSTM).

Andreas Look

Posted 2017-08-03T10:49:30.910

Reputation: 863


Instead of spectral features and moving average, I would recommend wavelet features. You could either do a continuous wavelet transform (CWT) or a Short Wavelet Transform (SWT) and identify the peaks / drop where the potholes show up. The advantage with wavelet is that it is well resistant to noise and you can also preserve the time axis information to pin point the location of peaks.

The trick is just to choose an appropriate wavelet. Those which have spike shapes like Daubechies and Symlets would be ideal.

With proper feature extraction, you can even do the detection without machine learning.

Arun Aniyan

Posted 2017-08-03T10:49:30.910

Reputation: 183

Thankyou for your help. I have applied DWT with the Daubechies filter using the wavlet package on R, unfortunately, the classification did not improve. – lfs – 2017-08-04T13:55:50.017

Curious to know what features did you extract with DWT? Because simply applying DWT and using the coefficients directly will not help. – Arun Aniyan – 2017-08-09T13:51:42.930