0

Let's assume I have local measurements of the temperature, wind speed, air pressure, humidity and so on, in the form of time series, and thats all I know from the world. From time to time, a tornado goes over my probe.

Because a tornado is not just a random stuff, there is a pattern, that a trained eye can recognize in the time series... some changes in the temperature, wind speed etc. correlated together in some fashion, with unpredictable fluctuations around.

I'd like to do that in some automatic way to recognize intervals in the times series that would corresponds to periods when a tornado was "seen" by my detector.

Which machine learning method would be more appropriate to recognize them, and give me some corresponding "reliability coef".

Note that, because the tornado is a inherently unsteady object, which furthermore moves in some erratic way, the detector do not always see the same variations of temperature, wind speed etc. as the tornado can move back and forth over the detector, locally changes its shape etc. I guess what I want to say is that the time series measurements do not correspond to the actual spatial profiles of these quantities one could plot in the "rest frame" of the tornado. However, it always see "kind of" the same features with some randomness around that my eye alone would recognize, and which makes me think it is an appropriate task for ML.

Other question : is there a python ML library that would implement the recommended method? (PyBrain, Scikit ? ...?)

tm8cc
  • 1,111
  • 2
  • 12
  • 26

2 Answers2

1

It's probably possible to extract some time-series features over moving windows over your data, label manually some events as positive examples of tornado occurrences and treat the rest of the samples as negative and then fit a classifier to tell apart positive tornado event from random weather conditions as measured from your sensors.

How many such events do you have in your data? If think you would require at least 100 tornado events to be able to train a reliable enough model with a good enough estimation of its predictive accuracy.

Would be open to distribute this dataset publicly, e.g. on http://figshare.com?

ogrisel
  • 39,309
  • 12
  • 116
  • 125
  • I agree that this is the way to go. You could easily apply a neural network to this problem. – Narthring Jan 05 '13 at 02:07
  • If the OP is new to machine learning if would advise him/her to start with other models such as penalized linear classifiers (e.g. linear suppor vector machines or logistic regression), Support Vector Machines or Random Forests. Those models are simpler to use correctly (less hyperparameters to grid search) than neural networks. Be careful though that both logistic regression, support vector machines and neural networks need input scaling (e.g. to the [0-1] range for instance). Input scaling is not required (both does not hurt) when using random forests. – ogrisel Jan 05 '13 at 16:53
1

There are a wide variety of machine learning algorithms. The information you have provided does not suggest any one group of algorithms as superior without further investigation. An extended time series does suggest that you may need an algorithm to create reduced feature vectors. A neural network will not automatically provide you with a reliability coefficient. If you are willing to distribute a dataset publicly of several hundred random positive and negative examples then it's likely a number of different groups would apply various algorithms over time. Various "contest" sites exist which would speed up this process.

Clif Davis
  • 11
  • 1