Issues with a single sensor SVM

It would be very useful if we could train an SVM on a single sensor and detect picks on that sensor correctly. There are a few issues with doing that. 

We classify non-quake times as negative picks and quake times as positive picks. The training data is separated at the first time domain kSigma pick. Suppose it occurs at 34 seconds for some earthquake and some station. The algorithm takes an FFT for each 5 second interval before and after the pick, i.e. 0-5, 5-10, 10-15, 15-20, 20-25, 25-30 seconds for non-quake period and 35-40, 40-45, 45-50, 50-55, 55-60 seconds for quake period. We're not sure at this point if this is infact a good idea. It might not capture the information correctly. Also, since 5 seconds is a very small interval, and therefore is lesser data, even slight abnormalities can lead to disastrous results. 

We also face data insufficiency issues.There are only a few sensors for which we have the data for all earthquakes during the past few years. Even among those, a few stations did not pick for some of the earthquakes. Also, some stations picked at the 54th second out of the 60 seconds data. This causes there to be more quiet time data than quake time data. The SVM, in such cases, is highly biased towards negative picks and is very likely to predict a positive pick as a negative one.