Time series Augmentation Toolkit

A simple toolkit that helps researchers or specialists to generate more training dataset by augmentation!

Requirements

• numpy>=1.12

Implementations

1. Distributed Noise (DN):

   adding a noise with defined distribution

2. Window Slicing (WS): ~TODO

   extracting slices from time series and recombine the them together in orders

3. Window Warping (WW): ~TODO

   warping a randomly selected slice of a time series by speeding it up or down

4. Dataset Mixing (DM): ~TODO

   tune the extrapolating between time series in feature space

Functions

def DN(ts, sigma=0.0, mu=1.0, alpha=1.0, random_seed=None):
  '''
  ts: np.array, time series data
  sigma: float, standard deviation of the noise distribution
  mu: float, mean of the noise distribution
  alpha: float, the ratio of changes (new_value = alpha * delta + original_value)
  random_seed: int, set numpy random seed
  '''

################ TODO ################
def WS(ts_set, n=None, random_seed=None):
  '''
  ts_set: np.ndarray, time series dataset
  slicing_size: int, size of the sliced series, if None then default to 1/4 data length
  random_seed: int, set numpy random seed
  '''

def WW(ts, n=None, alpha=0.5, random_seed=None):
  '''
  ts: np.array, time series data
  n: int, influence range (+/-n values), if None then default to 1
  alpha: float, the ratio of streching
  random_seed: int, set numpy random seed
  '''
  
def DM(ts_set, random_seed=None):
  '''
  ts_set: np.ndarray, time series dataset
  
  //TODO
  
  random_seed: int, set numpy random seed
  '''

References

1. Le Guennec, Simon Malinowski, and Romain Tavenard: Data Augmentation for Time Series Classification using Convolutional Neural Networks
2. Terrance DeVries and Graham W. Taylor: Dataset Augmentation in Feature Space