ocean_wnn

OceanWNN

Citekey	WangEtAl2019Study
Source code	own
Learning type	semi-supervised
Input dimensionality	univariate

Dependencies

• python 3
• pytorch

Activation Functions

The model uses wavelet basis functions (WBF) as activation functions for the hidden layer. The following plot shows the 3 different options:

Lower Bound (a=-2.5, k=-1.5)

Upper Bound (a=2.5, k=1.5)

Detection Strategies

The model keeps an internal threshold, which is the upper percentile of the normal distribution of its training errors. When the model detects a point whose error exceeds this threshold, the next sliding window changes the outlier point with the predicted point.

Anomaly Scores

OceanWNN forecasts points based on a preceding window. Hence, the first window_size points do not have a score.

Future Ideas

Using LSTMs instead of Linear layers to allow multivariate data.