LSTM Based Anomaly Detection

Hypothetically it should be possible to autoencoder an LSTM in order to build an anomaly detector. This proves that this is possible.

Example

It is shown that a simple sin wave with added amplitudal noise can be detected if trained with a pure unsupervised LSTM autoencoder.

The entire signal as it is presented to the network [sliding window] :

The signal approximation (mean) :

The MSE output of the LSTM detector:

The number of noise samples added (and caught):