This strategy combines autoencoding and predictor.
For comparison, we have implemented two options:
- standard autoencoder + LSTM predictor + anomaly detector;
- variational autoencoder + LSTM predictor + anomaly detector;
It's possible to use just an autoencoder.
We use prediction/reconstruction error to define anomaly. Anomaly detector is based on the 99.7 rule, also known as the empirical rule. Means that 99.73% of the values lie within three standard deviations of the mean.
Using a combination of autoencoders&predictor
from models import V_AE_LSTM
input_shape = (X_train.shape[1], X_train.shape[2],)
intermediate_cfg = [64, 'latent', 64]
latent_dim = 10
model = V_AE_LSTM(input_shape, intermediate_cfg, latent_dim, 'VAE-LSTM') # or 'AE-LSTM'
model.fit(X_train, y_train, epochs=2, batch_size=124, validation_split=None, verbose=1)Using a normal autoencoder
from models import LSTM_Autoencoder
input_shape=(X_train.shape[1], X_train.shape[2],)
intermediate_cfg = None #[64, 'latent', 64]
latent_dim = 10
ae = LSTM_Autoencoder(input_shape, intermediate_cfg, latent_dim)
ae.fit(X_train, epochs=10, batch_size=32, validation_split=None, verbose=1)
reconstructed = ae.reconstruct(X_test)Using a variational autoencoder
from models import LSTM_VAutoencoder
input_shape=(X_train.shape[1], X_train.shape[2],)
intermediate_cfg = [64, 32, 'latent', 32, 64]
latent_dim = 10
vae = LSTM_VAutoencoder(input_shape, intermediate_cfg, latent_dim)
vae.fit(X_train, epochs=10, batch_size=32, validation_split=None, verbose=1)
reconstructed = vae.reconstruct(X_test)