energy_fault_detector.autoencoders.sequence_autoencoder

Base class for all sequence (windowed) autoencoders.

Note: not part of .core since it depends on SequenceDatasetBuilder.

class SequenceAutoencoder(sequence_builder=None, **ae_kwargs)

Bases: Autoencoder

Abstract base for sequence autoencoders (seq2one and seq2seq).

Subclasses must implement:

• create_model

• _build_dataset

• _assemble_predictions

encode(x, conditions=None, **kwargs)

Encode input time series into the latent space.

Parameters:

• x (DataFrame) – Input data (DataFrame with DatetimeIndex).

• conditions (DataFrame) – Unused (kept for API compatibility).

Return type:

ndarray

Returns:

Latent representations as NumPy array.

fit(x, x_val=None, **kwargs)

Fit the sequence autoencoder on time-series data.

Parameters:

• x (DataFrame) – Training data (DataFrame with DatetimeIndex).

• x_val (Optional[DataFrame]) – Optional validation data.

• **kwargs – Passed to model.fit.

Return type:

SequenceAutoencoder

Returns:

The fitted instance.

get_reconstruction_error(x, reconstruction=None, **kwargs)

Compute reconstruction error (predicted - actual) for main features.

Parameters:

• x (DataFrame) – Input data (DataFrame with DatetimeIndex).

• reconstruction (DataFrame) – pre-computed reconstruction. If None, predict() is called internally.

• kwargs – other keyword args for the keras Model.predict method.

Return type:

DataFrame

Returns:

DataFrame with reconstruction errors.

predict(x, **kwargs)

Predict/reconstruct input data.

Parameters:

x (DataFrame) – Input data (DataFrame with DatetimeIndex).

Return type:

DataFrame

Returns:

Reconstructed DataFrame.

tune(x, x_val=None, learning_rate=0.001, tune_epochs=5, **kwargs)

Fine-tune the model for additional epochs.

Parameters:

• x (DataFrame) – Training data.

• x_val (Optional[DataFrame]) – Optional validation data.

• learning_rate (float) – Learning rate for tuning.

• tune_epochs (int) – Number of additional epochs.

• **kwargs – Passed to model.fit.

Return type:

SequenceAutoencoder

Returns:

The tuned instance.

tune_decoder(x, x_val=None, learning_rate=0.001, tune_epochs=5, **kwargs)

Tune decoder only — encoder weights are frozen.

Return type:

SequenceAutoencoder