Model overview

This page summarises the main model types provided by energy_fault_detector.

For configuration details see Configuration and Sequence models.

Autoencoders

Autoencoders learn a model of normal behaviour and are configured under train.autoencoder in the YAML config.

Class	Typical config names (`train.autoencoder.name`)	Description
`MultilayerAutoencoder`	`"default"`, `"MultilayerAutoencoder"`	Dense symmetric autoencoder for tabular or time-series data treated as independent rows. Good default choice.
`ConditionalAE`	`"ConditionalAE"`, `"ConditionalAutoencoder"`	Dense autoencoder where selected features are used as conditions and not reconstructed (e.g. time-of-day).
`LSTMSeq2OneAutoencoder`	`"LSTMSeq2OneAutoencoder"`, `"lstm_seq2one"`	Sequence-to-one LSTM autoencoder. Uses windows of historical data to reconstruct the last timestep.
`BidirectionalLSTMSeq2OneAutoencoder`	`"BidirectionalLSTMSeq2OneAutoencoder"`, `"bilstm_seq2one"`	Bidirectional LSTM variant of the seq2one model for richer temporal context.
`CNNSeq2OneAutoencoder`	`"CNNSeq2OneAutoencoder"`, `"cnn_seq2one"`	CNN-based seq2one autoencoder. Efficient, focuses on local temporal patterns.
`LSTMSeqAutoencoder`	`"LSTMSeqAutoencoder"`, `"lstm_seq2seq"`	Sequence-to-sequence LSTM autoencoder. Reconstructs the full window, useful when localisation within the window matters.
`CNNAutoencoder`	`"CNNAutoencoder"`, `"cnn_seq2seq"`	Sequence-to-sequence CNN autoencoder using Conv1D/Conv1DTranspose layers.

All sequence models require a sequence_builder block in train.autoencoder.params. See Sequence models for details and examples.

Anomaly score classes map reconstruction errors to a scalar anomaly score per sample. They are configured under train.anomaly_score.

Class	Typical config names (`train.anomaly_score.name`)	Description
`RMSEScore`	`"rmse"`, `"RMSEScore"`	Root mean squared error over reconstruction errors. Default choice.
`MahalanobisScore`	`"mahalanobis"`, `"MahalanobisScore"`	Mahalanobis distance on (optionally PCA-transformed) reconstruction errors. Captures feature correlations.

Threshold selectors map anomaly scores to boolean anomaly decisions and are configured under train.threshold_selector.

Class	Typical config names (`train.threshold_selector.name`)	Description
`QuantileThresholdSelector`	`"quantile"`	Sets the threshold to a fixed quantile of the (normal) anomaly scores. Works without labels.
`FbetaSelector`	`"fbeta"`, `"FbetaSelector"`	Chooses the threshold that maximises the F-beta score (requires labels / `normal_index`).
`FDRSelector`	`"fdr"`, `"FDRSelector"`	Chooses the threshold to match a target false discovery rate (requires labels / `normal_index`).
`AdaptiveThresholdSelector`	`"adaptive_threshold"`, `"AdaptiveThresholdSelector"`	Learns an input-dependent threshold using a small regression NN on autoencoder inputs and scores.

For a full reference of model parameters, see the API docs: