Models

autoencoder_anomaly()

Train an MLP bottleneck autoencoder on normal-operation data. At inference, returns a single

'anomaly_score' column per row (squared reconstruction error on scaled features). Pair with

hotelling_t2_monitor when you need non-linear anomaly detection — PCA captures linear variance,

this captures non-linear structure.

Unsupervised: do not set target_feat. NaN values must be handled upstream.

Properties:

• hidden_layer_sizes: tuple/list of hidden layer widths (default (16, 8, 16) = bottleneck 8).

Pass as a comma-separated string from the UI (e.g. '16,8,16').

• max_iter: training iterations (default 500)

• random_state: RNG seed for reproducibility (default 0)

gmm_regime_classifier()

Train a Gaussian Mixture Model to identify process regimes from multi-sensor data. At

inference, returns a single 'regime' column per row (integer cluster index, 0..n_components-1).

Good for separating operating modes (e.g. startup / steady / shutdown) without labels.

Unsupervised: do not set target_feat. NaN values must be handled upstream.

Properties:

• n_components: number of regimes / clusters (default 3)

• covariance_type: 'full' (default), 'tied', 'diag', or 'spherical'

• random_state: RNG seed for reproducibility (default 0)

gradiant_boost_classifier()

Train a Gradient Boosting classifier on all features to predict a target column.

hotelling_t2_monitor()

Train a PCA-based multivariate SPC monitor on training data representative of normal

operation. At inference, produces two columns per row: T² (Hotelling's statistic — anomaly

within the principal subspace) and SPE (squared prediction error — novelty orthogonal to the

PC space). The classical industrial monitoring pattern — far more interpretable than a black-box

outlier score because T² and SPE can be decomposed back to per-sensor contributions.

Unsupervised: do not set target_feat. NaN values must be handled upstream.

Property:

• n_components: int (exact number of components) or float in (0, 1] (minimum explained

variance ratio, default 0.95).

isolation_forest()

Train an Isolation Forest for unsupervised anomaly detection using a sensitivity level (1-5).

linear_regression()

Train a linear regression model on all features to predict a single target column.

logistic_regression()

Train a logistic regression classifier on all features to predict a binary target column.

mahalanobis_anomaly()

Train a Mahalanobis-distance anomaly detector on normal-operation data. At inference,

returns a single 'mahalanobis_distance' column per row — larger distance = more anomalous,

interpretable because each feature contributes proportional to its deviation weighted by

the inverse covariance. No dimensionality reduction (unlike hotelling_t2_monitor), so useful

when you have few-to-medium sensors and want a single calibrated score.

Unsupervised: do not set target_feat. NaN values must be handled upstream.

Properties: (none — uses sklearn's EmpiricalCovariance defaults)

random_forest_classifier()

Train a Random Forest Classifier on all features to predict a target class column — robust,

handles multi-class, no scaling required, exposes featureimportances for interpretability.

Good default for process-regime or fault-type classification.

Properties:

• n_estimators: number of trees (default 100)

• max_depth: max depth per tree (default None = unbounded)

• random_state: RNG seed for reproducibility (default 0)

random_forest_regressor()

Train a Random Forest Regressor on all features to predict a single target column — non-linear,

robust to outliers, no scaling required. Exposes featureimportances on the trained model

for interpretability (root-cause hints).

Properties:

• n_estimators: number of trees (default 100)

• max_depth: max depth per tree (default None = unbounded)

• random_state: RNG seed for reproducibility (default 0)

ridge_regression()

Train a Ridge (L2-regularized) linear regression to predict a single target column.

A drop-in replacement for linear_regression when features are highly correlated (common with

multi-sensor industrial data) — produces more stable coefficients.

Property:

• alpha: L2 regularization strength (default 1.0 — larger means more shrinkage)

setpoint_optimizer()

Train a KNN-based setpoint optimizer that learns quality = f(telemetry + setpoints) and at

inference recommends optimal setpoint values. Setpoint columns are auto-detected via

BusinessType.SET_POINTS on context.datapoints; bounds are the 5th-95th percentile of train data.

Required MLModelConfig:

• train_script = 'wizata.models.setpoint_optimizer'

• target_feat = '

• function = 'recommend' (or 'predict' for quality forecasting only)

• output columns_names = ['recommended', '_recommended', ...] in the order setpoints appear.

The trained model exposes .output_names() to suggest a matching default.

Properties:

• k: KNN neighbors (default 5)

• grid_size: points per setpoint axis (default 10 — total cost is grid_size^n_setpoints per row)

• direction: 'minimize' (default) or 'maximize' the target