pdmlabs.method.PCA_uns

pdmlabs.method.PCA_uns#

Classes

PCA_uns(event_preferences, window[, n_bins, ...])

class pdmlabs.method.PCA_uns.PCA_uns(event_preferences: EventPreferences, window, n_bins=10, alpha=0.1, tol=0.5, sub_sequence_length=1, *args, **kwargs)#

Bases: UnsupervisedMethodInterface

get_all_models()#

Return reference to internal model(s).

Returns model instances for inspection, export, or further processing. Structure depends on method - may return single model or dict of models.

Returns:

Underlying model object(s). Examples:

  • Single model: sklLearn model instance

  • Multiple models: {‘bearing_1’: model1, ‘bearing_2’: model2}

  • None: If model is not accessible/applicable

Return type:

model or dict

get_library() → str#

Return the underlying library/framework name.

Returns:

Name of library used (e.g., ‘sklearn’, ‘torch’, ‘custom’).

Used for dependency tracking and method categorization.

Return type:

str

get_params() → dict#

Return hyperparameters and configuration.

Returns:

Dictionary of hyperparameters (e.g.,

{‘n_neighbors’: 5, ‘contamination’: 0.1}). Useful for logging, model comparison, and reproducibility.

Return type:

dict

predict(target_data: DataFrame, source: str, event_data: DataFrame) → list[float]#

Predict anomaly scores for batch of samples (offline mode).

Computes anomaly score for each row in target_data independently. Higher scores indicate more anomalous behavior.

Parameters:

  • target_data (pd.DataFrame) – Feature matrix with dates in index, features in columns. Must have same features as training data.

  • source (str) – Source identifier (e.g., ‘bearing_1’). Used to select source-specific model if method maintains multiple models.

  • event_data (pd.DataFrame) – Event log with columns ‘date’, ‘type’, ‘source’, ‘description’. Can be used for context-aware scoring.

Returns:

Anomaly scores (float) with length = target_data.shape[0].

Score range and semantics depend on method: - Distance-based: typically [0, ∞) where higher = more anomalous - Probability-based: typically [0, 1] or (-∞, 0] log-likelihood - Reconstruction-based: typically [0, ∞) reconstruction error

Return type:

list

Examples

>>> method = SomeAnomalyDetector(event_preferences={...})
>>> method.fit([df_train], ['bearing_1'], events_df, labels)
>>> df_test = pd.DataFrame([feature values], index=[dates])
>>> scores = method.predict(df_test, 'bearing_1', events_df)
>>> print(len(scores), scores[0])  # (100, 0.45)
Raises:

NotImplementedError – If method hasn’t been fit (for supervised methods).

predict_one(new_sample: Series, source: str, is_event: bool) → float#

Predict anomaly score for single sample (online/streaming mode).

Computes anomaly score for one observation at a time. Useful for: - Real-time anomaly detection - Online learning scenarios - Memory-efficient processing

May maintain internal state (buffers, windows) for context-aware scoring.

Parameters:

  • new_sample (pd.Series) – Single observation with feature values. Index should contain feature names matching training data.

  • source (str) – Source identifier for source-specific models.

  • is_event (bool) – Whether this sample is from an event timestamp. Can affect how method processes the sample (e.g., special handling for known events vs normal operation).

Returns:

Anomaly score for this single sample (same scale as predict()).

Return type:

float

Examples

>>> method = SomeAnomalyDetector(event_preferences={...})
>>> method.fit([df_train], ['bearing_1'], events_df, labels)
>>>
>>> # Online scoring
>>> for idx, row in df_test.iterrows():
...     is_event = idx in event_timestamps
...     score = method.predict_one(row, 'bearing_1', is_event)
...     print(f'{idx}: {score}')
Contents