pdmlabs.preprocessing.record_level.feature_selector

pdmlabs.preprocessing.record_level.feature_selector#

Feature selection preprocessor for dimensionality reduction.

FeatureSelector filters a DataFrame to keep only selected columns/features. Useful for: - Removing noisy/irrelevant features - Dimensionality reduction - Domain-specific feature sets

Classes

FeatureSelector(event_preferences, ...)

Select a subset of features for anomaly detection.
class pdmlabs.preprocessing.record_level.feature_selector.FeatureSelector(event_preferences: EventPreferences, selected_features: list[str])#

Bases: RecordLevelPreProcessorInterface

Select a subset of features for anomaly detection.

This preprocessor is stateless and deterministic: given a fixed list of feature names, it returns only those columns from input data. Useful for: - Domain knowledge: select physically meaningful sensors - Dimensionality reduction: drop highly correlated features - Noise filtering: remove sensors with too much noise - Feature importance: keep top-K features from prior analysis

selected_features#

Column names to keep from input DataFrames.

Type:

list[str]

Examples

>>> from pdmlabs.preprocessing.record_level.feature_selector import FeatureSelector
>>> import pandas as pd
>>>
>>> # Data with 5 features, but we only want 3
>>> df_train = pd.DataFrame({
...     'vibration': [1, 2, 3],
...     'temp': [50, 60, 70],
...     'pressure': [100, 102, 104],
...     'noise1': [0.1, 0.2, 0.3],
...     'noise2': [0.05, 0.07, 0.09]
... })
>>> df_test = pd.DataFrame({...})  # Similar structure
>>>
>>> selector = FeatureSelector(
...     event_preferences={'failure': [], 'reset': []},
...     selected_features=['vibration', 'temp', 'pressure']
... )
>>> selector.fit([df_train], ['bearing_1'], events_df)
>>> df_test_selected = selector.transform(df_test, 'bearing_1', events_df)
>>> # df_test_selected now has only 3 columns: vibration, temp, pressure
fit(historic_data: list[DataFrame], historic_sources: list[str], event_data: DataFrame, anomaly_ranges=None) → None#

Fit feature selector (no-op, just placeholder).

Feature selection is stateless, so fit() does nothing. The selected features are fixed at initialization.

Parameters:

  • historic_data (list[pd.DataFrame]) – Ignored.

  • historic_sources (list[str]) – Ignored.

  • event_data (pd.DataFrame) – Ignored.

  • anomaly_ranges – Ignored.

get_params()#

Return hyperparameters.

Returns:

{‘features’: list of selected feature names}

Return type:

dict

Examples

>>> print(selector.get_params())
{'features': ['vibration', 'temp', 'pressure']}
transform(target_data: DataFrame, source: str, event_data: DataFrame) → DataFrame#

Select subset of features from target data.

Parameters:

  • target_data (pd.DataFrame) – Test data to filter.

  • source (str) – Source identifier (unused).

  • event_data (pd.DataFrame) – Event log (unused).

Returns:

Subset of target_data with only selected_features columns.

If selected_features is empty, returns input unchanged (fallback).

Return type:

pd.DataFrame

Raises:

KeyError – If any selected feature not in target_data.columns.

Examples

>>> df_test_selected = selector.transform(df_test, 'bearing_1', events_df)
>>> print(df_test_selected.columns.tolist())
['vibration', 'temp', 'pressure']
>>> print(df_test_selected.shape)
(100, 3)  # 100 rows, 3 columns
transform_one(new_sample: Series, source: str, is_event: bool) → Series#

Select features from a single sample.

Parameters:

  • new_sample (pd.Series) – Single row (Series with feature names as index).

  • source (str) – Source identifier (unused).

  • is_event (bool) – Event flag (unused).

Returns:

Subset of new_sample with only selected_features.

Return type:

pd.Series

Examples

>>> new_row = pd.Series({'vibration': 1.5, 'temp': 65, 'pressure': 103, 'noise1': 0.15, 'noise2': 0.06})
>>> selected = selector.transform_one(new_row, 'bearing_1', False)
>>> print(selected.tolist())
[1.5, 65, 103]
Contents