Predictive maintenance platform

The anomsmith.platform subpackage is the operational PM layer: rolling and spectral feature extraction, RandomForest-based RUL and failure models, threshold alerts, streaming ingestion, and matplotlib dashboards. It was merged from the former standalone Anomaly Detection Toolkit (historically packaged as anomaly_detection_toolkit) so that one install delivers both low-level detectors and higher-level maintenance workflows.

Detectors and scorers remain in anomsmith.primitives (single BaseDetector hierarchy with LabelView / ScoreView outputs). The platform stack calls those detectors when you pass an optional anomaly_detector into PredictiveMaintenanceSystem.

Key modules

  • anomsmith.platform.predictive_maintenance — FeatureExtractor, RULEstimator, FailureClassifier, AlertSystem, PredictiveMaintenanceSystem, RealTimeIngestion, DashboardVisualizer, and dataset helpers (calculate_rul, prepare_pm_features, …).

  • anomsmith.platform.evaluation — evaluate_detector, compare_detectors, lead-time and confusion helpers using the same label convention as primitives (1 = anomaly).

  • anomsmith.platform.visualization — optional matplotlib plots wired to anomsmith detectors (e.g. plot_pca_boundary()).

Public re-exports also appear on the root anomsmith package for a single import story (from anomsmith import PredictiveMaintenanceSystem, …).

Migration from anomaly_detection_toolkit

  • Replace import anomaly_detection_toolkit with import anomsmith.platform or root anomsmith imports.

  • Do not import duplicate detector names from the old toolkit; use anomsmith.primitives.detectors instead.

  • prepare_pm_features(..., add_degradation_rates=...) was renamed to include_degradation_rates=... to avoid shadowing the add_degradation_rates function inside that helper.

The standalone asset_health / anomaly_detection_toolkit repository is no longer required for any documented capability; see Retiring anomaly_detection_toolkit (asset_health).