📜 Our Manifesto

📜 Our Manifesto#

PdMLabs is an open-source Python framework for predictive maintenance (PdM) experimentation. It is designed to help teams detect early warning signs and model time-to-event behavior, while keeping experiments reproducible, comparable, and extensible.

This project is not just a collection of isolated models. It provides a full experiment orchestration layer around data preparation, method execution, post-processing, thresholding, hyperparameter optimization, and evaluation.

What Problem It Solves#

In real industrial settings, teams usually face one or more of these challenges:

  • It is unclear which modeling flavor is appropriate for a given use case.

  • Data arrives from multiple sources/assets with different event semantics.

  • Evaluation must reflect predictive-maintenance reality, not only generic metrics.

  • Reproducibility and fair comparison across methods are difficult.

PdMLabs addresses these by standardizing the experiment workflow and exposing clear interfaces for each stage of the pipeline.

Design Philosophy#

PdMLabs follows a few core principles:

  1. Pipeline-first design Experiments are built from explicit steps: preprocessor -> method -> postprocessor -> thresholder.

  2. Interface-based extensibility New methods and processing components can be added by implementing lightweight interfaces, without rewriting the experiment engine.

  3. Evaluation aligned with PdM goals Metrics emphasize practical early warning behavior (predictive horizon, lead time, episode-aware evaluation), with additional range/VUS-style metrics available.

  4. Optimization and tracking by default Parameter search (Bayesian/Random via Mango) and MLflow logging are integrated into the experiment lifecycle.

  5. Multi-source realism The dataset contract and event preference mechanism support multiple assets and source mappings, matching real deployment scenarios.

How The Framework Is Organized#

At a high level, the framework is organized around these packages:

  • pdmlabs.pipeline: Pipeline composition and event-aware utilities.

  • pdmlabs.experiment: Experiment flavors and execution orchestration.

  • pdmlabs.method: Detection/prediction techniques (semi-supervised, unsupervised, supervised).

  • pdmlabs.preprocessing: Input transformations before model scoring.

  • pdmlabs.postprocessing: Score transformations after model scoring.

  • pdmlabs.thresholding: Threshold inference or score-to-target conversion.

  • pdmlabs.evaluation: PdM-focused metrics and evaluation utilities.

  • pdmlabs.utils: Helper utilities, including automatic parameter generation.

Supported Experiment Flavors#

PdMLabs supports multiple experiment modes, including:

  • Online auto-profile semi-supervised anomaly detection.

  • Incremental semi-supervised anomaly detection.

  • Semi-supervised anomaly detection.

  • Unsupervised anomaly detection.

  • Supervised classification.

  • Supervised RUL / survival-oriented workflows.

This allows users to compare fundamentally different strategies under a common execution and evaluation framework.

Extending PdMLabs#

You can extend PdMLabs by implementing one of the provided interfaces:

  • MethodInterface and its supervised/semi-supervised/unsupervised variants.

  • RecordLevelPreProcessorInterface.

  • PostProcessorInterface.

  • ThresholderInterface.

After implementation, your custom component can be plugged into run_experiment just like built-in components.

Current Scope#

PdMLabs is strong in batch experimentation and comparative evaluation. Streaming-related modules are present but currently more limited than batch workflows.

The recommended path for new users is:

  1. Read 📚 Concepts.

  2. Follow Quickstart.

  3. Explore 💡 Examples.

  4. Use 📖 API Reference for implementation details.

Contents