Experiment Fabric

The Experiment Fabric treats an experiment as a first-class, reproducible object rather than a notebook cell or a chart. Its contract connects the engineering question to the exact evidence returned:

Complete experiment anatomy

  1. Asset

    Versioned machine identity and twin context.

  2. Problem

    Bearing degradation as the bounded failure mode.

  3. Dataset

    Synthetic, versioned fixture with explicit provenance.

  4. Operating regime

    The envelope in which evidence is interpreted.

  5. Target and horizon

    Seven-day detection objective and outcome.

  6. Feature set

    Process, vibration, physics, or combined inputs.

  7. Algorithm

    Candidates compete under identical conditions.

  8. Configuration

    Exact selections, versions, and random seed.

  9. Validation

    Time-aware evaluation without future leakage.

  10. Metrics and cost

    Performance, false alarms, lead time, and compute.

  11. Qualification

    Robustness, explainability, and uncertainty.

  12. Evidence result

    Limitations and human authority remain attached.

The experiment record binds the engineering question, reproducible configuration, validation discipline, qualified result, and decision boundary.

P-101 example

| Field | Configuration | | ------------------------ | -------------------------------------------------------------- | | Asset | Boiler Feed Water Pump P-101 | | Objective | Detect bearing degradation | | Prediction horizon | 7 days | | Candidate models | Isolation Forest, XGBoost, Autoencoder, Physics Residual Model | | Candidate feature groups | temperature, vibration, process, physics-derived, temporal | | Validation | Walk-forward validation | | Operational constraint | Maximum one false alert per month |

The search target is the best solution for a particular machine, operating regime, failure mode, and operational constraint—not the algorithm with the highest accuracy. A detector with a high aggregate detection rate may be unusable if false alerts overload maintenance, if lead time disappears in a high-flow regime, or if its required sensors are unreliable.

Validation discipline

Time split, walk-forward, and leave-one-regime-out validation answer different questions. The split must respect temporal ordering and prevent future information from leaking into training. Metrics should be reported by operating regime and with uncertainty, not reduced to a single flattering average. Failed candidates remain part of the experiment history because they constrain later hypotheses and prevent expensive repetition.

Evidence Package

Every result binds the model, dataset, feature set, operating regimes, validation method, performance, false-alarm rate, lead time, robustness, uncertainty, explainability, limitations, data provenance, experiment ID, twin version, and code version. It is a model card for an engineering decision, not a deployment certificate.

If an industrial AI result cannot be reproduced, it should not be trusted.

The conceptual demonstrator compares deterministic fixtures so the interface and evidence contract can be inspected. It does not train models or claim measured performance.

Conceptual demonstration — synthetic fixture results.

P-101 bearing-degradation Evidence Package

EXP-P101-BD-COMBINED-XGBOOST-WALKFORWARD

Conceptual demonstration — synthetic fixture results.

Experiment record

Model
MODEL-XGB-0.1.0
Model status
experimental
Dataset
DATASET-P101-SYN-0.1.0
Feature set
combined
Validation
walk-forward
Twin version
TWIN-P101-0.1.0
Asset version
ASSET-P101-0.1.0
Dataset version
DATASET-P101-SYN-0.1.0
Simulator version
SIM-P101-0.1.0
Feature pipeline
FEATURES-P101-0.1.0
Provenance model
MODEL-XGB-0.1.0
Code version
ITL-PHASE-1-0.1.0
Configured asset
P-101
Problem
bearing-degradation
Configured feature set
combined
Algorithm
xgboost
Configured validation
walk-forward
Random seed
101
Timestamp label
Synthetic fixture

Qualified evidence

EvidenceSummaryStrengthOrigin
EXP-P101-BD-COMBINED-XGBOOST-WALKFORWARD-SYNTHETIC-EVIDENCEDeterministic synthetic fixture for conceptual comparison.limitedSynthetic fixture

Metric results

MetricResult
Detection Rate86 %
False Alarms0.9 alerts/month
Lead Time5 days
Inference Cost18 ms
Sensor Count11 sensors
Robustness80 /100
Explainability78 /100

Operating regimes

  • Nominal flow: 220–260 m³/h
  • Rated speed: 2,900–3,000 rpm
  • Ambient temperature: 5–45 °C

Limitations

  • Synthetic fixture results do not establish plant performance.
  • No control decision, alarm threshold, or maintenance action is authorized.
  • Simulation fidelity and transferability require independent engineering validation.

Uncertainty, explainability, and provenance

Uncertainty
Illustrative uncertainty only; no confidence value is derived from an operating machine.
Explainability
XGBoost compared under Walk Forward.
Dataset source
Industrial Twin Lab synthetic experiment fixture
Dataset disclosure
This dataset is a deterministic synthetic fixture for conceptual comparison only; it is not plant data.
Experiment source
Industrial Twin Lab deterministic experiment fixture lookup
Experiment disclosure
Conceptual demonstration — synthetic fixture results. No model is trained or executed and no real plant data is used.
Author agent
Industrial Twin Lab synthetic fixture agent

Human engineer retains decision authority. This package does not authorize control, maintenance action, or safety certification.

The fictional P-101 example keeps model and feature competition, validation, provenance, uncertainty, limitations, and human review in one reproducible record.