Open datasets: condition monitoring and physical AI#

Ownership decision#

Dataset

Primary repository

Deliverable

InfraredSolarModules

Energy Data Insights

Labeled solar anomaly classification and review evidence

AI4I 2020

Energy Data Insights

Synthetic industrial failure classification benchmark

PushT

Physical AI

Episode-held-out action prediction

DROID

Physical AI

Recorded robot-motion audit before policy learning

Solar images belong here while their use is asset analysis. Physical AI can later reuse the trained perception method inside a robot. Energy Agents consumes metrics and identified evidence to draft investigation proposals; it does not own these dataset models. See Physical AI’s docs/datasets.md for robot examples.

Run the examples#

From the repository root, in an existing EDI environment or a separate benchmark virtual environment (Python 3.12+):

python -m pip install -r examples/datasets/requirements.txt
python -m pip install --no-deps ./analysis
export PYTHONPATH=analysis/src
# Windows PowerShell: $env:PYTHONPATH = "analysis/src"

AI4I: failure screening from operating measurements#

Download the official UCI ZIP and extract ai4i2020.csv to data/ai4i/.

python examples/datasets/run_condition_benchmark.py ai4i data/ai4i/ai4i2020.csv --output outputs/ai4i

Uses five measurements: air/process temperature, rotational speed, torque and wear. Product type is deliberately omitted in this numerical baseline. IDs and TWF/HDF/PWF/OSF/RNF failure-mode labels are never input features: using them would leak the target. A fixed 75/25 stratified split and random forest are compared against the majority-class predictor. This is contemporaneous failure classification, not a future-failure forecast or remaining-useful-life model.

The 10,000 records are synthetic industrial data. They do not validate a model for PV inverters, wind turbines or BESS. Use the example to learn class imbalance, precision/recall, reproducibility and evidence reporting before collecting appropriate equipment data. The random row holdout is not a temporal validation. Source: UCI AI4I, CC BY 4.0, DOI 10.24432/C5HS5C.

InfraredSolarModules: classify solar inspection anomalies#

Download the publisher archive and extract it to data/infrared/, retaining its InfraredSolarModules directory.

python examples/datasets/run_condition_benchmark.py infrared data/infrared/InfraredSolarModules --output outputs/infrared

Loads module_metadata.json and 20,000 source images; removes exact decoded-image duplicates before the split and excludes all images in contradictory-label duplicate groups. A random forest on flattened normalized grayscale pixels is a baseline for the 12 classes. It is not a deep-learning model. The source images have width 24 and height 40 pixels.

The available metadata does not supply plant-level groups, so this random-image holdout cannot establish performance at unseen plants. Near-duplicate scenes may remain. Report macro F1 and per-class recall instead of accuracy alone. The images do not establish lost kWh, financial recovery, fault causality or remaining life. Source: Raptor Maps, MIT repository license. Retain its license and paper attribution with any data copy.

Outputs and Energy Agents handoff#

Both examples write metrics.json, split.csv, predictions.csv and input_sha256.json. The metrics include a confusion matrix and majority baseline. Source IDs and exact input hashes allow evidence to be traced to the downloaded snapshot. Dataset downloads are explicit and excluded from repository commits.

The JSON is an evidence artifact, not a native Energy Agents API payload. A future adapter can attach it to a reviewed case alongside asset mappings, measurement dates and uncertainty. Those fields cannot be invented from these datasets. Reports explicitly request human review and set dispatch_authorized=false. No maintenance dispatch or claimed revenue recovery occurs.

Use a reviewed task proposal only after evidence quality, operating context and model applicability are established. Robot mission execution belongs in Physical AI.

Executed baseline results#

Dataset

Evaluated rows

Model macro F1

Majority baseline macro F1

AI4I

10,000

0.841

0.491

InfraredSolarModules

19,972 after duplicate/ambiguity filtering

0.396

0.056

The solar result is a modest baseline, not a deployment-ready classifier. Full per-class reports, input archive hashes and environment versions are committed under examples/datasets/results/. These results are for the fixed random split only. Dataset filtering removes 28 source records in total; this includes redundant copies and all members of conflicting-label groups.