Sustainability · Technology · predictive maintenance · live
Six days of real plant data, taken blind. Diagnose them, then face the lab's answer key.
This is real fault-injection data from two research facilities: LBNL's FLEXLAB air-handling unit and ORNL's instrumented rooftop unit, with faults physically imposed and ground truth recorded (published CC-BY 4.0). Pick a day blind. Residual rules, with their baselines computed live from the fault-free days, flag the anomalous readings; the diagnosis then walks the airflow path and refrigeration circuit upstream and downstream to isolate the component. Only after the engine commits do you reveal the lab's label and score it. One of the six is a trap the reveal explains. No machine learning, no black box: every rule and threshold is on the second tab.
Real data · LBNL FLEXLAB + ORNL RTU · OEDI submission 910 · CC-BY 4.0
The whole diagnostic brain, on one page. Baselines are computed in your browser from the two fault-free days; thresholds are fixed engineering margins, declared before any faulty day is examined. A rule that cannot run on a given day says so rather than guessing: the damper rule, for example, is unidentifiable when outdoor and return air are at the same temperature.
Runs fully in your browser over the published extract. Source: Granderson et al., Scientific Data 7, 65 (2020) · OEDI submission 910 · CC-BY 4.0. · willytai.com