Athmane BenarousHydro-Québec Research Institute (IREQ) · Machine Learning Research Intern · May – Sep 2026
Slow-leak detection for high-voltage circuit breakers
0.867 vs 0.733sensitivity vs the physical baseline on synthetic leaks, at an equal false-event budget; the gain is conditional
Industrial research project. No code, data or report figures appear here; the diagrams are schematic.
Problem
Circuit breakers rated 145–315 kV are filled with SF₆, a gas that insulates the contacts and quenches the arc when the breaker interrupts a fault. A slow leak has to be inferred from quarter-hourly pressure and temperature records.
The difficulty is scale. Pressure moves about 40 mbar per °C, while a 1%/year leak removes only about 0.27 mbar per day. Could a learned detector find that signal better than physical rules: a temperature-corrected pressure baseline and temperature-slice methods?
Constraints
- One known leaking device. Positive examples are synthetic: known declines injected into healthy records, with physically unvalidated fidelity.
- Overlapping windows. Consecutive 30-day windows share 29 days, so a random split puts near-copies on both sides and makes false alarms look rarer than they are.
- Alarms cost dispatches. What matters is false events per device-year at a fixed budget, not accuracy.
What I built
An evaluation protocol designed for overlapping windows, and a detector selected under it:
- Segment-grouped splits: windows are grouped by continuous measurement segment, so no test window shares days with a training window.
- An invariance screen keeping 34 of 93 features stable across healthy time blocks.
- Thresholds from healthy data only, with exact Poisson upper bounds and a multiplicity correction.
- A 7-day persistence rule before an alarm is confirmed.
- Dispatches, not episodes: an alarm that never clears is one episode but keeps sending crews.
I compared an L1-regularized logistic regression (scikit-learn) with the physical baseline, temperature-slice methods and a temporal convolutional network (PyTorch), over multiple seeds on a cluster scheduler. The selected model was delivered as a frozen, integrity-checked package with reproducible configurations.
- 93 candidate features
- Invariance screenkeeps 34 stable across healthy time blocks
- L1-regularized logistic regression
- Threshold from healthy data onlyexact Poisson upper bound
- 7-day persistence rule
- Dispatchcounted per dispatch, not per episode
Results
- Sensitivity 0.867 vs 0.733 for the physical baseline, at an equal synthetic false-event budget.
- The gain is conditional: below 1% loss per year, the physical baseline was stronger.
- 3.41 false events per device-year over 429 healthy device-days (exact 95% upper bound: 11.35).
- All 4 segments of the one known leaking device confirmed, after 6, 16, 6 and 6 days.
The model is not certified for operational use.
What went wrong and how I fixed it
- Features that recognized the device, not the leak. 44 of 93 features shifted by more than one standard deviation between two healthy devices. I excluded device-identity features, and the invariance screen keeps only features that stay stable across healthy time blocks.
- A sequence model that looked blind. At a frozen threshold, the TCN seemed to miss the leaking device; a threshold sweep showed it confirmed all 4 segments at the common budget. That was a calibration problem, not a transfer problem, so I compare models at an equal false-event budget.
- Too much capacity. The TCN had 224,642 parameters for 1,477 training sequences and lost 0.063 AUC from simulation to field. I selected the regularized linear model.
What I’d do next
- Have specialists validate the leak-injection physics; every sensitivity number depends on it.
- Add devices: with one leaking device, a leak and a device’s identity are hard to separate.
- Collect more healthy device-days to tighten the 11.35 upper bound, then run a prospective evaluation before any operational use.