Published research
Our claims trace to a published paper and measured experiments. The model, the math, and the numbers are on this page — and in the PDF.
The paper
Peer-reviewed and published at IEEE SoutheastCon 2026.
SIGMA: A Physics-Informed AI Framework for Predictive Fault Probability Modeling in Microgrids
R. Monemi and S. Monemi · IEEE SoutheastCon 2026
SIGMA (Smart Infrastructure Grid Monitoring AI) computes a real-time fault probability from power-quality features using an interpretable logistic core and sliding windows, with an optional Kalman filter for noise rejection. The approach bridges classical reliability analysis and modern AI-based predictive maintenance — probabilistic learning grounded in the physics of how faults actually develop.
The method
SIGMA sits between deterministic observers and opaque deep nets: a lightweight, interpretable logistic model over five power-quality signals — voltage sag, total harmonic distortion, power-factor deviation, \(\dot V\), and \(\dot I\) — merging measurable physical indicators with probabilistic learning.
Fault-probability model
A logistic core maps the feature vector to a calibrated fault probability \(p(t)\) every sample.
Coefficients are constrained \(\beta_i \ge 0\) to enforce monotonicity: worse power quality always implies higher fault risk.
Regularized training
Parameters are learned by \(\ell_2\)-regularized maximum likelihood over labeled windows.
A sliding window updates \(p(t)\) on every sample for continuous, online estimation.
State machine
The probability stream is classified into operating states that drive alerting.
Probability rises smoothly in anticipation of a fault rather than spiking only after it occurs.
Kalman residual preprocessing
An optional extended Kalman filter models nominal feeder behavior; residuals emphasize deviations and suppress noise.
Residual filtering cut the false-alarm rate by roughly half in testing.
Measured results
On a 12 V DC microgrid testbed and Pandapower digital-twin simulations, leave-one-run-out cross-validation.
lead time before onset
average detection rate
ROC AUC
inference per window
| Fault type | Lead time (s) | Detection (%) | False alarms (/h) |
|---|---|---|---|
| Voltage sag | 12.4 | 95 | 0.09 |
| Overload | 11.8 | 92 | 0.10 |
| Line-to-line / line-to-ground | 13.1 | 94 | 0.08 |
Kalman preprocessing reduced the false-alarm rate by roughly 50%, and probability output rose smoothly in anticipation of each event — prognostic, not reactive.
Validation environment
The published results were measured on a physical 12 V DC testbed and a Pandapower digital twin — not one convenient dataset.
12 V DC microgrid testbed
A physical testbed instrumented with Arduino + INA226 sensing at 1 Hz, streaming live voltage, current, power-factor, and harmonic measurements into the model.
Pandapower digital twin
A Pandapower digital twin generates labeled fault runs — line-to-line, line-to-ground, overload, and harmonic distortion — for controlled, repeatable training and evaluation.
Cal Poly Pomona Smart Grid Laboratory
Beyond the published experiments, SIGMA runs on live data at the Cal Poly Pomona Smart Grid Laboratory, where the system prototype has been demonstrated in a relevant environment.
From research to practice
The same rigor runs through our commercial studies. Acceptance criteria agreed up front. Results that are auditable line by line. No black boxes — every number on a Monegrid deliverable can be traced, checked, and defended.
Talk to the authors
If you work on microgrids, power-system protection, or data-center power and want to discuss SIGMA — or put its rigor to work on your studies — reach out directly.
Book a discovery call