See the fault
before it happens
SIGMA is a physics-informed engine that computes a real-time fault probability from power-quality signals — published, measured, and interpretable, not a black box.
What SIGMA does
Four capabilities, one engine: SIGMA turns the power-quality data your system already produces into advance warning, health scoring, and maintenance foresight.
Fault prediction
A calibrated fault probability, updated every sample — flagging incipient faults ~12 seconds before onset, while there is still time to act.
Anomaly detection
EWMA-based drift detection plus Isolation Forest multivariate scoring — catching abnormal behavior even when it does not match a known fault signature.
Node Health Index
A composite 0–100 health score for every monitored node — one number that tells operators where the system is strong and where it is degrading.
Remaining useful life
Degradation-curve fitting estimates remaining useful life per asset — so maintenance is scheduled by evidence, not by calendar or by failure.
The five power-quality features
Incipient faults leave fingerprints in power quality long before a relay would trip. SIGMA watches five physically meaningful signals — each one chosen because grid physics says it matters.
Short RMS dips from arcing or partial shorts.
Arcing injects nonlinear harmonics; THD rises sharply.
Arc or insulation faults shift the V–I phase angle.
Highlights abrupt sags and recoveries.
Captures the current spikes typical of arc ignition.
How it works
A real-time pipeline from sensing through inference to alerting — every stage inspectable, every output explainable.
The fault-probability model
An interpretable logistic core maps the five features to a calibrated fault probability \(p(t)\) over sliding windows, on every sample.
Coefficients are constrained \(\beta_i \ge 0\) to enforce monotonicity: worse power quality can only mean higher fault risk. An optional extended Kalman filter preprocesses the signals, modeling nominal behavior so that deviations stand out and noise is suppressed. Every coefficient is a number an engineer can read and defend.
The operating states
The probability stream drives a four-state machine that turns a continuous risk signal into unambiguous operator actions.
Because the model is grounded in physics, the probability rises smoothly in anticipation of a fault rather than spiking only after it occurs — the AHEAD state is the early warning. Full thresholds, training formulation, and Kalman residual math are on the research page.
Measured results
These are instrumented numbers, not projections — measured on a 12 V DC microgrid testbed and Pandapower digital-twin simulations under leave-one-run-out cross-validation.
lead time before fault 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 residual preprocessing cut the false-alarm rate by roughly 50% — and the probability output rose smoothly in anticipation of each event. Prognostic, not reactive.
Full methodologyDeployment & integration
SIGMA rides on the measurements your system already takes. No rip-and-replace, no proprietary sensors — it routes into the monitoring you have.
Serving stack
Python, FastAPI, and TimescaleDB with operator dashboards — a lightweight service that ingests telemetry and streams probabilities, health scores, and alerts.
Fits your existing stack
Outputs route directly into existing monitoring and SCADA systems. SIGMA consumes power-quality signals your infrastructure already measures — no new hardware required.
Validation status
System prototype demonstrated in a relevant environment — running on live data at the Cal Poly Pomona Smart Grid Laboratory. Municipal pilot discussions are underway in Southern California.
prototype · live lab dataWhere SIGMA goes next
The published core is the foundation. The roadmap scales it up — and points it at the loads where a missed fault costs the most.
Ensemble & adaptive learning
Ensemble models with adaptive online learning — recursive Bayesian updating — for robustness as operating conditions drift.
Distributed multi-node
A distributed implementation for multi-node, grid-scale microgrids — health and fault probability across the whole network, not one feeder.
AI data centers & industry
Extension to AI data centers and industrial facilities, where nonlinear loads are heaviest and the cost of a missed fault is highest.
Ask about a SIGMA pilot
If you operate a microgrid, campus, or critical facility and want predictive reliability on the data you already collect, let's talk about what a pilot looks like.