Asoba Asoba
Predictive Fault Detection

Catch the Fault. Skip the False Alarm. Get the Fix.

Static threshold rules can't tell a failing inverter from one running fine under cloud cover — so teams drown in false alarms until they learn to ignore the dashboard. Asoba instead builds a behavioral model of each asset's own normal operation, scores deviations continuously instead of flagging on/off, and hands your team a cited fix pulled from your own OEM documentation — not another alert to triage.

Per-Inverter Behavior Model Live telemetry · predicted vs. actual state Live model 100 75 50 25 0 kW Streak: 6 · Inverter 3 00:00 06:00 12:00 18:00 24:00 Severity: Critical Predicted vs. actual: −235 kW Diagnosis: Nehanda v3
Up to 97% Fewer false alerts at equal fault-catch rate
0.69–0.98 AUROC across evaluated production inverters
Up to 260% F1 improvement over rule-based thresholds
0 On-site visits or new hardware required to start

Based on evaluation across seven production inverters at a commercial solar site; results vary with data availability and fault history.

From Anomaly to Action, Automatically

Four steps, no manual dashboard-correlation in between.

1

Learn Normal Behavior

Per-Asset Model

A behavioral model is built for each inverter individually, conditioned on weather and time of day — not one fixed rule applied fleet-wide. What counts as "normal" for a shaded rooftop unit and a fully exposed ground-mount unit is learned separately, because it isn't the same.

2

Score the Deviation

Continuous, Not Binary

Deviation from expected behavior becomes a continuous severity score and a streak length, not a single on/off flag. Your team can prioritize the handful of detections that matter instead of triaging every alert equally — including the ones caused by a cloud, not a fault.

3

Ground the Fix

Cited, Not Generated

High-confidence detections are matched against your own OEM manuals, troubleshooting guides, and maintenance history, then synthesized into a cited recommendation by Nehanda v3 — a model fine-tuned specifically to cite its sources and refuse to fabricate a fix it can't support.

4

Route to Your Team

Deterministic, Auditable

The anomaly score and the recommendation stay separate from any automated dispatch action. A deterministic downstream layer decides what happens next, so the control loop stays auditable and free of generative-model guesswork where it matters most.

Proven on a Live Fleet

Tested Against Real Faults, Not a Lab Benchmark

Evaluated across seven inverters on a 2+ MW commercial rooftop site over six months of production telemetry, benchmarked directly against the site's existing rule-based alerting — not a synthetic dataset built to flatter the model.

  • 216 labeled faults caught on the fleet's highest-signal inverter, at a 41–56% precision and recall lift over rule-based thresholds
  • Two inverters showed weak signal during evaluation — diagnosed as an upstream data-backfill gap, not equipment failure, before it was ever forwarded as a false lead
  • Same production platform, security model, and validation pipeline used for a pilot as for ongoing monitoring — nothing stripped down to run a trial
Maintenance intelligence screenshot

Built for Your Engineering & Security Team

The pilot runs on the same platform, security model, and validation pipeline as ongoing monitoring — nothing about a trial is a stripped-down version of the product.

Deployment

Serverless · Isolated Failure Domains
  • Serverless, event-driven architecture — auto-scaling, no idle-cost overhead
  • In-region processing for local data residency requirements
  • Failure domains are isolated: an issue in one ingestion pipeline can't cascade into the detection engine

Data & Governance

Tenant-Isolated · Audited
  • Multi-OEM ingestion into a canonical schema — no manual spreadsheet wrangling on either side
  • Row-level, cryptographically enforced tenant isolation
  • Every access and manual override captured in an audit trail; infrastructure benchmarked weekly against CIS, ISO 27001, and SOC 2 controls

Free 3-Week Diagnostic Pilot

See Your Fleet's False-Alarm Number, Before You Commit to Anything

Send at least three months of inverter or SCADA production data (15-minute intervals ideal) and basic site details. We'll tell you within the first week whether your data clears the bar to start — and if it doesn't yet, we'll say so up front rather than take your data and hand back a weak answer.

  • No on-site visit, no new hardware, no engineering time from your team
  • No cost, and no obligation to continue afterward
  • Your data stays in-region and isolated for the duration of the pilot; request full deletion anytime after
Three Weeks, Three Outcomes
WEEK 1 — Find the Signal
We build a per-asset behavior model from what you send us and hold it up against actual output.
WEEK 2 — Rank the Faults
Every flagged asset is scored and ranked by severity, so you know exactly where to look first.
WEEK 3 — Show the Number
You get your false-alarm reduction rate and a ranked, cited list of what to check first — no obligation to continue.

Stop Learning About Faults From a Shortfall Report

See how per-asset fault detection and cited troubleshooting recommendations fit into your existing monitoring stack.

Get Help & Stay Updated