What it does
Capabilities ofAgenticGrid Pro
Nine capability areas, in the order the work happens: take the evidence in, gate its quality, interpret it against standards context, draft the reasoning, and get it signed.
Intake
Evidence intake with provenance kept
Every record entering the workbench keeps the answer to "where did this come from?"
AgenticGrid Pro reads the files your team already produces: laboratory DGA and oil-quality reports, online monitor exports, PRPD and SFRA test-set files, radiometric thermal imagery and survey reports, field inspection forms and photos, nameplate data, work-order history, and loading extracts from a historian.
Each parsed record is pinned to an asset with its source system, file identity, timestamp and owner. Free text is preserved verbatim alongside the normalised fields, so nothing is silently rewritten on the way in. When a reviewer later asks which report a gas value came from, the answer is one step away, not a forensic exercise.
Quality gate
Correlation and the quality gate
Before any reasoning runs, signals are aligned per asset and screened. The gate checks units and detection limits, sampling-date consistency, duplicate lab reports, sensor-drift and calibration flags, SFRA baseline matching, emissivity and ambient context for thermal work, timestamp alignment on historian series, and asset identifiers against the master hierarchy.
Corrections are applied in the open and rejections are logged with the reason they failed. A clock drift on a thermal camera is recorded as a correction, not absorbed. The point is that the reasoning stage never sees a value nobody vouched for.
Evidence model
What each stream contributes
The workbench treats a diagnostic value as incomplete until its context arrives with it.
| Evidence stream | What the workbench reads | Quality gate |
|---|---|---|
| Laboratory DGA and oil quality | Gas concentrations and generation rates, ratios, moisture, acidity, furan context, laboratory identity | Units and detection limits, sampling-date consistency, duplicate-report reconciliation |
| Online gas and moisture monitors | Continuous channels, sensor status, alarm context, alignment to the latest laboratory benchmark | Drift and calibration flags; monitor values cross-checked against the most recent lab sample |
| Partial discharge (PRPD) | Pattern families, phase-reference quality, calibration and noise context, activity over time | Calibration and noise-floor notes required before a pattern is trended |
| SFRA | Frequency-response sweeps, per-band deviation against a named baseline | Each trace matched to an identified baseline before any deviation is scored |
| Thermal and loading | Top-oil and hot-spot estimates, cooling state, ambient context, loading profile, aging assumptions | Emissivity and ambient recorded; spot temperatures tied to the load at survey time |
| Inspection and maintenance history | Field notes and photos, leaks, bushing and tap-changer observations, work orders, nameplate context | Dated, asset-tagged entries; free text preserved verbatim with the record |
Interpretation support
Diagnostics the workbench is built around
Each of these is deterministic, local, and reported with its inputs and assumptions visible.
- DGA interpretation
- Multi-gas trends and generation rates screened against IEEE C57.104 and IEC 60599 context, with Duval Triangle placement and the basic gas ratios shown alongside the measured values they came from. Interpretation stays reviewable; the workbench does not issue a final DGA conclusion of its own.
- Partial discharge
- PRPD patterns tracked across tests to separate persistent defect activity from noise and load artefacts, with phase reference and noise floor carried as part of the finding.
- SFRA
- Sweeps overlaid on a named baseline with deviation bands highlighted by frequency range, so winding-movement questions are argued from the comparison rather than a single number.
- Thermal and loading
- Hot-spot and top-oil estimates placed in the context of the IEEE C57.91 loading guide, with the loading profile that produced them, so a chemical or electrical signal can be read against what the unit was actually doing.
- Failure-mode context
- CIGRE's international transformer reliability survey (Technical Brochure 642, WG A2.37) analysed 964 major failures over 167,459 transformer-years contributed by 56 utilities in 21 countries and found windings, tap changers and bushings to be the leading failure locations — which is why those are the components the evidence model is organised around.
Health index
A score that shows its drivers
AgenticGrid Pro computes a condition score per unit, but the score is never the deliverable on its own. Each one is presented with its driver attribution, trend movement, the weighting assumptions used, and an explicit note on uncertainty and missing evidence.
That matters at fleet level. Prioritising units for an outage window or a capital case is only defensible if a reviewer can say which evidence moved a unit up the list, and what would change if a missing test arrived. A score with no drivers behind it is a number nobody can defend in a review meeting.
Agent reasoning
Chain of evidence, and the signal that disagrees
The agentic layer drafts the condition assessment: what the evidence suggests, which records support it, what the residual uncertainty is, and what it recommends investigating or doing. Every assertion in the draft is bound to the records it rests on, so the draft can be read backwards.
Contradictions are surfaced deliberately. When one stream points to a developing fault and another does not support it, the draft flags the disagreement and says so in the summary rather than resolving it silently in favour of the more alarming reading. A confidence band accompanies each draft, and the draft marker stays attached until an engineer clears it.
The human gate
The engineer sign-off state machine
Four decisions, one of which must be recorded before a draft can become a work package.
-
Approve
The reviewer accepts the draft as written. The assessment is stamped with their identity and the timestamp, and becomes eligible for work-package assembly.
-
Edit
The reviewer changes the finding, the rationale or the recommendation. The original AI draft is retained alongside the edit, so the difference between what was proposed and what was approved stays visible.
-
Reject
The reviewer rejects the draft with a reason. The rejection and its reason are recorded; the draft does not proceed and does not quietly re-appear.
-
Escalate
The reviewer routes the case to a more senior engineer or a specialist, with the evidence chain and their own comment attached, when the call should not be theirs alone.
Audit log
What the log preserves
Entries are append-only. An edit adds to the record; it never rewrites it.
- Reviewer identity and role
- Each decision names the engineer who made it, with their role and any comment they left.
- State changes with timestamps
- Draft, edit, approval, rejection and escalation are each dated, so the review history reads as a sequence rather than a final state.
- Evidence and assumptions
- The source records, weighting assumptions and uncertainty notes stay attached to the finding they informed.
- Final decision state
- The approved, rejected or escalated outcome closes the record and travels with the exported report.
Outputs
Work packages and evidence-pack reports
Approved assessments become a maintenance work package: a prioritised task list with asset, priority and work-type fields, parts context, an outage-window note, and the evidence chain behind each line. Work packages export as CSV for CMMS or EAM import and as PDF for review, so no write access to your systems of record is needed.
The evidence pack is the fuller document — condition evidence, charts, the AI rationale, the reviewer comments and the complete sign-off trail — assembled so that a third party can follow every conclusion back to a source record. The specimen published on this site is built on labeled synthetic data, not a customer fleet.
The other GridAPM product
GridAPM also builds ProtectionAI
ProtectionAI
Protective-relay testing, settings management and reporting in one Windows workspace, with an agentic AI copilot that works to your approval.
Put your own evidence through the workflow
A bounded evaluation starts with the records you already have and ends with an evidence pack your reviewers can inspect line by line. Your units, your engineers at the gate.