Predictive Maintenance Workflow Guide

Use this guide to build an AI Agent workflow around the FactVerse AI Agent predictive maintenance module. The workflow combines equipment identity, signal history, anomaly context, maintenance records, and field feedback so engineers can review equipment health and plan the next inspection or maintenance step.

Prerequisites

Requirement	Details
Equipment model	Stable equipment ID, component hierarchy, location, operating mode, and criticality.
Signal history	Time-series signals, data quality flags, sampling interval, unit normalization, and known gaps.
Maintenance context	Inspection records, work orders, replaced parts, failure notes, and operator feedback.
Knowledge base	Manuals, standard operating procedures, troubleshooting notes, and site-specific constraints.
Review owner	Maintenance engineer or reliability owner who can accept, revise, or reject a recommendation.

Endpoint and scopes

Use /mcp/pdm/ for predictive maintenance health and anomaly tools. Use /mcp/base/ when the workflow also needs asset records, documents, work-order context, or approved action drafts.

Endpoint	Scope	Use
/mcp/pdm/	pdm.read	Read health summaries, anomaly context, component signals, and model outputs available to the key.
/mcp/base/	base.read	Read asset identity, documents, work records, and related operational evidence.
/mcp/base/	base.action.write	Draft a follow-up inspection or maintenance action after human approval.

For a detailed map of pdm.read tools and AI Engine predictive maintenance endpoints, see Predictive Maintenance AI Tools.

Tool and surface selection

Start with the user's operating question, then choose the narrowest tool or product surface that can answer it.

Question	Tool or surface	Expected handoff
Which assets need attention this shift?	get_pdm_summary, Predictive Maintenance dashboard, command-center view	Ranked fleet summary with health, alert, data-readiness, and open-action context.
Which anomaly should be reviewed first?	list_pdm_anomalies, advisory inbox	Anomaly list grouped by severity, status, source window, and current advisory state.
Why is this asset unhealthy?	get_equipment_health, equipment detail, health history, vibration and energy views	Health explanation with contributing signals, recent trend, baseline state, and missing data.
Is a recommendation ready for action?	Advisory inbox, work-order history, Inspector context through /mcp/base/	Accepted, rejected, or snoozed recommendation with reviewer note.
Can the model be trusted or retrained?	Model console, DFS quality, model assignment and readiness blockers	Model state, active assignment, data-gate blockers, and required data correction.
Are parts or components driving the issue?	Component intelligence and circular recovery tools when enabled	Component-level context, recovery outlook, and sourcing or repair notes.
Is more data preparation required?	DFS Lite quality, DFS Pro datasets, rejection queue, dataset lineage	Readiness report before any maintenance recommendation is made.

Predictive maintenance workflow flow

Workflow steps

1. Define the equipment set: choose the site, asset group, equipment ID, operating mode, and review window.
2. Check data readiness: confirm signal freshness, missing intervals, unit mapping, and known sensor issues.
3. Retrieve health context: use runtime-discovered tools to summarize health, anomalies, trend changes, and recent events.
4. Connect maintenance history: compare anomalies with inspections, part changes, alarms, and operator notes.
5. Prepare an engineering view: describe probable causes, confidence limits, evidence gaps, and suggested next checks.
6. Route a reviewed action: draft inspection, spare-part review, or planned maintenance only after the engineer accepts the recommendation.
7. Feed back outcomes: capture confirmed cause, false alarm, repaired component, and follow-up notes for future model improvement.

DFS setup for predictive maintenance data

Prepare signal history and maintenance context before relying on predictive maintenance output.

Data need	DFS workflow
Time-series signal history	Prepare Signal History for Predictive Maintenance
Connector and source configuration	Connector Types
Unit and identity mapping	Mapping Source Fields
Missing intervals and stale values	Data Quality
Work-order, inspection, and signal fusion	Fuse Inspection, Work-Order, and Sensor Data
Reusable reviewed dataset	Dataset Lifecycle
AI tool selection and review checks	Predictive Maintenance AI Tools

Create a readiness package before model training, anomaly review, or recurring Agent workflows:

Readiness item	Required evidence
Asset identity	Equipment ID, aliases, location, component hierarchy, owner, and criticality.
Signal definition	Sensor tag, axis or channel, unit, sampling interval, expected range, and operating-state meaning.
Data quality	Freshness, missing intervals, stale values, flatlined sensors, unit conversions, and source reset notes.
Operating context	Running, stopped, standby, partial-load, cleaning, batch, or maintenance state when available.
Maintenance outcome	Work-order close date, confirmed cause, replaced part, action taken, false-positive flag, and reviewer.
Model context	Template, model ID, model version, baseline date, assignment state, and retraining blocker if present.

Recommended output structure

Section	Content
Health summary	Equipment, current condition, trend direction, anomaly window, and severity.
Evidence	Signals, model outputs, maintenance records, inspection notes, and source timestamps.
Possible causes	Ranked engineering hypotheses with supporting and conflicting evidence.
Recommended checks	Field inspection items, operating checks, data checks, and parts to review.
Review result	Engineer decision, accepted action, rejected recommendation, or required data correction.
Feedback handoff	Outcome label, root cause, false-positive reason, model-readiness note, and follow-up data task.

Continuous improvement loop

Predictive maintenance workflows improve when completed work feeds back into the data layer. Keep the loop explicit:

• Confirm whether the anomaly matched a real equipment issue.
• Record replaced parts, adjusted settings, and inspection findings.
• Mark false positives and sensor-quality issues.
• Keep 24x7 signal ingestion visible so model training and rule tuning can use updated operating history.
• Review model output against engineer feedback before expanding automation.
• Use the model console to check whether data gates, stale baselines, fallback assignments, or missing model assignments are blocking model use.
• Update rule sets and templates only after the maintenance owner accepts the evidence and the change is recorded.

Common failure modes

Symptom	Likely cause	Response
Health output is weak	Signal history is short, sparse, or unmapped	Return a data-readiness report before making maintenance suggestions.
Anomaly appears without cause	Maintenance and operating context are missing	Pull work records, inspection notes, operating mode, and recent alarms.
Too many warnings	Query boundary is too broad or thresholds are not tuned for the asset	Narrow the equipment group and review thresholds with the engineer.
Recommendation is rejected	Confidence is low or evidence is incomplete	Keep the output as an engineering note and request the missing data.

Validation checklist

• The workflow uses pdm.read for predictive maintenance tools and writes user-facing output with readable terms such as predictive maintenance module.
• Each recommendation includes evidence, confidence limits, and missing-data notes.
• Engineer approval is recorded before creating an action.
• Completed maintenance outcomes are captured for model and knowledge improvement.