Skip to main content
MAN-01 · Manufacturing

Maintenance Troubleshooting Copilot

Outcome​

Get a maintenance technician from a fault code to a verified repair by combining equipment history, sensor evidence, manuals, and prior successful fixes, with supervisor authorization before any safety-critical intervention.

Portfolio attributeValue
Primary personaMaintenance technician
Primary aggregateTroubleshootingSession
Cosmos DB containers2
Grounding datamanuals, work orders, sensor events, fault codes
Development estimateUSD 60-220 monthly dev estimate
Runtime optionsPython FastAPI and C# ASP.NET Core

Productivity hypothesis​

These measures reflect the workflow, evidence, decision boundary, and industry risk for this use case.

Impact typeMeasureConservative internal assumptionCustomer measurement
ProductivityFault evidence discovery time20-35% reduction(baseline median fault-evidence search minutes - assisted median minutes) / baseline median minutes * 100
Quality or riskRepeated troubleshooting8-15% reduction(baseline repeated troubleshooting-step rate - assisted rate) / baseline rate * 100

Important: Conservative internal pilot assumptions only; not public benchmarks, commitments, or expected customer results. Validate against a comparable baseline or matched control cohort with mandatory human review.

Rationale: PLC alarms, historian trends, OEM procedures, spares, and CMMS history support faster fault isolation.

Pilot guardrail: Require verified equipment state, evidence, safety procedures, and technician approval before action.

Customer validation plan​

  1. Baseline the current workflow with representative users and tasks.
  2. Run the accelerator on the same task mix with human review retained.
  3. Measure fault evidence discovery time and repeated troubleshooting with the documented formulas.
  4. Replace internal assumptions with customer-observed results before a business case.

Why Azure Cosmos DB​

This AI solution needs an operational data plane, not only a vector index:

  • Hybrid vector and full-text retrieval combines semantic recall with exact identifiers, codes, names, and domain language.
  • Hierarchical partition keys and metadata predicates keep tenant and domain boundaries in the normal query path.
  • Operational JSON, source chunks, workflow and agent state, feedback, and audit records remain together for low-latency reads and governed writes.
  • Change feed reacts to source and operational updates without polling.
  • Autoscale throughput, configurable consistency, and multi-region distribution support variable demand and enterprise availability.

Domain workflow​

Domain workflow

StageActorInput → outputApproval
1. Report the fault at the machineMaintenance technicianFault code, symptom description, and asset tag scan → TroubleshootingSession bound to the assetPolicy controlled
2. Correlate sensor and fault evidenceDiagnostic agentRecent fault events and process signals → Evidence timeline for the failure windowPolicy controlled
3. Rank causes from history and manualsHypothesis agentEvidence timeline, repair history, procedures → Ranked root-cause hypotheses with supporting fixesPolicy controlled
4. Guide the repair step by stepMaintenance copilotSelected hypothesis and applicable procedure → Cited step sequence with tools and safety precautionsPolicy controlled
5. Authorize safety-critical interventionMaintenance supervisorProposed lockout, bypass, or parameter change → Authorized work with a signed permit recordHuman required

Technical architecture​

Technical architecture

Cosmos DB data model​

ContainerHierarchical partition keyDocumentsSearch paths
troubleshootingSessions/tenantId, /assetIdTroubleshootingSession, AssetProfile, FaultEvent, RepairHistoryRecord/symptomVectorEmbedding, /rootCauseNarrativeVectorEmbedding, /symptomVector, /rootCauseNarrativeVector
maintenanceProcedureLibrary/tenantId, /procedureIdMaintenanceProcedure/stepVectorEmbedding, /stepVector

Query contract​

QueryContainerScopePurpose
Asset diagnostic contexttroubleshootingSessionssingle-partitionLoad asset profile, recent faults, and repair history for the machine in one partition read.
Recurring fault patterntroubleshootingSessionssingle-partitionDetect the same fault code repeating on the asset within the reliability window.
Model-applicable procedure lookupmaintenanceProcedureLibrarycross-partition-boundedKnowledge-container query against maintenanceProcedureLibrary (bounded fan-out across the /tenantId + /procedureId partitions named by the shortlist): retrieve only the manual steps valid for this equipment model, over the procedure ids that the asset profile and fault signature shortlist.
Repeat-failure repairstroubleshootingSessionssingle-partitionFind past repairs on the asset where the fix did not hold, to avoid repeating them.

Python and C# services route each document type to its declared container and hierarchical partition key. ETag-based writes protect mutable aggregates, while append-only records and bounded cross-partition queries preserve scale.

Azure deployment​

Azure deployment

Clone the source repository, authenticate with Azure, and run:

Set-Location accelerators/manufacturing/maintenance-troubleshooting-copilot
az login
./Deploy.ps1 -Runtime python -Location eastus2

Change python to csharp to deploy the .NET implementation.

Explore the implementation​

Return to all accelerators