Manufacturing
MES data for AI: what to export from your manufacturing execution system
By SourceX Editorial · Updated
Short answer
For AI, export the MES records that tie production events to outcomes: work order and lot genealogy, operation history, downtime with reasons, quality results, inspection images and rework. An MES data export built on those links is useful to developers, while machine tags without lot or quality context are not. Keep the IDs that join MES to ERP.
Key takeaways
- Genealogy, operation history and quality results joined by work order or serial number are the core of a useful MES export.
- Quality-control content, such as labeled defect images, CMM results and SPC records, can be among the most useful parts of an export.
- AI-ready for a licensing buyer means stable pseudonymous IDs, documented codes and no customer-owned designs, which is stricter than internal use.
- Export from a replica or reporting database, test a sample period first, and ship a manifest with every extract.
- Operator personal data, customer recipes and export-controlled programs are removed or scoped out.
What should you export from an MES first?#
The MES records to export first are the ones that connect a unit or lot to how it was made and how it turned out. Genealogy, operation history and quality results joined by the same work order or serial number give AI developers cause and effect. Machine tags alone give them signals without meaning.
An MES sits between the ERP, which knows orders and costs, and the machines and historians, which know signals. That middle position is what makes its records worth exporting, as long as the identifiers that link up to the ERP and down to the equipment survive the export intact.
MES record types, value and export method#
MES record types vary by product and vendor, but most systems hold some version of the families below, whether the MES is a packaged product such as Siemens Opcenter, AVEVA MES or Plex, a shop floor module inside the ERP, or a homegrown build. Use the table to decide what to inventory first and how each family usually leaves the system.
| Record type | What it captures | Value to AI developers | Typical export method |
|---|---|---|---|
| Work order and lot genealogy | Materials, components, machines and tools behind each lot or serial | Traces defects back to causes | Database query or genealogy report |
| Operation and routing history | Start, stop, station, status and operator role by operation | Cycle time, flow and scheduling models | Database query or vendor reporting tool |
| Downtime and state events | Machine states with reason codes and comments | Maintenance prediction and loss analysis | Event table export joined to maintenance work orders |
| Quality results | Inspection and test data, measurements, pass or fail | Yield and defect prediction | Database query or QMS integration |
| Nonconformance and rework | Defect codes, dispositions, rework steps, retest results | Defect reasoning and repair guidance | Export with links to NCR and CAPA records |
| Inspection images | Vision or manual inspection photos with decisions | Training data for visual inspection models | File export with an index linking image, unit and result |
| Electronic work instructions | Instruction versions, sign-offs and deviations | Procedure and deviation models | Document export with version history |
| Material consumption and scrap | Quantities, lots and scrap reasons | Yield and material planning | Database query |
Quality-control data is not the same as data quality#
Quality-control data and data quality are different things, and confusing them leads to weak exports. Data quality describes whether records are complete and consistent. Quality-control data is the inspection, test and defect content a plant generates, and it can be among the most useful parts of an MES export.
Vision-inspection model builders in particular look for labeled defect images: photos or scans with the defect type, location and the human or machine decision attached. A defect library that a quality team curated over years, with consistent categories, can be more useful than a much larger pile of unlabeled images.
CMM reports and SPC records add measurement context. Keep the characteristic, nominal, tolerance, measured value and the decision made, and record which drawing revision applied, without exporting the customer drawing itself.
What AI-ready means for a licensing buyer#
AI-ready for a licensing buyer means a static, documented, de-identified extract whose scope has been cleared, which is a stricter bar than readiness for internal analytics. The differences show up in identifiers, people, customers and documentation.
| Question | Internal analytics | External licensing |
|---|---|---|
| Identifiers | Internal IDs used as they are | Stable pseudonymous IDs that still join across tables |
| People | Operator names often visible | Names and badge IDs removed or replaced with role labels |
| Customers | Customer names and part numbers used freely | Customer identity and customer-owned designs removed |
| Documentation | Tribal knowledge fills gaps | Data dictionary, code tables and change history supplied |
| Scope | Whatever helps today | Set by a rights review and supplier approval |
| Format | Live dashboards and reports | Versioned extracts with a manifest |
How to run the export without disrupting production#
Running the export safely means working from a copy, proving the joins on a sample and documenting everything that leaves. The steps below follow a practical order for a combined IT and OT team.
- Inventory MES tables or objects, the years they cover and their links to ERP, QMS, historian and maintenance systems.
- Export reason code and defect code tables, including how they changed over time.
- Query a replica, a restored backup or a reporting database, never the live production database during shifts.
- Keep work order, lot, serial and part identifiers, and record how each maps to the ERP.
- Normalize time stamps and time zones, and note clock changes and daylight-saving shifts.
- Export a sample period first and check that joins hold before pulling full history.
- Write a manifest describing every file, field and code table.
What to leave out of an MES export#
An MES export usually leaves out three kinds of content: personal data about operators, customer-owned technical content and anything under export controls. Operator logins, badge numbers, training records and time stamps tied to named people are removed or pseudonymized.
Customer drawings, recipes, test programs and process parameters that a customer specified may belong to that customer. Records from export-controlled programs, such as ITAR defense work, are excluded altogether. When in doubt, scope a line or product family out rather than redacting record by record.
Free-text comments need their own pass. Operators and supervisors often type names, phone numbers or customer references into downtime and rework comments, so scan those fields for personal and customer details rather than assuming coded fields are the only risk.
Illustrative: a controls maker exports quality-linked MES history#
Illustrative: a fictional maker of industrial pump controllers runs an MES on its assembly lines, a separate vision system for solder joint and connector inspection, and an ERP for orders. The IT/OT director starts with an inventory and finds that serial numbers link MES operations, test results, inspection images and rework records.
The team exports a sample period from a reporting replica: genealogy, operation history, test results, the defect library with an image index and rework steps, with operator IDs replaced by role labels. Products built to customer designs are scoped out. Reason code tables changed during an MES upgrade, so the team exports both versions with a mapping.
The sample shows that joins hold across systems, so the full history export follows the same pattern, staged on company-owned storage.
How SourceX approaches MES data#
SourceX starts an MES review with metadata, not exports. The fit check asks which MES you run, which records it holds, how far back they go and how they link to ERP and quality systems. In the SourceX five-step transaction, Supply and Rights come before any Preparation work, followed by Approval and Delivery.
Large MES histories and image libraries stay in the company's own storage or ship on encrypted drives, and the SourceX Evidence Packet records provenance, licensing rights, permitted use, the privacy record and release authorization for everything delivered.
Frequently asked questions
Do we need historian data along with MES records?
Not always. Historian data adds detail when it can be aligned to work orders or lots and joined to outcomes. If that alignment is hard, an MES package with genealogy and quality results can stand on its own, and historian data can be added later for specific lines.
Our MES is a custom build. Does that matter?
Custom and homegrown MES systems can hold excellent records. The extra work is documentation: what each table means, how codes are defined and how the system changed over time. A clear data dictionary matters more than the brand of the system.
Should inspection images be raw or annotated?
Both can be useful. Raw images with the pass or fail decision and defect type attached are the minimum. Existing annotations, such as defect boxes drawn by inspectors, add value when their meaning is documented. Avoid images that show customer-proprietary designs unless rights are confirmed.
How do we handle MES data from a plant we acquired?
Treat it as a separate source with its own rights review. Acquired plants may run a different MES, different codes and customer agreements signed before the acquisition. Check what the purchase agreement transferred and whether any transition services or vendor terms limit exports.
Can we export MES data while still under a vendor support contract?
Usually yes, since the records are your own, but check the MES license and any hosting or support terms for limits on database access, bulk extraction or use of vendor-provided reports. Cloud-hosted MES products may route exports through the vendor's own tools or APIs.
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.