MES Production Data for AI: Orders, Genealogy and Downtime Records from Real Plants

Real execution records teach what simulated plants cannot: late material, a changeover that ran 40 minutes over standard, a "minor stop" coded as "other" for three shifts. Trade-press and vendor commentary both point at contextualized, harmonized production data as a main bottleneck for manufacturing AI [3][4]. This page covers execution flow. Inspection and nonconformance data are covered on the manufacturing quality records page.

Which MES records are worth licensing

The records worth licensing are the ones that connect a plan to what actually happened on the line, with timestamps and identifiers that survive the export. A list of machine states alone is a sensor feed; an MES dataset adds the order, operation, material and reason context around each event.

• Production orders (work orders): order ID, material or part number, BOM and routing version, planned quantity, due date, release, start and close timestamps, order status history.
• Routing operations: operation sequence, work center, planned setup and run time, actual setup and run time, yield, confirmations by shift.
• Labor and machine bookings: clock-on and clock-off per operation, crew size, equipment ID, indirect time.
• Material consumption and genealogy: consumed lot or serial, produced lot or serial, quantity, parent-child links across operations.
• Scrap and rework: quantity, scrap code, operation where detected, rework routing.
• Downtime events: equipment ID, start and end, state (planned, unplanned, starved, blocked), reason code at each tree level, operator free-text comment, who coded it and when.

For regulated process plants, electronic batch records add structure: drug and food plants under GMP rules typically record equipment used, component lot identities, weights and in-process checks per batch. That makes these plants a rich source of genealogy, though trade-secret and confidentiality concerns are higher.

Mapping exports to ISA-95 objects

Describe your request in ISA-95 terms so suppliers on different MES platforms can map their tables to it. IEC 62264, published in the US as ANSI/ISA-95, models Level 3 manufacturing operations and its interface with Level 4 enterprise systems such as ERP, which gives buyers and suppliers a neutral vocabulary.

In practice, ask for production requests and schedules (orders), segment responses (operation confirmations), material lots and sublots (genealogy), and the equipment hierarchy (enterprise, site, area, work center, work unit). Exports from commercial MES platforms and in-house systems differ widely in table names and grain, so the data dictionary is not optional. If you plan process-mining or agent work, ask whether events can be delivered as an object-centric event log; OCEL 2.0 defines relational (SQLite), XML and JSON formats where one event can reference several objects such as an order, a lot and a machine [1].

Downtime reason codes and why their history matters

Downtime reason codes are only usable for OEE loss classification if you also get the code tree, its change history and the free-text comments. Plants restructure trees after continuous-improvement projects, merge codes, or let operators default to "other"; without the history, a model learns the reorganization rather than the loss.

Ask the supplier for:

1. The reason-code tree per plant with effective-from and effective-to dates for each code.
2. Mapping notes where codes were merged or renamed.
3. Share of unplanned downtime minutes coded "other", "unknown" or left blank, by year.
4. The rule for micro-stops (for example, stops under a threshold auto-coded or dropped).
5. Operator comments in raw form, with language and any shorthand glossary.

Free text is often where the real cause sits ("waiting on forklift", "label roll changed early"). That makes these records useful for weak-supervision relabeling and for evaluating an LLM that proposes reason codes, as described in outcome-labeled evaluation data.

Linking plan, execution and outcome with shared keys

Shared keys turn MES records from logs into training signal. Order IDs that match ERP orders tie execution to demand and due dates; lot and serial IDs that match quality records tie execution to defects; equipment IDs and time ranges that match historian tags tie stops to sensor behavior.

Before licensing, test join rates on a sample: the share of MES orders with an ERP match, the share of nonconformances with a resolvable lot, and clock alignment between MES and historian (time zone, daylight saving, server drift). If you need sensor context, scope it separately through sensor and IoT data; inbound and outbound flows belong with supply chain and logistics datasets. Assess quality with a shared vocabulary; ISO/IEC 5259-1 gives terminology for assessing data quality for analytics and ML across the data life cycle [2].

Use cases and what each needs from the data

Each use case stresses a different part of the record, so scope fields by use case rather than asking for "everything".

Illustrative example: invented to show structure; it does not describe an available dataset.

For the text-to-SQL row, see text-to-SQL training data from real enterprise schemas. If you build a held-out test set from plant history, the method in golden evaluation datasets from business records applies.

Request template for an MES dataset

A good request names grain, history, systems and restrictions so a supplier can say yes or no quickly.

Illustrative example: invented to show structure; it does not describe an available dataset.

request: mes_execution_records
plant_profile: discrete assembly or packaging, US, 2+ lines
history: 3+ years, including at least one reason-code tree revision
grain: one row per operation confirmation and per downtime event
objects: [production_order, operation, material_lot, equipment, downtime_event]
keys_required: [order_id -> ERP order, lot_id -> quality record, equipment_id -> historian tag]
text_fields: [downtime_comment, rework_note]  # raw, with glossary
pseudonymize: [customer_name, customer_part_number, operator_id, supplier_lot_vendor]
restricted_fields: [recipe_parameters, ideal_cycle_time]  # field-of-use limits
format: Parquet or SQLite; OCEL 2.0 export optional
documentation: data dictionary, reason-code tree with effective dates, shift calendar
intended_use: OEE loss classifier training and evaluation

Privacy, trade secrets and license scope

MES data carries less personal data than support logs, but operator IDs, badge numbers and comments can identify people, and customer names and customer part numbers can expose commercial relationships. Pseudonymize them consistently so an operator or customer stays joinable across tables without being named; check free-text comments, where names tend to leak. Broader methods are covered in the de-identified data guide.

Recipes, setpoints, ideal cycle times and yield by product can be trade secrets. Expect suppliers to ask for field-of-use limits, aggregation, or withholding of specific products, and decide early whether your model truly needs product-level parameters. For delivery, read-only warehouse shares, such as Snowflake Secure Data Sharing, let a supplier grant access without copying data between accounts [5].

How SourceX approaches MES data requests

SourceX sources operational datasets from US companies on request and manages the commercial process through licensing and ongoing purchases; it does not hold MES data in stock, and a request does not guarantee a match. You describe the records you need, not the plants, and every release is approved by the supplying company. Datasets are rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery, through private, access-controlled workflows after an executed agreement. Manufacturing buyers can review the manufacturing buyer overview, browse other industry-specific operational data, or start a request at SourceX for buyers.

Request MES production records for your model

Describe the orders, genealogy and downtime records you need, the history and the intended use. SourceX will look for US businesses that hold matching data and assess data and licensing permissions before anything is agreed. Start at https://sourcex.si/buyers.