AI System for Matching Process Deviations with Production Events

The problem

When process-series data, alarms, lots, recipes and event records are held in separate sources, their fields must be matched by hand before time-matched explanation candidates can be prepared.

The process engineer must compare every time-matched explanation candidate with its source record; missing or conflicting evidence must not replace the final decision.

If several simultaneous process changes are not set aside for review, a correlation could be treated as a root cause; the record must not proceed without a human decision.

What we set out to improve

• Time taken for draft time-matched explanation candidates and their supporting evidence to reach the process engineer for review
• How records set aside for human review because several process changes occurred at the same time are resolved
• Number of suggestions corrected after identifying a risk that a correlation could be treated as a root cause

How the system works

1. 01 Retrieve process-series data, alarms, lots, recipes and event records from permitted sources using the record identifier
2. 02 Link source, date and version information to the shared work record
3. 03 Validate fields and units
4. 04 Prepare time-matched explanation candidates with their supporting evidence
5. 05 Check for an exception: several process changes occurred at the same time
6. 06 Have the process engineer accept, revise or reject the suggestion
7. 07 Record the approved decision and its source in the processing log

Methods we used

• A source-identified common data schema for process-series data, alarms, lots, recipes and event records
• A decision structure that separates deterministic rules from AI suggestions
• A review screen showing time-matched explanation candidates alongside the underlying source evidence
• A mandatory human review queue when several process changes occurred at the same time
• An event identifier that prevents duplicate processing

Where people stay involved

Data and security

Only the fields required for this decision should be processed from process-series data, alarms, lots, recipes and event records. Access to the source systems must be authorised separately from the process engineer role, and personal data, commercial information and technical secrets must not be passed into the model context unnecessarily. Reading, suggestions and human decisions must be logged separately.

Who this suits

Frequently asked questions