LLMs link spare-part orders to installations in multilingual logs

Challenge

Determining when ordered spare parts are actually installed — and classifying the actions that precede a replacement — required analysing multilingual free-text maintenance logs with inconsistent terminology across five languages.

Solution

Maintenance logs are processed in Databricks through translation, classification and matching stages, linking spare-part orders to installations and enabling analysis of repair-before-replace patterns and lead times.

• Autonomous terminology learning. An LLM extracts worker terminology from logs where known part numbers appear, automatically building a multilingual vocabulary of parts and maintenance actions across all plants.
• Semantic classification. Each maintenance action is classified by the LLM to separate actual replacements from inspections, cleaning and other actions.
• Order matching. Every spare-part order is linked to its best candidate installation within a configurable time window, based on equipment and temporal proximity.

Results

• 5 languages processed automatically
• Order-to-installation lead times extracted
• Action taxonomy built from raw multilingual logs
• Cross-validated detections