UNLOCKING MAINTENANCE DARK DATA WITH EMBEDDING-BASED SEMANTIC SEARCH

Textual maintenance logs from mining operations provide information relevant to predictive maintenance, but Swedish free-text narratives make automated classification difficult. Existing methods are limited by domainspecific terminology and multilingual contexts. A hybrid semantic lexical search system is presented that maps freetext maintenance descriptions to a standardized taxonomy of 72 categories by combining contrastively fine-tuned multilingual embeddings (GTE) with TF-IDF lexical matching. The system uses weighted score fusion ( α = 0.90 ) to integrate semantic similarity with exact term evidence. Results on 919 real-world Swedish maintenance logs from mining drill rigs show that the hybrid approach achieves 45.9 % Recall@1, 75.4 % Recall@5, and 0.585 MRR on real-world data, substantially outperforming semantic-only (73.1 % → 45.9 % validation-to-real drop) and lexicalonly baselines (52.7 % R@1 on validation). Per-category analysis shows strong performance for fire safety systems (100 % R@1) while indicating persistent challenges for motor failures (0 % R@1).

Keywords: text embeddings, taxonomy classification, predictive maintenance.