Schema-Agnostic Feature Extractor Using LLM and Transformer for RUL Prediction

RUL (Remaining Useful Life) prediction is an essential technique for preventing equipment failures, reducing maintenance costs, and improving overall operational efficiency. Most existing RUL prediction methods assume that all samples in the dataset share a common schema, i.e., a common set of variables. However, in practical industrial environments, datasets with multiple schemas often exist due to differences in data collection systems and component diversity. To address this issue, we propose the Schema-Agnostic Feature Extractor using LLM and Transformer (SAFELaT), a feature extraction framework applicable to datasets with different schemas, and apply it for RUL prediction. SAFELaT utilizes variable semantic embeddings of brief descriptions derived from variable names via Large Language Models (LLMs), and aggregates variable tokens with a Transformer-based feature aggregator. LLM-based Enrichment enables the generation of rich and accurate semantic embeddings even when only short variable names are available, while Variable Dropout on token aggregation enhances robustness to multi-schema variations during training. We evaluated the effectiveness of SAFELaT on the simulated turbofan engine dataset and a real-world server dataset. SAFELaT outperformed prior methods under both single-schema and multi-schema settings, and variant analysis confirmed the benefits of each component in SAFELaT.