Exécution de notre outil de synthèse sur un nouvel article

Résumé : Modern IT infrastructures generate vast volumes of system logs essential for monitoring and failure diagnosis. While conventional index-based log anomaly detection methods achieve strong performance in controlled settings, they produce high false positive rates when log messages change due to software updates, a problem known as log template evolution. Generic semantic encoders miss domain-specific meaning and execution order, while LLM baselines are too slow for production. We propose a hybrid LogBERT-LSTM framework that combines a domain-adapted BERT encoder, refined through SimCSE contrastive training, with an LSTM that predicts the next log event in continuous semantic embedding space. Evaluated on HDFS and BGL under synthetic evolution, the proposed model maintains F1-scores of 0.8205 and 0.1841 respectively. The nearzero performance delta between static and evolved evaluations empirically validates that the framework effectively adapts to template evolution without false positive inflation.