A Fine-Tuned BERT-Based Model for Individual Log Anomaly Detection in Operational Monitoring at Paranal Observatory

Abstract

In operational environments such as astronomical observatories, continuous monitoring of system logs is critical yet challenging due to the vast volume of data generated. Manual inspection of these logs is impractical, needing automated methods capable of accurate and scalable anomaly detection. This paper proposes a novel sentiment-aware anomaly detection framework, leveraging Bidirectional Encoder Representations from Transformers with In-Task Pre-Training and Fine-Tuning (BERT-ITPT-FiT), to classify textual logs based on their inherent sentiment polarity. Specifically tailored for the Paranal Observatory, our approach incorporates a computationally efficient, character-based preprocessing strategy that retains essential semantic elements, such as acronyms and technical terms, thereby eliminating the need for traditional parsing. This method efficiently processes a substantial dataset comprising 7,359,506 log entries from three VLTI instruments in an average runtime of only 114.23 seconds. Extensive experiments utilizing real operational data demonstrate that our architecture achieves an effective throughput of approximately 3,958 logs per second during the combined embedding generation and model training phases while consistently delivering an excellent classification performance. Besides, it significantly outperforms existing deep learning methods combined with conventional embeddings in both in-and cross-instrument scenarios, achieving F1-Scores exceeding 99.99% and 99.96%, respectively. This work emphasizes a critical balance between computational efficiency and classification performance, providing a robust and scalable solution for anomaly detection in high-stakes observatory operations.