Research on Web System Anomaly Detection and Intelligent Operations Based on Log Modeling and Self-Supervised Learning

Traditional log alerting systems suffer from high false positive rates and delayed anomaly diagnosis. This paper proposes an intelligent log analysis framework integrating self-supervised temporal modeling with vectorized semantic retrieval. The system constructs a log collection pipeline using the ELK Stack, employs BERT-derived models for semantic encoding of log fragments, and utilizes a Temporal Contrastive Learning module to capture cross-temporal anomaly patterns. By integrating Cluster-based Outlier Detection and an Attention-based visualization mechanism, it enables interpretable diagnosis of complex system behaviors. Experiments conducted on a production dataset of 120 million logs achieved a 14.7% improvement in F1 score, reduced detection latency by 48%, and attained an average alert accuracy of 92.3%. This framework significantly enhances the intelligent operations and maintenance capabilities of full-stack systems in AIOps environments.