Multimodal Knowledge Graph for Database Alarm based on Enhanced Fusion Strategy

Database alarms are a pivotal monitoring mechanism, designed to notify system administrators of potential database performance issues, security risks, or failures in a real-time context. Traditional database alarm systems typically depend on pre-established rules and thresholds, which curtails their efficacy when confronted with unknown or intricate anomalies. Moreover, these systems often neglect the multimodal characteristics of alarm data, such as textual logs and visual charts, further impairing the precision and promptness of the alarm. We put forward an innovative database alarm management model, ImageKnowledgeRetriever and the Advanced Textual Knowledge Retriever module are designed to extract the image and text information and design a fusion strategy to merge multiple modal characteristics. Our model is adept at comprehending and processing textual logs, as well as analyzing and interpreting image data, including performance charts and architectural diagrams. This method enables our model to conduct a more exhaustive analysis of alarms, thereby pinpointing and diagnosing database issues with heightened accuracy. Our model demonstrates superiority over existing database alarm systems regarding accuracy and F1 score. Specifically, our model achieves an average accuracy enhancement of 8.1% and surpasses the current leading model by 0.5% in the F1 score metric. These findings affirm the advanced capabilities and effectiveness of our model in managing database alarms.