Betreuer/in: (Al Sardy)
Smart Grid applications generate large volumes of heterogeneous log data that comes from the variety of components of a grid infrastructure, including field devices and the corresponding industrial communication protocols that they utilize to communicate with each other, monitoring devices and systems, etc. Efficient log analysis is necessary to maintain system’s continuous performance and reliability, detect anomalous behaviour in a timely and correct manner, and support cybersecurity of a smart grid. However, traditional log parsing approaches may not perform effectively in Smart Grid environments due to their inability to model domain-specific semantics and conditions.
Hence, improper processing of logs can lead to further degraded anomaly detection and system investigation. This thesis aims to address the problem of log parsing for Smart Grid systems by proposing the log parsing framework named HyParse-SG. The proposed system combines structural and contextual parsing approaches and employs selective large language model refinement. Initially HyParse-SG executes domain-specific preprocessing of log data. Structural parsing performs template extraction and contextual parsing captures sequential relationships between events on a system/device level. After parsing, an LLM is applied to correct structurally incorrect templates. Selectivity of reparation stems from the confidence mechanism which decides if LLM should be triggered to repair templates, reducing computational cost and preserves parsing effectiveness. I evaluate the proposed framework on the several Smart Grid-specific datasets. Experimental evaluation on these datasets demonstrated that HyParse-SG consistently achieved equal or higher parsing and grouping accuracy values and highest efficiency across all datasets compared to the baselines. The contextual analysis showed a good level of performance, and selective LLM-based refinement demonstrated a strong ability to correct degraded templates.
Raum 04.137, Martensstr. 3, Erlangen
oder
Zoom-Meeting beitreten:
https://fau.zoom-x.de/j/68350702053?pwd=UkF3aXY0QUdjeSsyR0tyRWtLQ0hYUT09
Meeting-ID: 683 5070 2053
Kenncode: 647333