<![CDATA[The growing complexity of enterprise network infrastructures has increased the importance of predictive maintenance for network security devices, particularly firewall systems. In operational environments using Fortinet firewalls, large volumes of firewall logs are continuously generated, while existing monitoring tools such as FortiAnalyzer remain limited to descriptive analysis and lack predictive capabilities. This study aims to evaluate the effectiveness of artificial intelligence, specifically Large Language Models (LLMs), for predictive maintenance through automated analysis of firewall logs. Four open-source LLMs-Gemma 2B, Mistral 7B, DeepSeek-R1 7B, and Qwen 2.5-Coder 7B-were benchmarked using a standardized Indonesian-language prompt designed to extract high-severity events, including emergency, alert, and critical conditions, from multi-severity Fortinet log data. The evaluation focused on AI benchmarking metrics such as severity filtering compliance, reasoning accuracy, linguistic consistency, structural clarity, and processing efficiency. The results indicate that Qwen 2.5-Coder 7B provides the most reliable overall performance, demonstrating strong adherence to severity constraints, consistent Indonesian-language output, and well-structured analytical results suitable for operational predictive maintenance. Mistral shows superior contextual reasoning but exhibits language inconsistency, while Gemma offers the fastest processing time with moderate severity accuracy. DeepSeek performs least effectively due to instruction non-compliance. This study addresses an existing research gap by demonstrating how large language models can support predictive maintenance for firewall-based network security systems and provides a comparative framework for future AI-driven firewall and log-analysis research.]]>
