@techreport{oai:ipsj.ixsq.nii.ac.jp:00234254,
 author = {久恒, 泰地 and 受田, 賢知 and Taichi, Hisatsune and Tomoharu, Ukeda},
 issue = {17},
 month = {May},
 note = {システム運用時の障害分析に対し生成 AI を活用した際の正答率向上手法を検討した．仕様書とユーザーガイドをナレッジベースへ格納した RAG に対し障害情報を入力し要因・対策を出力する従来手法を IoT データ収集基盤にて発生した障害事例に対し評価した所，要因・対策を明記した障害以外では正答率 25% と低い点が課題となった．本報告では，処理，対象，設定の複数知識を問うプロンプトを生成し，ナレッジベースから各知識を RAG により抽出し，GPT-4 の一般解と結合し分析する関連知識統合手法を提案した．IoT データ収集基盤の障害事例 20 件で回答精度を評価し，70% まで向上した結果を報告する．, We have studied methods to improve the accuracy of fault analysis during system operation by utilizing AI. We inputted fault information into a RAG that stored specifications and user guides in a knowledge base, and evaluated the conventional method of outputting causes and countermeasures against failure cases that occurred on the IoT data collection platform. The problem was that the accuracy rate was as low as 25% for faults other than those that clearly stated causes and countermeasures. In this report, we propose a method of integrating related knowledge that generates prompts that ask for multiple knowledge of processing, subject, and setting, extracts each knowledge from the knowledge base by RAG, and analyzes it in combination with the general solution of GPT-4. We evaluated the answer accuracy with 20 failure cases on the IoT data collection platform, and report that it improved up to 70%.},
 title = {障害要因分析における生成AIを活用した関連知識統合手法の提案},
 year = {2024}
}