Study on Anomaly Detection Algorithm Using Stochastic Model with Periodically Updating Activity Sound of Living Alone

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K02145
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 08030:Family and consumer sciences, and culture and living-related
• Research Institution: Akita University
• Principal Investigator: Tanaka Motoshi 秋田大学, 理工学研究科, 准教授 (50261649)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 生活活動音 / 時間-周波数解析 / 異常検出 / 機械学習 / 更新

Outline of Final Research Achievements

In order to develop a detection system of abnormal situations (such as accidents) for a person living alone, an anomaly detection method for periodically updating a stochastic model of daily (activity) sounds and by observing the occurrence probability of the sound was investigated in accordance with changes in the living environment. In this method, additional daily sounds were periodically imported, the cluster of the daily sounds were updated, and the parameters of the stochastic model were recalculated accordingly. When the daily sounds changed significantly, all feature vectors in some clusters might be lost during the update. Therefore, ways to deal with this were also considered. The results indicated the feasibility of continuously updating the stochastic model of daily life sounds corresponding to the living environment changes.

Free Research Field

情報通信工学

Academic Significance and Societal Importance of the Research Achievements

家屋内の日常生活音から事故などの異常検出するために作成した確率モデルについて，定期的に生活音の特徴量を入れ替えて再学習し，確率モデルのパラメータを更新していく方法を提案した。これにより，利用状況や生活環境の変化に対応して，定期的な学習内容の更新が可能になったと考えられる。また，家庭内事故等の検出システムのみならず，経年変化などに対応が必要な装置・システムの故障診断など他の異常音検出への応用も期待できる。