| ABRAHAM OLUFEMI ABIODUN | D, 中間発表 | サイバーレジリエンス構成学 | 門林 雄基, 安本 慶一, 林 優一, 妙中 雄三 |
|
title: Unauthorized Power Usage Detection Using Gradient Boosting Classifier in Disaggregated Smart Meter Home Network.
abstract: For the smart grid unauthorized power usage identification in a disaggregated smart meter network, this paper introduces different classification algorithms to detect anomalies in appliance consumption patterns. They are Logistic Regression(LG), Support Vector Machine (SVM), Decision Tree Classifier(DTC), Random Forest(RF), Ridge Regression Classi- fier(RRC) and Gradient Boosting Classifier(GBC). While most of the existing Machine Learning (ML) algorithms focuses on aggregated data in smart meter network or Advanced Metering Infrastructure(AMI). Our ML algorithms are based on load dis- aggregation, otherwise known as Non-Intrusive Load Appliance Monitoring (NIALM). This technique generates appliance-level power consumption data based on single, smart meter readings to improve detection performance and time complexity. Simulation results for preprocessed Almanac of Minutely Power dataset version 2 (AMPds2) demonstrate the reasonable improvement of the proposed machine learning model. Emphasis has been laid upon the practical application of the proposed ML to detect unauthorized power usage in AMI. GBC improves both detection and False positive rates (FPR) compared with other classifiers. language of the presentation: English | |||
| 松井 智一 | D, 中間発表 | ユビキタスコンピューティングシステム | 安本 慶一, 中村 哲, 諏訪 博彦, 藤本 まなと |
|
title: Multi-person Daily Activity Recognition using Non-contact Sensors based on Activity Co-occurrence
abstract: Recognition of daily activities is important for health care, monitoring of the elderly, and offering highly useful services. Conventional activity recognition is generally based on the assumption that the activity is inferred based on the identification of the action owner. Therefore, few sensing and action recognition method has been proposed for multiple residents based on non-contact sensor data, which is difficult to identify individuals. However, the activity recognition of an entire household without individual identification is useful for measuring activities of each household and for outing support and recommendation services, etc. In this paper, we propose a non-contact sensing system and a method of activity recognition of multiple residents. In this presentation, we propose a method for recognizing activities in the entire household without personal identification based on data obtained by non-contact sensors. In a multi-resident household, there are cases in which one resident performs an activity for another resident, such as cooking and eating, or performs an activity at the same time, such as sleeping. The proposed method focuses on the co-occurrence of the activities of each resident as described above, and inputs the recognition results of the activities of each resident inferred from sensor data by a deep learning model to an individual machine recognition model (random forest) to recognize what activities are performed in the entire household. To evaluate the proposed method, we adapted it to one month of natural data from five homes. The results showed that the proposed method improved the recognition accuracy by about 5% from the baseline. language of the presentation: Japanese 発表題目: 非接触センサデータを用いた行動の共起性に基づく複数居住者の行動認識手法の提案 発表概要: 日常生活行動の認識はヘルスケアや高齢者の見守り，高機能サービスの提供のために重要である．従来の行動認識は，一般的に動作主を識別した上で行動の推論を行うことを前提としている．したがって，個人識別が難しい非接触センサデータに基づく，複数居住者を対象としたセンシング・行動認識手法は提案されていない．しかし，個人識別を伴わない家庭全体での行動認識は，家庭ごとの活動計測や外出支援・推薦サービス等に活用できるため有用である． そこで，本発表では，非接触センサによってセンシングされたデータに基づく，個人識別を伴わない家庭全体での行動認識手法を提案する．複数居住者の家庭では，料理・食事行動のように，ある居住者が他の居住者のために行動を実施する場合や，睡眠のように同時に行動を行う場合が存在する．提案手法では，上記のような各居住者の行動の共起性に着目し，深層学習モデルによってセンサデータから推論した各居住者の行動認識結果を，個別の機械認識モデル（ランダムフォレスト）へ入力し，家庭全体でどのような行動が行われているのかを認識する． 提案手法を評価するために，5軒の一般家庭から得られた1ヶ月間の自然なデータに対して適応した．その結果，提案手法によってベースラインから5％程度の行動認識精度の改善が見られた． | |||