(A) study on data measurement interval control and data imputation methods for rapid detection of abnormal situations in IoT environments

Abstract

This dissertation presents a comprehensive study on data measurement interval control and data imputation methods for the rapid detection of abnormal situations in IoT environments. The study initially introduces a novel deep reinforcement learning-based adaptive data measurement interval control method specifically designed to detect abnormal situations and enhance the energy efficiency of IoT devices. By leveraging three essential features extracted from the measured data, namely the trend feature, dynamics feature, and data outlier feature, the proposed method achieves efficient identification of abnormal occurrences while conserving energy. Through the integration of two distinct reward functions, namely the dynamics similarity reward function and energy reduction reward function, the method demonstrates superior response time for notifying abnormal situations, outperforming existing techniques. Additionally, the proposed methodology showcases exceptional performance in terms of energy consumption, yielding significant data volume savings during normal situations. This research contributes to the advancement of IoT-based monitoring by providing improved energy efficiency and timely abnormal situation detection, resulting in enhanced operational effectiveness and resource utilization. Furthermore, this study presents a pattern similarity-based missing data imputation method for large-scale data measurement in IoT-based applications, addressing the challenge of restoring missing data caused by factors such as battery depletion or communication failure. The proposed method encompasses a sophisticated framework that integrates several essential components, including the automatic selection of the k-nearest patterns, an advanced approach for incorporating temporal relations through data positioning optimization, and a meticulously designed pattern score function for accurate imputation. Notably, the pattern score function, formulated based on two RMSE calculations, enables the selection of the most appropriate pattern for imputation, resulting in remarkable accuracy. Extensive experimentation and evaluation demonstrate the superiority of the proposed method, surpassing benchmark methods with substantial reductions in RMSE across diverse datasets, particularly at the largest missing intervals. Additionally, the proposed method offers notable advantages such as exceptional computational efficiency and seamless scalability to handle large-scale datasets, addressing the critical need for accurate imputations in IoT applications with data loss challenges. This research significantly contributes to the development of robust data analysis techniques, ensuring the integrity and reliability of IoT-based monitoring applications.