Machine learning methods for contextual anomaly detection

Abstract

This dissertation proposes and explores several models for anomaly detection, which is frequently found in industrial fields. Among them, this dissertation focuses on contextual anomaly detection in case there are contextual variables as well as response variables directly used for anomaly detection. The contextual variables are not directly related to the system health, but they influence the response variables obtained from the system. These contextual variables violate the assumption that the response variables obtained in a normal state system follow a single distribution. In many industrial fields, the distribution of response variables varies due to these contextual variables. Therefore, it is important to detect anomalies in consideration of the contextual variables. In the first chapter of this dissertation, a contextual anomaly detection model using dimensional reduction and conditional density function estimation is proposed for high-dimensional data having complex relationship in response variables. In the second chapter, we propose a VAE-based contextual anomaly detection model that additionally considers the clustering effect of contextual variables for high-dimensional and complex data. Finally, in the third chapter, we propose a contextual anomaly detection model that reflects auto-correlation in data with multidimensional and complex relationships between variables. This dissertation aims to present the possibility of solving various anomaly detection problems occurring in the actual industrial fields by proposing contextual anomaly detection models based on machine learning for data with high-dimensional and complex relationships between variables.