A Method to Handle Unstable Time Series in Anomaly Detection Problem
Andrey Gaev, Anna V. Lantsberg · 2021 3rd International Conference on Control Systems, Mathematical Modeling, Automation and Energy Efficiency (SUMMA) · 2021
Time series analysis is a well-known problem solved in many spheres of human activity. This analysis is most widely used in the tasks of tracking the state of any processes - checking stability of working machines in factories, tracking the rhythm of the heartbeat, and so on. One of the important tasks is decrease time to identify deviations in these processes, which reduces the cost of their elimination. A well known name for such tasks is anomaly detection and they belong to one-class classification problem. It is known that in order to effectively solve this type of problem, the spread of data within the training dataset should be taken into account. In this paper, we present a method that allows us to reduce intra-class data variation by automatically searching for and excluding unstable parts from the time series. The application of this method is considered as part of an algorithm that can be used to represent raw samples of time series as vectors suitable for further work with machine learning models. The algorithm has been tested on boundary-based machine learning methods such as One-class Support Vector Machine (OCSVM), Isolation Forest and Local Outlier Factor (LOF). Our dataset consists of time series from gas sensor, which are reflect sensor‘s sensitive surface reaction to several gases appearing in the environment.