Extended Framework and Evaluation for Multivariate Streaming Anomaly Detection with Machine Learning

Streaming anomaly detection in multivariate time series is an important problem relevant for automatic monitoring of various devices. This paper tackles the problem of streaming anomaly detection by extending a framework for the purpose of incorporating model-based approaches and evaluating previously uncombined methods for a total number of 26 distinct machine-learning-based algorithms. The framework identifies four fundamental components inherent to many streaming anomaly detection algorithms and one or more methods are presented for each component. It is found that a simple and computationally less expensive strategy for detecting concept drift yields almost identical results to the "KSWIN" strategy, when applied to measuring concept drift in a training set relevant for training a machine learning model. A secondary experiment supports the effectiveness of finetuning a machine learning model after the detection of concept drift for the purpose of detecting anomalies.