Deep anomaly detection: Benchmarking machine learning models for multivariate time series

Detection of anomalous patterns in streams of sensor data is an important challenge in today’s industries. Previously unseen patterns can be a sign of misconfiguration, increasing mechanical wear-out or simply unforeseen situations which may eventually lead to failures. Identification of anomalous patterns can facilitate the proactive detection of failures, as modern industrial settings rely extensively on sensor data for monitoring mechanical or electrical behaviour of devices. Currently, this process is often manually handled by engineers with advanced knowledge of the machinery in the specific domain. Industrial systems, however, often record massive amounts of sensor data, which makes reliable manual detection of anomalous samples impossible and immensely time-consuming. Machine learning- based automated detection offers an attractive alternative to sensor anomaly detection. In recent years, deep learning algorithms have shown robust anomaly detection capabilities in complex domains. The primary reason for the success of these techniques lies in their ability to capture relevant information directly from raw data, with minimal need for manual feature engineering. However, they require a lot of data and training time, and do not always outperform shallower and less complex architectures. For this reason, it is important to understand in which conditions, and for which scenarios, deep learning models might bring an advantage.