Anomaly Detection in Remotely Collected Drive and Motor Data

Abstract

Today more and more devices, both commercial and industrial, are connected to the internet. On the industry’s side, IIoT devices, such as AC Drives, transmitting data about their state and performance make novel analytics possible. The aim of this thesis is to detect anomalies in remotely collected drive and motor data, mainly coming from erroneous configuration during commissioning.

For that purpose, unsupervised anomaly detection methods are applied to a dataset provided by ABB, consisting of the setup parameters of a fleet of almost a thousand ABB’s ACS880 drives. Four methods were tried, namely: Isolation Forest, Local Outlier Factor, Hotelling’s T2, and K-means, and their performance in detecting manually created anomalies was evaluated.

The thesis concluded that errors in the setup of drives during commissioning can be detected by applying unsupervised anomaly detection on the main numerical setup parameters. However, the numerical error on the input has to be greater than 20 % for Current, Voltage, and 10% for Frequency to be detected, as found by examining the manually created anomalies. The rest of the errors in parameters were more insensitive to detection or inconsistent. Isolation Forest outperformed LOF and K-means and had similar detection capabilities as Hotelling’s T2 method, but with fewer false positives. Last but not least, as a by-product of the k-means anomaly detection method, the drives were separated in three clusters of motors with low, medium and high power.