Unsupervised Machine Learning for Anomaly Detection in Multivariate Time Series Data of a Rotating Machine from an Oil and Gas Platform

Abstract

Deep Learning (DP) models have been successfully applied to detect and predict failures in rotating machines. However, these models are often based on the supervised learning paradigm and require annotated data with operational status labels (e.g. normal or failure). Furthermore, machine measurement data is not commonly labeled by industry because of the manual and specialized effort that they require. In situations where labels are nonexistent or cannot be developed, unsupervised machine learning has been successfully applied for pattern recognition in large and multivariate datasets. Thus, instead of experts labeling a large amount of structured and/or non-structured data instances (also referred to as Big Data), unsupervised machine learning allows the annotation of the dataset from the few underlying interesting patterns detected. Therefore, we evaluate the performance of six unsupervised learning algorithms for the identification of anomalous patterns from a turbogenerator installed and operating in an oil and gas platform. The algorithms were C-AMDATS, Luminol Bitmap, SAX-REPEAT, k-NN, Bootstrap, and Robust Random Cut Forest. The evaluation performance was quantitatively calculated with seven classification metrics. The C-AMDATS algorithm was able to effectively and better detect the anomalous patterns, and it presented an accuracy of 99%, which leverages the further development of supervised DL models.

Keywords

• multivariate time series
• rotating machinery
• pattern recognition
• anomaly detection
• unsupervised machine learning