An End-to-End Unsupervised Anomaly Detection Pipeline for Electrical Motors
Aydin, Ugur (2023)
Aydin, Ugur
2023
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2023052313435
https://urn.fi/URN:NBN:fi:amk-2023052313435
Tiivistelmä
Due to the pervasive use of electrical machines in the industry, it is important to ensure reliable operation of these devices while avoiding unplanned downtimes. Typically, the condition monitoring of electrical machines is done by utilizing various sensors that requires retrofitting with lengthy installation and setup processes. ABB developed Smart Sensor™ that is tailored for condition monitoring of electrical motors and easy installation to tackle this issue. The Smart Sensor™ provides cloud connectivity for data gathering and analytics where advanced algorithms can run to detect machine health status. Considering the condition monitoring algorithms, in the past decades, in addition to traditional methods, machine learning approaches have been utilized. Although various machine learning approaches were proven to be extremely effective for a single machine with labeled data, the generalization of these models are questionable in typical industrial setting where there are various machine sizes, loads and operating conditions. Collecting labeled data to cover wide range of applications is impractical and costly.
Due to impracticality of collecting labeled data for wide range of applications, in this work an end-to-end unsupervised anomaly detection solution was developed utilizing Smart Sensor™ data. The constructive research methodology and the CRISP-DM approach were chosen as the research and development methods, respectively. After choosing the appropriate evaluation metrics and features, experiments with various unsupervised machine learning approaches were performed with the Smart Sensor™ data on various machines. It was shown that for the problem at hand, simple machine learning approach, namely One-Class SVM is highly effective at detecting anomalies with very low computational effort. After the model selection, a highly scalable, reliable and reproducible end-to-end machine learning pipeline was constructed and deployed to Azure cloud platform following the MLOps practices. The machine learning pipeline was tested on 16000 connected machines and shown to perform well considering the computation time.
Due to impracticality of collecting labeled data for wide range of applications, in this work an end-to-end unsupervised anomaly detection solution was developed utilizing Smart Sensor™ data. The constructive research methodology and the CRISP-DM approach were chosen as the research and development methods, respectively. After choosing the appropriate evaluation metrics and features, experiments with various unsupervised machine learning approaches were performed with the Smart Sensor™ data on various machines. It was shown that for the problem at hand, simple machine learning approach, namely One-Class SVM is highly effective at detecting anomalies with very low computational effort. After the model selection, a highly scalable, reliable and reproducible end-to-end machine learning pipeline was constructed and deployed to Azure cloud platform following the MLOps practices. The machine learning pipeline was tested on 16000 connected machines and shown to perform well considering the computation time.