Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
One-class Support Vector Machines for Real-time and Unsupervised Condition Monitoring of Rotary Machines
Submission Author:
Alexandre Henrique Pereira Tavares , MG
Co-Authors:
Alexandre Henrique Pereira Tavares , Aldemir Ap Cavalini Jr, Antonio Veiga, C. Nataraj, Amirhassan Abbasi
Presenter: Alexandre Henrique Pereira Tavares
doi://10.26678/ABCM.COBEM2023.COB2023-1929
Abstract
Constant monitoring of industrial rotating machines is of great importance to ensure their health, reliability, and safety. In addition, costs with unplanned downtime due to preventive maintenance are reduced, as maintenance will only be performed if necessary. Numerous systems perform constant, real-time health monitoring of rotating machines. However, these systems require domain experts to define their parameters based on the type of rotating machine being monitored. Alternatively, some systems use machine learning algorithms with supervised or unsupervised learning techniques, to allow the algorithm to learn by itself the behavior of the rotating machine, reducing the need for an expert to set the system parameters to work ideally with each machine. Supervised learning algorithms require labeled data of normal and abnormal behavior, which is rarely available in a real-world system as it requires an enormous amount of time from experienced personnel to label enough data to train learning algorithms. The current study develops an algorithm that uses a one-class support vector machine algorithm to perform unsupervised conditioning monitoring. The condition monitoring problem was approached in a way that is similar to detecting anomalies, except that the algorithm focuses on identifying variations in the condition of the machine rather than detecting any condition that deviates from the normal. The inputs for the algorithm are the signals collected from proximity sensors attached to the rotor shaft processed by a discrete wavelet transformation. The algorithm was validated with different signals, including synthetically generated ones, and signals collected from sensors attached to real rotating machines. During the validation, the performance of the algorithm was measured using the F1 score, which showed a value of close to 0.9 for most of the Tests. This high value demonstrates the high capability of the algorithm in detecting variations in the machine conditions.
Keywords
One-class Support Vector Machine, machine learning, Anomaly detection, Rotary machine, condition monitoring
