Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Application of sinusoidal analysis to feature extraction in rotating machine vibration signals
Submission Author:
Dionísio Henrique Carvalho de Sá Só Martins , RJ
Co-Authors:
Dionísio Henrique Carvalho de Sá Só Martins, Diego Haddad, Amaro Lima, Milena Pinto, Denys Pestana-Viana, Ulisses Admar Barbosa Vicente Monteiro
Presenter: Dionísio Henrique Carvalho de Sá Só Martins
doi://10.26678/ABCM.COBEM2021.COB2021-1704
Abstract
Rotating machines as turbines, pumps, conveyors and compressors are frequently subject to a wide range of harsh conditions that induce mechanical faults and performance degradation. The extraction of relevant features of vibration signals from rotating machines has been the subject of intense efforts during the last few years. A fault signature can be seen as a set of characteristics related to a particular failure. These characteristics should be extracted from specific vibration signals information in an adequate manner. This paper presents a novel feature extraction method for fault diagnosis problems in rotating machines. The proposed method utilizes sinusoidal analysis, which comprises several methods traditionally employed in the audio signal processing field. This paper is the pioneer, up to the authors knowledge, in utilizing such family of methods to address pattern recognition problems encountered in rotating machines under industrial settings. The main goal of the paper is to differentiate isolated faults composed only of imbalance, vertical misalignment and vertical misalignment from combined faults formed by imbalance associated with horizontal misalignment, imbalance associated with vertical misalignment and horizontal misalignment associated to vertical misalignment. In this research, the vibration signals were described by the sinusoidal analysis, which aims to represent a vibration signal through a finite set of components formed by time-varying amplitude, phase and frequency. After computing this novel time-frequency representation, the method extracts the main signal tracks, which are responsible for containing the most relevant information about a signal. Some peculiar features are then obtained from these tracks, such as the birth frequency, the length of the largest track, the energy of the largest track, the variance of birth frequencies, among others. The performed experiments revealed that such sinusoidal features can be beneficial for classification purposes. In the database utilized, one reaches a global accuracy of 83.30%, when such features are combined with the Random Forest algorithm as the pattern recognition method. Further, the selection of the optimal number of features obtained through the sinusoidal representation by intelligent algorithms and the selection of the best classification algorithm model capable of distinguishing classes through the recognition of sinusoidal features patterns are investigated
Keywords
sinusoidal analysis, Feature Selection, Fault Diagnosis, combined fault
