FAULT DIAGNOSIS OF MOTOR BEARING USING ENSEMBLE LEARNING ALGORITHM WITH FFT-BASED PREPROCESSING

Abstract

Rolling bearings are one of the pivotal mechanical elements in rotating machines like the electric motor. However, they are liable for the majority of the faults encountered by rotating machines. Detection or estimation of these faults at an early stage can help to eliminate them and prevent the machine from malfunctioning or failing during operation. The recent developments in the field of Machine Learning (ML) have brought a radical change in the way we interpret and analyze these faults, and certain learning techniques have enabled us to predict motor bearing faults almost impeccably. This paper proposes a method to diagnose bearing fault signals that employ an ensemble learning method named Random Forest (RF). The procedure associated with this method requires simple preprocessing using Fast Fourier Transform (FFT) that explore bearing vibration signal to reveal intrinsic features about fault which are used with RF for classifying fault types. The potency of the proposed method is demonstrated using the practical motor vibration data obtained from the Case Western Reserve University (CWRU) Lab. This supervised learning algorithm is able to classify and predict various types of bearing faults with almost 99% accuracy.

Existing System

Support Vector Machines (SVMs), feed forward neural network, one-class m-SVM and Hilbert Transform

Proposed System

In this paper, we are going to use Fast Fourier Transform (FFT) to preprocess the fault data; then we will normalize the preprocessed signals and use Random Forest (RF) classifier to analyze various types of motor bearing faults. FFT is not a new transform, rather a variant of the original Fourier transform which is widely used for countless applications related to science, engineering, and mathematics. It is a method for efficiently quantifying the Discrete Fourier Transform (DFT) of a series of data samples (typically in the time domain) which allows us to conveniently analyze them in the frequency domain and design systems to process them further. RF is a classifier that is made of multiple tree structured classifiers. It creates a number of decision trees at the time of the training and results in the class which is either the mode of the classes or mean prevision of the distinctively generated trees

Architecture

BLOCK DIAGRAM