09 Apr 2016

A Sensitivity Comparison of Neuro-fuzzy Feature Extraction Methods from Bearing Failure Signals

Jonny Latuny, July 2013
Curtin University

Abstract
Bearings play an important role in the operation of rotating machinery. They are critical machinery components that are subject to continuous load and harsh operational conditions and hence prone to failure during its lifetime. Bearing failure can lead to breakdown of the whole machine which could in turn lead to unwanted stoppage of an industrial production line. Therefore, the operational condition of a bearing must be monitored for the purpose of maintenance and avoidance of unwanted machinery stoppage that might be caused by such a bearing failure.

Better understanding of a bearing failure condition is useful in maintaining a continuous production line. In this context, a scheduled maintenance event, based on bearing fault diagnosis, is an advantage, preventing unwanted production line maintenance events. This is an important requirement in the prevention of revenue loss due to stoppage of the production line or in any rotating machinery related to human safety (i.e., transportation, etc)

The importance of bearing fault analysis and classification in relation to maintenance cost reduction and safety issues have been the motivation for intensive and wider research aimed at providing better methods for bearing fault analysis and classification.

A bearing fault diagnosis system which utilises advanced combinations of vibration signal processing and artificial intelligent methods has gained attention in recent years. In past decades, the trend was to combine digital signal processing with available artificial intelligence (AI) techniques to produce better and reliable bearing fault diagnosis systems. The application of signal processing and artificial intelligent methods is an open research field for investigation and exploration for the purpose of obtaining a new combined methodology.

A review of the literature, which includes vibration signal analysis in fault diagnosis, statistical parameter applications in feature extraction methods, wavelet transforms and artificial intelligence systems in fault diagnosis, is presented. Based on the literature review, it was found that there is a possibility of proposing a new combined method for the purpose of building a bearing fault diagnostic system. In particular, the application of new feature extraction methods combined with artificial intelligence (AI) systems. There were no standard guidelines available in finding better systems for bearing fault analysis and classification through the utilization of combined feature extraction methods and AI applications.

This research reports an investigation process of building bearing fault classifiers for outer race, inner race and ball fault cases using a wavelet transform, statistical parameter features and an Artificial Neuro-Fuzzy Inference System (ANFIS). The building process started by acquiring and processing raw vibration signalfrom a bearing under investigation. The data acquisition process was carried out for both normal (fault-free) and faulty operation of a double row self-aligning ball bearing.

An accelerometer was used to collect the vibration data from a faulty bearing. The raw vibration data was processed using a wavelet transform employing a Daubechies wavelet filter to produce wavelet coefficients and their energy levels. The result was then processed to extract the statistical parameters (i.e., kurtosis, RMS, variance, standard deviation). The features generated from statistical parameters and wavelet transform scheme were then used to train the ANFIS.

In order to reduce the number of rules generated during the training process, only two inputs were used for the purpose of building the classifier. The selection of the most influential inputs for the training process of the ANFIS was achieved through the use of the ANFIS built-in capability of selecting the best correlation of two inputs towards one target output which best represented the bearing operating condition.

An extensive computation was used in the process of selecting the most influential input-output combination fromthe six inputs available. The number of input-output combinations tested was 720,being the permutations of six inputs. In the search for the best combination of input-output, the possible combinations of statistical parameters, wavelet coefficients and wavelet’s level of energy were investigated extensively in order to obtain the best classifier for bearing fault diagnosis.

The ANFIS was then implemented to capture the input-output relation of the selected inputs to generate a suitable classifier that could be used to classify bearing operating condition. The classifiers generated were then tested to evaluate their ability and accuracy in predicting a faulty bearing.

The test results show that the ball fault (BF) classifier successfully achieved 100% accuracy without mis-classification, while the outer race fault (ORF), inner race fault
(IRF) and no fault (NF) classifiers achieved mixed percentage between successful classification and mis-classification results.

Share this article on:

Related Article


Back to Top