Intelligent methods for condition monitoring of rolling bearings using vibration data

Abstract

Owing to the importance of rolling bearings in rotating machines, there has been great interest in the development of computational methods for rolling bearings condition monitoring over the last few decades. The aim of these methods is to determine early and automatically the occurrence of a fault condition in rolling bearings to avoid machine breakdowns that may lead to downtime, possibly safety incidents, production lost, and higher costs of repairs. Vibration-based methods are commonly used and have become well-accepted techniques of many condition-based maintenance (CBM) management. However, the vast amount of the collected vibration data requires large storage and time for signal processing and this also may limit the number of machines that can be monitored remotely across wireless sensor networks (WSNs) due to bandwidth and power constraints. To avoid the burden of much storage requirements and processing time of a tremendously large amount of vibration data, the scope of this thesis is the development and application of vibration analysis methods of a large amount of acquired vibration dataset for condition monitoring of rolling bearings. The main aim of this analysis is to obtain compressively-sampled dataset which possesses the quality of the original vibration dataset and then learn fewer features from these compressively-sampled signals directly without reconstructing the original signal. Regardless of its importance in addressing the challenge of learning from a large amount of vibration dataset, no existing method possesses the ability to learn directly from low-rate of compressed measurements. This thesis presents new and tested vibration condition monitoring methods for roller bearings that can greatly compress the amount of the original vibration and further learn features from this compressed amount of data to increase the identifying power of the compressed measurements in rolling bearing faults diagnosis. There are two most important contributions in this thesis. The first one is the formulation of a three-stage method, Compressive Sampling with Correlated Principal and Discriminant Components (CSCPDC) for classification of bearing faults. This method applies CS to obtain compressively-sampled signals from the raw vibration data and then adopts a multi-step feature learning algorithm to learn fewer features from the compressively-sampled signals. Finally, it employs a multi-class support vector machine (SVM) to classify bearing health conditions using these learned features. The second one is the design of intelligent condition monitoring method for bearing faults from highly compressed measurements using sparse overcomplete features. This method applies CS to produce highly compressed measurements of the original bearing vibration dataset and then it uses an effective deep neural network (DNN) with unsupervised feature learning algorithm based on sparse autoencoder to learn over-complete sparse representations of these compressed measurements. Finally, it employs two techniques to deal with the classification problem, namely, pre-training classification based on stacked autoencoder and softmax regression layer, and retraining classification based on backpropagation algorithm. These methods and other less performance methods proposed in this thesis have been validated and applied to several real vibration datasets of roller bearings. The experimental validation demonstrates improved bearing health condition classification accuracy, with highly reduced feature dimension, and much lower computational complexity, compared to state-of-the-art methods.