Guangxing Niu

Date of Award

Spring 2023

Document Type

Open Access Dissertation


Electrical Engineering

First Advisor

Bin Zhang


Rolling element bearings are critical components in industrial rotating machines. Faults and failures of bearings can cause degradation of machine performance or even a catastrophe. Therefore, it is significant work to perform bearing fault diagnosis and prognosis (FDP) reliably and effectively. Diagnosis aims to detect and estimate the fault state of bearing in real-time. Prognosis aims to conduct a long-term prediction to predict the fault evolution and estimate the remaining useful life (RUL) of the bearing.

Deep learning (DL) based FDP methods have become an important branch of bearing FDP methods due to their powerful capabilities in feature automatic learning, fault modeling, and fault identification. Although they have made significant achievements in bearing FDP, there still exist some challenges and open problems, especially for modern industrial systems that are often designed with multiple complicated functions and operated under varying operating conditions and environments. Most existing DL-based FDP methods are designed for stationary operating conditions and are purely data-based, which often cannot guarantee good results and high efficiency for FDP of modern industrial systems.

To overcome the challenges and improve the performance of the existing DL-based bearing FDP methods, this thesis proposes some improvements of DL-based bearing FDP methods in terms of structure optimization, adaptive learning strategy, FDP algorithm execution strategy, etc. The proposed DL-based FDP methods can generate better results in accuracy, efficiency, and robustness for bearing FDP tasks.

Deep belief network (DBN) and convolutional neural network (CNN) are two mainstream DL structures. They have different advantages and unique network characteristics in various applications. This thesis is conducted based on these two networks to improve their performance in FDP applications. This thesis first proposes a deep belief network (DBN) and principal components analysis (PCA) based approach for bearing fault classification. A particle swarm optimization (PSO) based DBN adaptive training procedure is employed to optimize the DBN structure. This approach provides an automatic, accurate, and effective bearing fault classification solution.

To improve the accuracy and learning efficiency of the existing CNN-based bearing diagnosis methods, a deep residual CNN is proposed for multi-task bearing fault diagnosis. In the proposed approach, domain knowledge is integrated with monitoring data to build the information map. Two attention modules are introduced to enhance the discriminate feature learning ability. Two classifiers are employed for multi-task diagnosis. This diagnosis method has significant improvements in terms of diagnostic accuracy and training efficiency.

For the bearing diagnosis under varying speeds, a novel multi-scale discriminate CNN based bearing fault diagnosis is proposed to deal with the challenges caused by the varying speeds. The varying speed information is integrated with monitoring data to build the information map. A multi-scale discriminate convolutional neural network-based method is designed to enhance the learning ability for signals with some specific characteristics. The extracted features are then employed to identify bearing fault modes. Experimental results and comparisons show that the proposed approach can achieve better performance in terms of accuracy and efficiency than some state-of-the-art methods.

For bearing fault prognosis, this thesis proposes a hybrid Bayesian estimation-based FDP framework with fault detection and multiple model fusion. In the proposed approach, convolutional neural network (CNN) is used to detect fault and select appropriate fault dynamic model. To improve the performance, continuous wavelet coefficient matrices (CWCM) power spectrum of vibration are fused with operating conditions to build information maps for fault detection and model selection. After a fault is detected, Bayesian estimation based FDP method is triggered to estimate the fault state and predict the remaining useful life. In the FDP process, Dempster-Shafer theory (DST) is employed to fuse prediction results from different models if necessary. The bearing state and RUL can be estimated based on PF-based estimation and fusion results.

The proposed methods are verified with different bearing case studies. The experimental results are analyzed and compared with other state-to-art approaches, which demonstrate that the proposed approaches have better performance. This thesis provides a successful exploration and attempt of DL algorithms in dealing with fault diagnosis and prognosis problems.