Development of Deep Learning-based Approach for Robust Fault Diagnosis and Failure Prediction of Bearings

Xiuyu, Li (2025) Development of Deep Learning-based Approach for Robust Fault Diagnosis and Failure Prediction of Bearings. PhD thesis, UNIMAS.

	PDF dsva_Li Xiuyu.pdf Restricted to Repository staff only Download (323kB)
	PDF Thesis PhD_Li Xiuyu.pdf Download (5MB)

Abstract

Bearings play a crucial role in rotating machinery. Once a bearing failure occurs, it may lead to severe casualties and economic losses. Therefore, diagnosing bearing faults and ensuring their smooth operation are essential steps in ensuring the safe and stable operation of modern mechanical equipment. However, traditional methods for diagnosing bearing faults typically involve complex manual denoising, filtering, and feature extraction processes, which are both tedious and lack intelligence, making them particularly challenging when dealing with large volumes of data. The operating conditions of bearings are complex and involve various factors, making manual extraction of fault features even more difficult. Thus, this motivates research in improvements of end-to-end bearing fault diagnosis and life prediction. Furthermore, the rapid development of information technology has made deep learning a powerful tool for automatically learning and diagnosing bearing faults under complex conditions. This research aimed to investigate the end-to-end bearing fault diagnosis and life prediction under complex scenarios such as high noise and different operating conditions, explore various deep neural networks for effective bearing degradation prediction and fault diagnosis. To achieve the aim, this research was segmented to 3 parts: (i) Firstly, to achieve end-to-end diagnosis of bearing fault types and severity, this thesis proposed a Multi-scale processing, Channel attention, Feature enhancement, Convolutional Neural Network (MCFCNN) network architecture. The MCFCNN network combines multi- channel parallel convolution, effectively capturing spatial information, and introduces channel attention mechanisms to adaptively recalibrate channel-level feature responses. Moreover, secondary neurons are introduced to enhance the model’s ability to capture complex nonlinear patterns related to bearing faults. In a high-noise environment, the proposed model outperforms existing models in fault diagnosis, with accuracy greater than 80%. (ii) Secondly, a bearing fault test platform was built to validate the effectiveness of the MCFCNN model. Rolling bearings vibration signals were generated using the test platform. In the experiments, the performance of the MCFCNN model in high-noise conditions and the results of different network structures and data augmentation strategies on the final results were verified. In the cross operating scenario, the average resolution in the Bearing Fault Test System (BFTS) dataset is 0.7595, compared to WDCNN with 0.6492 and QCNN with 0.7266. (iii) In the third and final part of this research, the problem of predicting the remaining life of bearings was approached as a classification problem. Rather than predicting the remaining life, the priority was to identify the stage of a bearing’s life cycle for better maintenance and management. Bearing life was divided into four different stages and used the bearing data to predict the current stage of the bearing’s life. Based on this, a method for early prediction of bearing degradation, referred to as Fourier Transform Autoencoder with K-means (FAEK) automatic labeling and Multi-scale Processing Channel Attention Classification (MCC) classification prediction model, was proposed. This method combines a convolutional autoencoder segmentation network with a high-performance classifier, effectively distinguishing bearing degradation processes and achieving accurate classification recognition. In the degradation stage prediction task, our model architecture achieves an accuracy of 0.9665. This thesis focuses on deep learning research applied to bearing fault diagnosis and prediction, which helps to reduce the workload of manual maintenance personnel, enhance the efficiency and accuracy of fault detection, and mitigate losses caused by equipment failures. Indirectly, this reduces maintenance costs, enhances equipment reliability and improves production efficiency, through timely maintenance and fault prevention of equipment.

Actions (For repository members only: login required)

View Item