A Generalised AI-based Model for Rolling Element Bearing Fault Diagnosis in Rotating Machinery

Abstract

Rolling element bearings (REBs) are crucial for the efficient and safe operation of industrial rotating machinery. Since REB faults can lead to unexpected breakdown, increased repair expenses, as well as safety risks, accurate fault detection and diagnosis methods are highly demanded. This PhD study aims to develop an intelligent fault diagnosis (IFD) model capable of detecting and diagnosing REB defects based on vibration data, signal processing techniques and artificial intelligence (AI) algorithms. There are significant challenges facing AI implementation in industries such as quality and quantity of data for model training and generalization across machines and different working conditions. In order to overcome these limitations, this research has identified a range of REB vibration datasets from experimental work and industry fieldwork sources for the AI model development. Through the study of existing signal processing methods, a new hybrid technique combining cepstrum pre-whitening (CPW) and high-pass filtering has been developed in this study to effectively extract bearing characteristic frequencies (BCFs), which has been a foundation for subsequent investigation in bearing fault diagnosis using machine learning (ML) methods. The novel hybrid method automates the envelope analysis and extracts useful features from vibration data, validated using simulated and different experimental and real-world datasets. To diagnose bearing faults, firstly a two-stage model using the novel signal processing method and traditional ML algorithms has been developed to classify healthy and faulty REBs in the first stage then classify fault types in the second. The model's effectiveness and accuracy is evaluated and compared with existing recent developed models. To improve the efficiency and interpretability of the method in the first stage, i.e., identifying a faulty from healthy bearings, a universal health indicator (UHI) has been developed by calculating the energy ratio of BCFs to background noise in vibration spectra using the novel hybrid method, then assess the bearing healthy/faulty conditions in a UHI model established using a variety of datasets. Finally, a convolutional neural network (CNN) fault classification model is trained using simulated vibration signals to improve fault classification accuracy in the second stage, particularly for complex cases. The model has been tested using diverse experimental and real-world datasets, demonstrating high accuracy compared to the traditional ML techniques and existing models in this domain. A generalised AI-based model for REB fault diagnosis, eliminating the need for training datasets for new case studies has thus been developed. This promises better machinery reliability, lower maintenance costs, and increased safety in industrial applications.

More information

Identifiers

ORCID for Amirmasoud Kiakojouri: orcid.org/0000-0001-5978-1970

ORCID for Ling Wang: orcid.org/0000-0002-2894-6784

ORCID for Zudi Lu: orcid.org/0000-0003-0893-832X

Catalogue record

Export record