Fault Detection and Diagnosis of an Industrial Process

Abstract

The suggested work in this thesis aims to develop a new contribution for fault diagnosis in an industrial process, specifically in bearings, based on signal processing and pattern recognition methods. The work presented in this thesis focuses on the detection and diagnosis of bearing defects by using vibration analysis and machine learning. In the first stage, data is acquired from a system or a test rig to be studied taking into account the various types of faults. In this work, two different datasets containing vibration signals are used for bearing fault diagnosis. Then, signal processing techniques such as Blind Source Separation methods will be applied for signal analysis. Among many methods of Blind Source Separation, the Independent vector analysis (IVA) is used for vibration signal analysis as a way to separate the sources of vibration from the observed signals, and to reduce the interference and noise. Next, develop an effective feature extraction technique using time-domain featuresto reduce the dimensionality of the data, remove irrelevant or redundant information, and enhance the meaningfulness and interpretability of the data. Then Feature Selection methods will be applied to reduce the dimensionality the complexity of the extracted feature, which will speed up a learning algorithm and improve the predictive accuracy of a classification algorithm, and avoid overfitting and noise. The second stage is condition classification based on machine learning algorithms. This stage will explore and compare various supervised learning algorithms like Support Vector Machines, Random Forests, Artificial Neural Networks and Extreme Learning Machines to determine the most suitable approach for bearing fault classification. Finally, conducting rigorous testing and validation of the developed machine learning models on real-world industrial machinery to ensure their effectiveness and practicality in a production environment. The effectiveness of the proposed methods in this thesis has been validated by simulated signals and experimental data.