Fault detection and diagnosis (FDD) can be defined as the identification of abnormal process behaviour or an unacceptable deviation in the characteristics that define the system under investigation, identifying the factors related to the fault, identifying the root cause of the fault and subsequently rectifying the abnormal behaviour. A significant proportion of the research work on FDD that has been done so far assume that an accurate model of the system (based on first principles) is known. However, absence of such a model renders majority of the techniques developed in the literature unsuitable for practical real world problems. The main focus of this thesis is development and application of data-based fault detection and diagnosis algorithms based on the state of the art machine learning technique known as support vector machines (SVM). In this thesis fault detection and diagnosis (FDD) has been viewed from two different perspectives: FDD in the field of medicine for disease diagnosis based on clinical data and FDD in process industries for process monitoring based on process data.
A detailed description of the significance of FDD algorithms that are used in the literature in the field of medicine and process industries is provided. The concepts are clearly explained with appropriate figures and examples. A detailed literature review of the statistical and machine learning data based FDD algorithms that have been used in the literature is also presented. An extensive tutorial introduction of support vector machines has been presented. The main characteristics and advantages of SVM over other machine learning techniques have been clearly explained with the help of appropriate examples. A novel methodology based on support vector machines is proposed to develop decision support system for disease diagnosis based on clinical metabolomic data. This thesis also has a component focusing on the signal processing aspect of high dimensional metabolomic nuclear magnetic reso nance (NMR) spectroscopy data. A novel, computationally efficient adaptive binning algorithm has been proposed to correct for peak shifts and binning of the NMR data, thereby reducing the dimension of the dataset. A novel algorithm based on 1-class SVM is proposed to do process monitoring based on normal operating data. It is shown that the proposed method had a superior performance in comparison with conventional algorithms used in the literature. A powerful and efficient algorithm based on 1-class SVM has been proposed to carry out fault identification in process data. The last component of this thesis deals with detection and diagnosis of faults in rotating machineries based on 1-class SVM applied to the spectral vibration data. The model that is developed is then tested on different levels of crack of the gear tooth. The 1-class SVM contribution metric will also be used to diagnose the fault in the gearbox system.