Fault Analysis is the detection and diagnosis of malfunction in machine operation or process control. Early fault analysis techniques were reserved for high critical plants such as nuclear or chemical industries where abnormal event prevention is given utmost importance. The techniques developed were a result of decades of technical research and models based on extensive characterization of equipment behavior. This requires in-depth knowledge of the system and expert analysis to apply these methods for the application at hand. Since machine learning algorithms depend on past process data for creating a system model, a generic autonomous diagnostic system can be developed which can be used for application in common industrial setups. In this thesis, we look into some of the techniques used for fault detection and diagnosis multi-class and one-class classifiers. First we study Feature Selection techniques and the classifier performance is analyzed against the number of selected features. The aim of feature selection is to reduce the impact of irrelevant variables and to reduce computation burden on the learning algorithm. We introduce the feature selection algorithms as a literature survey. Only few algorithms are implemented to obtain the results. Fault data from a Radio Frequency (RF) generator is used to perform fault detection and diagnosis. Comparison between continuous and discrete fault data is conducted for the Support Vector Machines (SVM) and Radial Basis Function Network (RBF) classifiers. In the second part we look into one-class classification techniques and their application to fault detection. One-class techniques were primarily developed to identify one class of objects from all other possible objects. Since all fault occurrences in a system cannot be simulated or recorded, one-class techniques help in identifying abnormal events. We introduce four one-class classifiers and analyze them using Receiver-Operating Characteristic (ROC) curve. We also develop a feature extraction method for the RF generator data which is used to obtain results for one-class classifiers and Radial Basis Function Network two class classification. To apply these techniques for real-time verification, the RIT Fault Prediction software is built. LabView environment is used to build a basic data management and fault detection using Radial Basis Function Network. This software is stand alone and acts as foundation for future implementations.
Library of Congress Subject Headings
Fault location (Engineering); Learning classifier systems; Classification--Data processing; Computer algorithms--Evaluation; Machine learning
- Please Select One -
Department, Program, or Center
Electrical Engineering (KGCOE)
Chandrashekar, Girish, "Fault analysis using state-of-the-art classifiers" (2013). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus