Study of pattern recognition based approaches for fault diagnosis

Abstract

Machine defects are behaviors in machines that do not conform to defined, standard behaviors in machines. By implementing health monitoring, anomaly detection and fault isolation on machines, machines can operate more reliably, and unexpected failures, unscheduled maintenance, economic loss, and even casualties caused by machine failures can be minimized. This thesis presents a study of pattern recognition based approaches for machine fault diagnosis based on experimental and synthetic data. The study objectives are carefully selected and are focused on four types of machines: inductionmotor, cooling fan, bearing, and wireless sensor networks (WSN). In this thesis, a health monitoring scheme, which is based on Mahalanobis distance (MD) with minimum redundancy maximum relevance features, is first proposed and applied to anomaly detection in cooling fans. The proposed method helps to avoid multicollinearity in the feature data set and tracks the degradation trends of the cooling fans. Second, Box-Cox transformation is used to convert the non-negative and non-Gaussian distributed variable, MD, into a normally distributed variable, such that the properties of normal distribution can be employed to determine the ranges of MDs corresponding to different health conditions. Mahalanobis-Taguchi system based tool is extended for classifying unbalanced electrical faults of induction motors. Then, supervised and unsupervised learning algorithms are introduced to deal with high-dimensional fault data for dimension reduction and fault classification. Comparative analysis within the algorithms is reported. Finally, autoregressive model and Kuiper test-based passive diagnosis approach is developed for WSN anomaly detection. A health indicator based on Kuiper test is proposed to indicate the health conditions of WSN.

Date of Award	14 Feb 2014
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Wai Shing Tommy CHOW (Supervisor), Kim Fung MAN (Supervisor), Lee Lung CHENG (Supervisor) & Michael Gerard PECHT (Co-supervisor)