DATA-DRIVEN FAULT DIAGNOSIS FRAMEWORKS USING ADVANCED SIGNAL PROCESSING, AND MACHINE LEARNING TECHNIQUES

Abstract

Continuously operating in extreme working conditions causes industrial systems to break after a period of use which leads to performance deficiencies and adverseffectsct on safety. For that reason, the reliable fault diagnosis (FD) of complex industrial systems a pressing demand in the industrial sector. Take a look at popular industrial systems, it is clear that devices such as bearings, power transformers, pressure vessels, or concrete structures are ubiquitous. The failure in these machines usually goes with the symptoms that generating abnormally vibration or elastic waves in form of the acoustic emission (AE). Such signals can be obtained in its entirety by modern vibration acceleration sensors or AE sensors. Because of the ability of efficiently capturing low-energy fault signatures in the elastic waves, the AE sensors have been chosen for all studies in this dissertation. The most important step in designing any FD system is detecting early fault signatures inhering in AE signals collected in advance. This step is effectively done by advanced signal and data processing techniques. Following this, state-of-the-art machine learning and deep learning techniques should be utilized to exploit patterns in the signals to classify fault symptoms or predict outcomes of interest. As a result, this dissertation mainly focuses on these techniques and presents the way to apply the techniques on every specific industrial devices. Major sections of this thesis are listed as follows:

Firstly, an efficient scheme for the early diagnosis of bearing defects using a convolutional neural network (CNN) and energy distribution maps (EDMs) of acoustic emission spectra. The CNN automates the process of feature extraction from the EDM. The features learned by the CNN are used by an ensemble classifier, that is, a combination of a multilayer perceptron that is integral to typical CNN architectures and a support vector machine to diagnose bearing defects.

Secondly, we present a reliable fault detection model for a pressure vessel under low pressure conditions. To improve the diagnostic performance, signals of different vessel health conditions are purified by eliminating noise so that signals of different categories are much more distinguishable. This de-noising technique uses a blind source separation (BSS) technique in which an initial noise-contaminated signal is separated into constituent sources. These individual sources are either device status-characterizing signal sources or interfering sources. Noise is removed and an unimpaired signal is regenerated from the characteristic sources.

Thirdly, the dissertation demonstrates a new approach to characterize fracture modes in a concrete structure using an acoustic emission (AE) technique and a data-driven technique. To clarify the damage fracture process, the specimens, which are of reinforced concrete (RC) beams, undergo four-point bending tests. During bending tests, impulses occurring in the AE signals are automatically detected using a constant false-alarm rate (CFAR) algorithm. For each detected impulse, its acoustic emission parameters such as counts, duration, amplitude, rise-time, energy, RA, AF are calculated and studied. The mean and standard deviation values of each of these parameters are computed in every 1-second AE signal and are considered as features demonstrating the damage status of concrete structures. The results revealed that as the damage level in concrete structures grows, these features also change accordingly which can be used to categorize the damage fracture stages.