| During the past decades, vibration response analysis has become a cornerstone among all others methods used for the condition monitoring of rotating machinery. Many researchers have shown that using the analysis of the dynamical behaviour of rotating components, it may be possible to monitor, detect and diagnose any incipient failure in the system. In practice, vibration signals measured by sensors contain many components which may not be useful in the characterization of the signal. This makes it difficult to interpret vibration analysis results, as the research has tuned towards the use of pattern recognition methods to represent machine condition in a high dimensional hyperspace, where we expect the characterization to be simpler.;This thesis addresses the problems posed by signal path transmission on the reliability of the predictions issued from pattern recognition methods. First, we show that signal path transmission has a great influence on the generalization ability of classifiers, regardless of the dimensionality of the hyperspace. Results show that the generalization ability of the classifiers will not exceed 60% in general, when new signals, having a different signal path transmission than those used during the learning, will be presented to the classifier. Solving that issue will open many breakthroughs in the monitoring of critical components such as aircraft engines, nuclear turbines and compressors. We will no longer have to worry about the location of monitoring sensors, as these components are under strict regulations.;This memoir proposes a combination of advance signal processing methods, like Time Synchronous Averaging, Spectral kurtosis, Adaptive Noise Cancellation, with pattern recognition methods, namely support Vector Machines, to overcome the problem posed by signal path transmission in signal identification. A diagnostic procedure that integrates these methods is implemented as follow: vibration responses analysis methods are first used to isolate the components of the vibration signal coming from the faulty components. Next, vibration parameters (22 and 40 parameters) are extracted from this component to from a feature vector that will represent machine state condition in a high dimensional hyperspace. Then Support Vector Machines are train to characterize these vectors and classify them according to their failure class. Later, these classifiers will be used to classify new vibration signals having different signal path transmission from those used during the learning.;The implemented diagnostic procedure has been tested and validated using vibration signals coming from two different tests rigs. Each test rig is composed by two bearings seats, one of them containing a faulty bearing. Accelerometers are used to pick-up vibration signals of each bearing. Using our diagnostic procedure, generalization performance of the classifiers vary between 46% and 100%, the mean being at 80%, which is a great improve regarding the complexity of these systems.;Even if, it was not first required for the completion of this thesis, this work is also addressing the use of a genetic algorithm to find features that are most useful during the learning when the transmission path is different. The genetic algorithm implemented in this memory is a basic algorithm. We are only interested in finding a feature vector that minimizes the classification error of svm-boundaries. Results show that among the 22 or 40 features extracted, there are boundaries where only 17, 6 or even 2 parameters would be sufficient to have a generalization performance higher than 90%.;Keywords. Condition monitoring, bearing fault, Vibration response analysis, Pattern recognition methods, Support Vector Machine, Spectral Kurtosis, Time Synchronous Averaging, Adaptive Noise Cancellation, Feature selection, Feature extraction, Genetic Algorithm. |
PDF Full Text Request