Maintenance of equipment at the required condition to ensure a reliable performance, as well as improvement of safety, are major concerns in the field of asset integrity management. Condition monitoring is a procedure that allows one to identify early signs of failures and implement efficient maintenance plans to eliminate the uncertainties in machine operation. In addition, vibration monitoring is known as a detection tool for early detection of degradation from the expected performance. It is often superior to other condition monitoring techniques, due to its high sensitivity and simplicity of implementation. Vibration analysis provides substantial information regarding the operating condition of components and aids to remedy problems. Therefore, it can be used to detect a wide range of fault conditions in rotating machinery, such as imbalance, misalignment of internal shafts, looseness, cracked shaft, gear failures, rolling element bearing damages, motor faults and impeller issues.
The primary intention of the research reported in this dissertation is to investigate the applicability of a neural network methodology for the detection and diagnosis of mechanical defects of impellers in centrifugal pumps. The study focuses on extracting appropriate features from vibration signals associated with pump impellers and the performance of artificial neural networks (ANNs) using these features. The second intention is to enhance maintenance decisions regarding the actual impeller condition. This leads to a transition from time based preventive maintenance to condition based maintenance, and also improving the safety and reliability of pumping systems, as well as reducing unexpected and catastrophic failures. Hence, vibration analysis techniques are used as a principal tool to characterise the impeller conditions under flow variation, with the requirements of data collection, data processing, transformation and selection of essential features corresponding to the running condition.
This dissertation presents a study of current vibration analysis techniques to extract the required features, namely time based features, frequency based features and wavelet based features. An experimental setup is developed to measure the impeller vibration. The experiment is performed using seven impeller fault conditions such as crack and imbalance under fluctuating flow conditions to simulate non-stationary conditions in the system. Also, the evolution of features over varying flow rates are evaluated in order to identify features that contain fundamental information corresponding the fault characteristics. Moreover, the collected features form non-dimensional training data sets are used to train ANNs. Comparisons of different training algorithms, network hidden nodes and effectiveness of different transfer functions are performed to select the most appropriate parameters of networks.
Validation of the results prove that the accuracy of ANN prediction improves considerably by using decomposed vibration signals and energy based features. Comparison of the network accuracy based on wavelet packet transform (WPT) features with time analysis and frequency analysis based features, indicate that WPT-ANN lead to lower mean square errors and higher correlation coefficients, as well as shorter training times. The WPT-ANN model can save computational time and provides better diagnostic information, which can be effectively used for classification of impeller defects under non-stationary conditions.
Dissertation (MEng)--University of Pretoria, 2016.