الفهرس Only 14 pages are availabe for public view
 |  | from | 89 |  |  |
| | | from | 89 | | |
Abstract
The fault detection relies on the relation between the deviation of a process parameter and the fault that causes this deviation, or on the appearance of a new parameter and the’ fault that generates it. The selection of a fault indicator is the critical step in any diagnosis system. The selected indicator must reflect as close as possible the primary effect of the fault and must contain as much as possible information about the fault. The objective of this thesis is to detect, locate (phase A, B or C) and estimate stator inter-turn faults. This task was successfully achieved after an extensive investigation on the effect of these faults on the machine electrical variables. All along the investigation, the inter-turns fault effect was compared with the effect of other faults. The positive sequence-current variatiqn is found specific to inter-turns but insufficient to the localization process. The 6ehavior of the stator current RMSs are found useful to locate the fault because they have a specific behavior with respect to the fault location. Also the investigation of the Concordia patterns indicated that they are directly related to the faults location and degree. These three fault indicator was found insufficient to identify and estimate the rotor asymmetries but the spectral analysis was very useful with these faults. Beside the detection of stator inter¬turns, the study and comparison of these features allowed the detection of supply imbalances and rotor faults. These features draw the guidelines of the fault diagnosis process. The neural network was investigated as a mean to the fault detection process. Neural networks were found to be a very efficient diagnosis tool. The input of the networks can be either the stator current RMSs or Concordia patterns. The interpretation of the network output is done using few tables. After the training, the network was able to evaluate and locate the faults. The study of the neural networks and the features extracted trom the investigation of the fault effects were used to build a diagnosis algorithm. The algorithm requires only the stator currents and generates a diagnosis report. The localization of the fault is done with a very good precision and the fault estimation was improved using two networks at the same time. The interpretation of the output of the algorithm as a diagnosis report can be done using tables that map the activated node to a specific fault location.