الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 131 |  |  |
| | | from | 131 | | |
Abstract
In recent years, the application of deep learning techniques in the field of three-phase induction motor fault diagnosis has gained significant attention. This thesis presents a novel approach utilizing Convolutional Neural Network (CNN) models specifically designed to diagnose faults occurring during the starting period of induction motors. The proposed model demonstrates its effectiveness in detecting a variety of faults, including locked-rotor, overload, voltage unbalance, overvoltage, and undervoltage, under different loading conditions, namely Light, Normal, and Heavy loads. To ensure high reliability and accuracy, the model utilizes image data derived from multiple motor signals such as voltages, currents, torque, and speed. One of the key challenges in fault diagnosis during the starting period lies in effectively analyzing the transient response of the motor, particularly when relying solely on raw time-domain images. To address this challenge, the measured signals are further represented as d-q and Lissajous images. This additional representation allows for a more focused exploration of the impact on the model’s performance when utilizing these representations. By utilizing multi-signal data, the study investigates the effectiveness of two different input shapes: single-channel and multi-channel. The goal is to determine the most efficient form that captures the relevant fault features. To evaluate the proposed fault diagnosis system, a series of tests were conducted using a simulated machine model. Eight distinct datasets were generated from the motor data, designed to cover various signal representations and input shapes. These datasets were then used for training and testing the CNN model. The experiments aimed to investigate and determine the optimal signal representation and input shape for the proposed fault diagnosis system. Through these experiments, valuable insights were gained into the performance and effectiveness of different signal representations and input shapes, providing crucial guidance for enhancing the accuracy and efficiency of the fault diagnosis system. Comparing the performance of different signal representations, it is observed that datasets incorporating both voltage and current d-q signal representations achieve the highest accuracy. These datasets consistently outperform other representations, yielding accuracies improvements ranging from 0.03% to 3.57%. Furthermore, the comparison between input shapes reveals that multi-channel datasets consistently outperform single-channel datasets. The multi-channel datasets exhibit superior performance, achieving accuracies improvements ranging from 0.77% to 4.06%. The findings of this study contribute significantly to the field of fault diagnosis in induction motors. By leveraging deep learning techniques and incorporating multiple signal representations, the proposed CNN model demonstrates its capability to accurately detect various faults during the starting period. Moreover, the evaluation of different input shapes provides valuable insights into optimizing the model’s performance.