Document Type
Dissertation
Date of Award
8-31-2022
Degree Name
Doctor of Philosophy in Computer Engineering - (Ph.D.)
Department
Electrical and Computer Engineering
First Advisor
MengChu Zhou
Second Advisor
Seyyedmohsen Azizi
Third Advisor
Nirwan Ansari
Fourth Advisor
Hieu Pham Trung Nguyen
Fifth Advisor
Qing Gary Liu
Sixth Advisor
Zhi Wei
Abstract
Over the past thirty years, the idea of computing based on models inspired by human brains and biological neural networks emerged. Artificial neural networks play an important role in the field of machine learning and hold the key to the success of performing many intelligent tasks by machines. They are used in various applications such as pattern recognition, data classification, stock market prediction, aerospace, weather forecasting, control systems, intelligent automation, robotics, and healthcare. Their architectures generally consist of an input layer, multiple hidden layers, and one output layer. They can be implemented on software or hardware. Nowadays, various structures with various names exist for artificial neural networks, each of which has its own particular applications. Those used types in this study include feedforward neural networks, convolutional neural networks, and general regression neural networks. Increasing the number of layers in artificial neural networks as needed for large datasets, implies increased computational expenses. Therefore, besides these basic structures in deep learning, some advanced techniques are proposed to overcome the drawbacks of original structures in deep learning such as transfer learning, federated learning, and reinforcement learning. Furthermore, implementing artificial neural networks in hardware gives scientists and engineers the chance to perform high-dimensional and big data-related tasks because it removes the constraints of memory access time defined as the von Neuman bottleneck. Accordingly, analog and digital circuits are used for artificial neural network implementations without using general-purpose CPUs. In this study, the problem of fault detection, identification, and location estimation of transmission lines is studied and various deep learning approaches are implemented and designed as solutions.
This research work focuses on the transmission lines’ datasets, their faults, and the importance of identification, detection, and location estimation of them. It also includes a comprehensive review of the previous studies to perform these three tasks. The application of various artificial neural networks such as feedforward neural networks, convolutional neural networks, and general regression neural networks for identification, detection, and location estimation of transmission line datasets are also discussed in this study. Some advanced methods based on artificial neural networks are taken into account in this thesis such as the transfer learning technique. These methodologies are designed and applied on transmission line datasets to enable the scientist and engineers with using fewer data points for the training purpose and wasting less time on the training step. This work also proposes a transfer learning-based technique for distinguishing faulty and non-faulty insulators in transmission line images. Besides, an effective design for an activation function of the artificial neural networks is proposed in this thesis. Using hyperbolic tangent as an activation function in artificial neural networks has several benefits including inclusiveness and high accuracy.
Recommended Citation
Mohammadi Shakiba, Fatemeh, "Artificial neural networks and their applications to intelligent fault diagnosis of power transmission lines" (2022). Dissertations. 1626.
https://digitalcommons.njit.edu/dissertations/1626
