ORCID Number

https://orcid.org/0000-0002-4532-8814

Date of Award

Summer 2026

Access Type

Thesis - Open Access

Degree Name

Doctor of Philosophy in Electrical Engineering & Computer Science

Department

Electrical Engineering and Computer Science

Committee Chair

Berker Pekoz

Committee Chair Email

pekozb@erau.edu

Committee Advisor

Berker Pekoz

Committee Advisor Email

pekozb@erau.edu

Committee Co-Chair

Radu F. Babiceanu

Committee Co-Chair Email

radu.babiceanu@wmich.edu

First Committee Member

Eduardo Rojas

First Committee Member Email

rojase1@erau.edu

Second Committee Member

Tianyu Yang

Second Committee Member Email

yang482@erau.edu

Third Committee Member

Laxima Niure Kandel

Third Committee Member Email

niurekal@erau.edu

Fourth Committee Member

Radu F. Babiceanu

Fourth Committee Member Email

radu.babiceanu@wmich.edu

Fifth Committee Member

Berker Pekoz

Fifth Committee Member Email

pekozb@erau.edu

College Dean

James W. Gregory

Abstract

The transition toward More Electric Aircraft (MEA) has introduced highly complex electrical architectures that impose strict requirements on reliability, safety, and real‑time operation. Yet most existing research on power quality disturbances (PQDs) and electrical fault diagnosis targets conventional utility‑scale power grids and relies on low‑frequency analysis, which limits accuracy and applicability in aircraft electrical systems that operate at higher frequencies. As a result, the use of data‑driven PQD classification in aircraft power networks remains largely unexplored. This paper addresses this gap by presenting a deep learning–based framework for automated multiclass detection and classification of electrical faults and PQDs in aircraft electrical systems with emphasis on classification metrics, robustness, and applicability under aerospace constraints. A high‑fidelity aircraft power system model inspired by the Boeing 787 electrical architecture was developed to represent operation at a 400Hz fundamental frequency. The model produced high‑resolution signals under a comprehensive set of fault and PQD conditions. Two datasets were produced. The first dataset consists of one‑dimensional time‑domain signals, further augmented through signal processing techniques and generative adversarial networks (GANs) to increase data diversity and robustness; this dataset is made publicly available on IEEE DataPort and supports the evaluation of one‑dimensional convolutional neural network (1D‑CNN) models. The second dataset consists of two‑dimensional time–frequency representations derived from the short‑time Fourier transform and targets two‑dimensional convolutional architectures (2D-CNN). Several deep learning architectures were evaluated, including 1D‑CNNs, 2D-CNNs, LSTMs, hybrid CNN–LSTM models, and established deep convolutional networks such as ResNet, MobileNet, and VGG. A compact ResNet architecture demonstrated the most favorable balance between classification performance and model complexity, achieving a software test accuracy of 96.94\% with only 175,685 parameters. After 8-bit quantization and deployment on a Xilinx Zynq UltraScale+ MPSoC ZCU102, the FPGA accelerator achieved a post-quantized accuracy of 95.87\%, with a mean on-board inference latency of 6.90 ms and a worst-case on-board latency of 14.40 ms per classifier input record. The implementation used 15.52\% of DSP slices, 63.71\% of Block RAM, and 93.77\% of CLB LUTs, indicating real-time feasibility for the deep-learning inference accelerator while leaving limited LUT margin for additional on-FPGA integration. The findings show that lightweight deep learning models can provide accurate multiclass classification under aerospace-relevant resource and timing constraints.

Additional Files
GS9_Acceptance_Guzman_LNK.pdf (430 kB)
Download

Share

COinS