Date of Award

Fall 2025

Document Type

Open Access Dissertation

Department

Mechanical Engineering

First Advisor

Yi Wang

Abstract

Ensuring the security, trustworthiness, and operational integrity of modern autonomous and cyber-physical systems presents a critical challenge. While widely utilized in various engineering applications, such as intelligent transportation and industrial manufacturing, these systems require robust monitoring frameworks to identify unexpected anomalies in real-time, thereby maintaining operational safety and efficiency. This dissertation develops advanced deep learning methodologies for anomaly detection and health monitoring, with a particular emphasis on semisupervised reconstruction-based approaches that identify anomalies in complex environments using models trained only with normal operational patterns.

Building on this theme, the first study focuses on analyzing pedestrian behavior and detecting anomalies at railroad grade crossings. A deep learning framework is developed that integrates pose estimation with skeleton trajectory analysis to capture rich spatiotemporal features of pedestrian motion. Two semi-supervised reconstruction models are proposed: a Generative Adversarial Network (GAN) leveraging adversarial learning on skeleton trajectories, and a Spatial-Temporal Graph Convolutional Network (ST-GCN) combined with an MLP-Mixer to enhance real-time inference efficiency. These models identify abnormal behaviors such as lingering and squatting by evaluating reconstruction errors, exhibiting robust and accurate detection across different camera angles and crossing sites.

Leveraging the reconstruction-based learning methodology established in the first study, the second research topic focuses on developing deep learning models for energy consumption auditing of robotic manipulators. Two deep learning models are developed to learn the energy consumption patterns captured through side-channel measurements, enabling secure and non-intrusive anomaly detection. A semi-supervised GAN framework with both single and multi-discriminator structures captures deviations in temporal energy profiles at the joint level. Meanwhile, a U-shaped Time- Frequency Fusion (U-TFF) network is presented to integrate time- and frequencydomain feature extraction via Fast Fourier Transform, enabling cross-domain feature fusion. Experiments on a multi-joint robotic manipulator demonstrate high accuracy and precision for cyber-physical health monitoring, which confirms the utility of the proposed models for energy-based anomaly detection in manufacturing.

To extend this approach to a broader manufacturing context, the third study specifically addresses the challenges associated with subtle variations in measurement variables in additive manufacturing processes. A transformer-based anomaly detection framework is proposed by incorporating a novel learnable positional embedding to capture both local and global temporal dependencies within energy consumption sequences. Through self-attention mechanisms and reconstruction-based decision metrics, the model detects subtle process anomalies caused by abnormal printing parameters such as nozzle temperature, bed temperature, print speed, and layer thickness. The framework achieves 98% accuracy and real-time performance on edge computing devices. Sensitivity tests are also conducted to verify its ability to identify even minor deviations in operating parameters.

In conclusion, this dissertation research demonstrates a unified, semi-supervised deep learning paradigm for anomaly detection across various domains, including human motion analysis, robotics, and additive manufacturing systems. By combining adversarial learning, graph-based modeling, cross-domain fusion, and transformer-based reconstruction, the present study advances secure, reliable, and deployable anomaly detection for next-generation intelligent engineering systems. Future work focuses on investigating both geographical and demographic information to identify  core factors that contribute to abnormal pedestrian behaviors, such as trespassing at grade crossings. Multimodal sensing, such as vibration, acoustic, or thermal signals,

Rights

© 2025, Ge Song

Download

Share

COinS