Date of Award

Spring 2026

Document Type

Open Access Dissertation

Department

Mechanical Engineering

First Advisor

Yi Wang

Abstract

Anomaly detection plays a pivotal role in ensuring the longevity and safety of engineered systems, particularly in domains where defects can have significant operational or economic consequences. From internal damage in structural parts to subtle visual defects in manufactured components, identifying abnormal patterns requires models that are both computationally efficient and suitable for real-world deployment. This dissertation explores deep learning-based solutions designed to address these challenges, with a focus on Structural Health Monitoring (SHM) and industrial defect inspection under real-world scenarios.

This dissertation first addresses the problem of efficient and reliable component- level perception with limited supervision and strict hardware constraints. The first part of this dissertation presents a lightweight object detection model designed for rail component inspection under limited annotated labels and strict deployment hardware constrains. To address the challenges posed by small and hard-to-distinguish targets such as spike and clips, an adaptively weighted loss (AWL) is introduced within a teacher-student training framework. This dynamic loss adjustment mechanism improves learning efficiency without altering the model architecture or increasing computational cost. Specifically, AWL evaluates the prediction quality of the teacher at each iteration: if the teacher’s output is unreliable, the student reduces its reliance on the teacher by down-weighting the distillation loss; conversely, when the teacher is accurate, the student learns more aggressively. This strategy enhances robustness during training and achieves high detection accuracy while maintaining real-time performance on edge devices such as the Jetson Orin.

Building upon this detection capability, the dissertation then extends from single- task perception to multi-task structural inspection, enabling simultaneous examination of component identity and surface condition within a unified framework. While the previous work targets a single task (detecting railroad components), this part extends the capability to simultaneously detect rail components and segment surface-level defects within a unified framework. A lightweight transfer-learning based architecture is proposed, where a frozen object detector first localizes key elements on the track, and a secondary segmentation network, referred to as TNet, leverages shared features to generate pixel- level defect maps. To compensate for feature misalignment and to maintain computational efficiency, TNet integrates a bias correction module that refines spatial consistency across feature stages and a channel-score attention mechanism that selectively amplifies informative channels while suppressing redundancy. The complete pipeline achieves high accuracy on both tasks and maintains real-time performance on edge platforms, making it well suited for onboard railway inspection under resource constraints.

Moving beyond visual inspection, the dissertation next addresses anomaly detection in time-series structural sensing, where anomaly is manifested as subtle deviations in dynamic signal patterns rather than visible defects. Extending the scope to SHM with time-series signals, this dissertation also focuses on detecting subtle damage patterns from Lamb wave propagation. A novel unsupervised dual-branch adversarial transformer architecture is introduced to process long-range dependencies in ultrasonic signals while enhancing sensitivity to anomalous waveforms. The model consists of a global attention branch that captures the overall waveform structure and a local attention branch that is explicitly constrained by a masked attention mechanism to focus only on short-range temporal contexts. Instead of a conventional training setup, the framework employs an adversarial learning objective: a distribution-level loss encourages the global branch to diverge from local representations, while simultaneously guiding the local branch to align with the global one. This training strategy drives the global attention module to extract robust global features. A segment replacement strategy further improves the global modeling capability of the global attention branch by randomly replacing local signal segments during training, which forces the model to rely on broader temporal context rather than memorizing local patterns. Together, these mechanisms enable the model to identify local anomalous segments that deviate from the global waveform pattern. Since the global attention branch fails to reconstruct these abnormal regions accurately, they result in higher reconstruction loss, which serves as a reliable signal for anomaly localization.

While reconstruction-based transformers provide strong anomaly localization in temporal sensing, industrial inspection often demands fine-grained spatial reasoning and scalable unsupervised segmentation. To address this need, the dissertation further introduces a memory-driven anomaly detection paradigm. To further advance unsupervised-learning based anomaly detection, the final part of this dissertation presents a novel framework based on a Memory and Retrieval Transformer. The model stores representations of normal patterns observed during training. During inference, each input is compared against this learned memory to compute patch-wise deviations, where larger deviations indicate potential anomalies. These deviations are further transformed into an anomaly map through a proposed Anomaly Gradient Interpretation mechanism, which  enhances spatial localization by highlighting regions that contribute most to the deviation. In combination with reconstruction, which inherently fails to reproduce anomalies because only normal patterns are stored, the framework produces accurate and fine-grained anomaly masks. The system operates without any abnormal supervision and generalizes well across diverse industrial inspection scenarios.

Finally, recognizing that many SHM and prognostics tasks require continuous health estimation rather than binary anomaly detection, the dissertation extends the anomaly-learning principle to label-efficient regression under sparse supervision. A Hard- label Combination for Hierarchical Regression (HACHI), a reference–query differential learning framework, is proposed. Instead of directly predicting absolute target values, HACHI transforms sparse scalar supervision into dense relational signals by modeling the difference between paired samples. During training, structured reference-query pairs enable the model to learn relative patterns between them, significantly enlarging effective supervision under low-diversity label distributions. During inference, absolute predictions are recovered by anchoring each query to a known reference. This relational reformulation enhances regression accuracy in weakly supervised settings without relying on discretization or synthetic oversampling.

Future research will advance vision-based anomaly detection toward real-world deployment by improving robustness under complex operating conditions and further reducing reliance on labeled data through stronger unsupervised and self-supervised learning. We will also explore multimodal sensing by fusing RGB with complementary modalities such as depth, thermal, ultrasonic, and Lamb wave signals to better localize subtle or subsurface defects, align spatial visual cues with temporal waveform signatures, and provide more physically meaningful interpretations. Finally, leveraging robotic platforms, we will develop task-driven path planning and active perception strategies that adapt viewpoints based on uncertainty and anomaly likelihood, enabling more efficient and reliable autonomous inspection in safety-critical environments.

Rights

© 2026, Jiawei Guo

Download

Share

COinS