Latent representations and fusion techniques for probabilistic structural health monitoring and state awareness
Loading...
Authors
Fan, Yiming
Issue Date
2025-05
Type
Electronic thesis
Thesis
Thesis
Language
en_US
Keywords
Mechanical engineering
Alternative Title
Abstract
Structural Health Monitoring (SHM) aims to detect and characterize damage in critical components and systems, thus helping prevent catastrophic failures. As infrastructure ages and advanced aerospace and mechanical systems operate under increasingly demanding conditions, the need for reliable SHM solutions continues to grow. This dissertation presents a comprehensive approach that integrates machine learning and signal processing techniques to tackle both forward and inverse SHM challenges. Emphasizing guided waves, vibration-based diagnostics, and multi-fidelity data, the study explores how data compression methods, the fusion of deep learning with time series analysis, and statistical modeling contribute to a more accurate, resilient, and computationally efficient damage assessment framework. By integrating diverse feature extraction methods, probabilistic models, and a multimodal system, this research enhances detection and characterization capabilities, even in complex and uncertain environments. A key motivation lies in the difficulties posed by guided wave signals, which often have high dimensionality and are influenced by multiple environmental and operational conditions (EOCs). The forward problem, in which observed sensor signals must be predicted for a hypothesized damage state, is essential when available data set has limited size especially under extreme conditions. The inverse problem—inferring the damage properties from measured signals—is equally critical, particularly in real-time monitoring contexts where fast and accurate damage estimates can reduce risk and maintenance costs. Traditional forward and inverse approaches often require complex physics-based simulations or extensive data, and can be limited by computational overhead. This dissertation leverages data-driven methods with different feature extraction approaches and time series models to capture essential wave patterns while reducing signal dimensionality. After the introduction of the first chapter, Chapter II examines a machine learning-based framework that addresses both forward and inverse tasks using ultrasonic guided waves in an aluminum plate structure. Various neural network architectures were systematically tested, including multiple layer configurations, differing filter sizes, and alternative data representations designed to handle time-series signals. The latent space generated naturally by this type of model paves the way to efficient computation with reduced data dimension. Different structures of input matrices allow for the creation of time-invariant and time-varying latent space. A detailed comparison among performance of different structures are presented. The models with best-performing structures reconstruct waveforms with high fidelity, indicating that they capture critical features related to damage states. Moreover, when used in an inverse configuration, the models predict structural parameters such as damage extent and external factors simultaneously, illustrating their potential to expedite diagnostic evaluations in practical SHM systems. The approach is assessed via experiments with various model structures on the dual function of the scheme, i.e., to solve forward and inverse problems. The results of the chapter confirm the high potential and effectiveness of using the proposed scheme for multi-purpose SHM missions. Building upon the initial framework, Chapter III focuses on extending this work by comparing various data compression and data expansion techniques on the same lightweight aluminum structure under varying environmental states from the previous chapter. In this scenario, CAEs and diffusion maps (DMaps) both effectively preserves damage-relevant and envorionment-relevant information in the latent space. The near-perfect prediction accuracy leads to a more challenging test case introduced in Chapter IV where a wing is tested in a wind tunnel. Rather than collected in a static condition without significant disturbance, this time the data is recieved with high levels of noise due to airflow and structural complexity. Additional dimensionality reduction and generative methods using VAEs are employed to address the amplified uncertainty. The results reveal that the compression-expansion hybrid approaches can maintain diagnostic accuracy even when ambient conditions vary considerably, demonstrating the importance of adaptable methods for real-world SHM deployments. Recognizing that multiple sensing modalities can further enrich damage detection, the investigation in Chapter V combines information from guided wave data and vibration-based parameters. Rather than directly fusing the high-dimensional signals, it extracts lower-dimensional descriptors—latent representations from autoencoders for the guided waves and autoregressive (AR) coefficients for vibration responses. This parameter-level fusion reduces computational cost while retaining the complementary nature of local wave-based damage sensitivity and global vibration-based structural response. The effectiveness of this fusion is validated through experiments suggesting that properly combined parameters outperform either modality in isolation, especially when noise or other uncertainties challenge single-sensor methods. While these techniques can offer substantial improvements in accuracy and speed, they also rely on sufficient data—whether collected experimentally or generated through simulations. To address this challenge, Chapter VI of the dissertation introduces a multi-fidelity Gaussian Process (GP) model that links a damage index to a damage severity measure. In many SHM contexts, only a limited amount of high-fidelity experimental data is available, whereas simulation data is more plentiful but may not perfectly reflect reality. By employing a hierarchical GP framework, it becomes possible to leverage the benefits of both lower-fidelity simulation data and higher-fidelity experimental measurements. The multi-fidelity GP approach yields a more reliable mapping between measured indices and actual damage levels, thus expanding the feasibility of data-hungry machine learning strategies in practice. Collectively, these five studies illustrate comprehensive approaches to handling complex SHM scenarios. First, CAEs are shown to be effective in producing compact representations for both forward signal generation and inverse condition estimation. Second, the inclusion of diffusion maps provides a detailed comparison between the nonlinear data compression techniques. Third, VAEs expands the range of feasible dimensionality reduction approaches to tackle more disruptive noise environments. Fourth, the fusion of latent features from guided waves with AR-based vibration parameters offers an efficient means of leveraging multiple sensors without incurring the prohibitive cost of raw multi-channel data. Finally, multi-fidelity GP regression addresses the persistent gap between simulation-based modeling and real-world measurements, elevating the predictive accuracy of damage levels across a range of structures. Throughout these contributions, a consistent emphasis is placed on practical considerations, such as computational efficiency, adaptability to different environmental or operational conditions, and scalability to large or complex structures. The methods advanced here pave the way for future intelligent monitoring systems, in which data-driven models respond dynamically to evolving structural states and environmental factors. In addition to aerospace and mechanical engineering, these techniques have relevance for civil infrastructure, automotive safety, and energy sector applications, where the ability to detect damage before critical failure can save resources and lives alike. In summary, this dissertation demonstrates how data-driven approaches such as deep learning and statistical modeling, when carefully integrated, can overcome well-known bottlenecks in guided wave and vibration-based SHM. By combining forward models for signal generation, inverse methods for damage quantification, multi-modal data fusion, and multi-fidelity inference strategies, the work aims to realize accurate, robust, and operationally efficient monitoring solutions. The results confirm that these novel approaches enable more reliable and fine-grained assessments of structural health, even under conditions of strong noise or limited high-fidelity data. Such advances mark an important step toward practical deployment of next-generation SHM systems across a broad spectrum of industrial and research sectors, ultimately reinforcing the integrity and safety of critical structures worldwide. Finally, Chapter VII presents the concluding remarks and outlines future directions for this dissertation.
Description
May2025
School of Engineering
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY