Advanced machine learning methods for damage detection in nonlinear dynamic systems

Structural Health Monitoring (SHM) is an interdisciplinary engineering field dedicated to the autonomous, real-time assessment of structural integrity, with critical applications ranging from civil infrastructure to aerospace, mechanical or more general engineering systems. A primary objective within SHM is the early detection of damage to ensure operational safety and optimize maintenance. In its modern standing, SHM leverages dense arrays of sensors that generate complex, multivariate data streams. These data are often captured under random excitations that constitute real-world operational conditions, and the structures however frequently exhibit nonlinear dynamic behavior. This combination of high-dimensional data and inherent nonlinearity poses a significant challenge, as traditional monitoring techniques based on linear system theory struggle to deduce meaningful insights from such complex responses. While deep learning has emerged as a powerful paradigm uniquely capable of extracting salient features from these datasets, its application has largely been confined to linear systems. This represents an important research gap, compounded by the practical scarcity of labeled data for damaged states, which necessitates the development of robust unsupervised learning methods. This dissertation directly addresses this gap by presenting a systematic exploration of advanced deep learning architectures applied to a suite of challenging nonlinear systems. These benchmarks, subjected to random excitations, encompass numerical models with cubic stiffness, path-dependent hysteresis, and complex hybrid nonlinearities, and are further validated on physical experiments including a magneto-elastic beam and a large-scale lattice tower. To analyze these complex systems, the research follows a sought architectural progression. It starts with an investigation into 1D Convolutional Neural Network (CNN)-based models, where the application of a Convolutional Autoencoder and a Generative Adversarial Network successfully establishes the fundamental viability of using unsupervised feature learning to detect the onset of damage in nonlinear dynamics. This foundational study provides a critical contribution by validating these architectures in a domain where they have been seldom explored. Building upon the insights from the convolutional models, which are powerful but offer limited diagnostic interpretability, the research culminates in the development of the SensorFusion Temporal Fusion Transformer (SensorFusionTFT). This novel and interpretable architecture is specifically adapted for multi-sensor based SHM. This progression from detection to diagnostics is enabled by a robust, multi-metric framework. Damage is identified through a synergistic suite of three distinct metrics: the foundational reconstruction error (RMSE), a highly sensitive Interval Score that quantifies the violation of learned uncertainty bounds, and a novel Attention Anomaly Score derived from the model's internal reasoning. The successful implementation of this framework is demonstrated across all benchmarks, showing a clear advancement in both sensitivity and diagnostic capability. This work thus establishes a complete pathway from foundational unsupervised models to a state-of-the-art, interpretable, and multi-metric Transformer-based solution, providing a powerful and practical framework for the next generation of SHM systems for real-world, nonlinear structures.