etallic structures are the key backbone of the society and economy, which are often subjected to different types of loadings resulting cracking, corrosion, and other material discontinuity, and affecting structural integrity and safety. Therefore, ultrasonic guided wave (UGW) has been widely used for structural health monitoring (SHM) to gain a deep understanding of structural performance, assess the current state of structural conditions, and avoid potential catastrophic events. Despite advances in technologies and methods in data process, microdamage detection still posts great challenges in their detectability. Different from conventional physics-based methods, artificial intelligence and machine learning (AI/ML) has recently fueled profound automation solutions toward signal process and data fusion, thereby dramatically overcoming the limits. Along this vein, this study aims to propose AI-empowered SHM framework by decoding the UGW to uncover complex interconnected information among data, models, uncertainty, and risk for enhanced structural diagnosis and prognosis to improve metallic structural integrity and safety. Several structural cases, from one-dimensional plates/rods to three-dimensional pipes, were deliberately selected to demonstrate the real-world applications. Three different levels of the AI/ML approaches, from shallow learning to deep learning, are used to explore the effectiveness of the data fusion and data representation. Meanwhile, noise interference and structurally initial nonlinearity as typical structural uncertainty are included in data collection to understand the effects of data quality and uncertainty on the robustness of the proposed methods. The results showed that the proposed method was an efficient and accuracy way to identify the damage characteristics. Results from the shallow learning demonstrated that different features had certain levels of sensitivity to damage, while the feature selection method in the shallow learning revealed that time-frequency features and wavelet coefficients exhibited the highest damage-sensitivity. However, with the increase of noise level, the shallow learning failed in detectability. By taking advantage of higher automation in feature extraction, the deep learning exhibited significant improvement in accuracy, robustness, and reliability for structural diagnosis and prognosis. Particularly, the higher-layer architecture could outperform the shallow learning in terms of higher effective and efficient data fusion, and enhanced their capability in decoding information over noise interference and structural uncertainty.