Recent Advances in AI and Signal Processing for PZT-Based Structural Health Monitoring

Structural health monitoring (SHM) systems fundamentally rely on effective sensing technologies for reliable damage detection and structural condition assessment. Among the available sensing approaches, piezoelectric (PZT)-based transducers are widely used in civil engineering due to their dual actuation–sensing capability, high sensitivity, low cost, and suitability for real-time monitoring. However, SHM performance not only depends on the sensing hardware, but also on the signal processing techniques that extract meaningful damage-related information from measured responses. Recently, Artificial Intelligence (AI), particularly machine learning (ML) and deep learning (DL), has shown strong potential to enhance automation and improve the performance of SHM systems. This paper provides a critical review of signal processing and data-driven learning approaches for PZT-based guided-wave (GW) SHM and nondestructive testing (NDT), with applications to metallic, composite, and concrete structures. The review covers developments from early ML-based GW SHM methods to recent advances in DL, hybrid frameworks, and physics-informed approaches. Although emphasis is placed on civil infrastructure, developments in other fields such as aerospace and energy engineering are also reviewed due to their role in validating GW-based SHM methodologies. The fundamental theory of PZT sensing and guided wave propagation is introduced to establish the required background for monitoring techniques. Classical signal processing methods are then reviewed, followed by AI-based SHM frameworks, with particular emphasis on hybrid approaches that integrate physics-based signal processing with data-driven models to improve robustness, accuracy, and generalization. Key challenges such as environmental variability, sensor degradation, limited labeled data, and model transferability are discussed, along with future research directions including physics-informed machine learning (PIML), transfer learning, explainable AI, and baseline-free SHM. The review highlights that hybrid and physics-informed frameworks offer strong potential for field deployment by improving robustness, reducing data dependency, and enhancing generalization capability. A key contribution of this work is the comparative synthesis of signal processing, ML, DL, and hybrid methodologies across different material systems and structural types, together with a structured discussion of the challenges and future research directions for real-world implementation.