Recent Advances in Deep Learning and Transfer Learning for Structural Health Monitoring and Condition Monitoring

Dear Colleagues,

Structural health monitoring (SHM) and condition monitoring (CM) have become critical for ensuring the safety and longevity of infrastructure and equipment in civil, mechanical, and energy systems, and aerospace, offshore structures, and other engineering domains. Modern SHM/CM systems integrate advances in materials, sensing, and computation—for example, self-sensing smart materials, embedded sensor and actuator networks, advanced signal and image processing, wireless telemetry, and data fusion—to continuously assess the state of structures and assets. Such transdisciplinary approaches enable on-line detection, localization, and severity assessment of damage or faults in a wide variety of systems (e.g., wind turbines, bridges, buildings, aircraft, solar panels, power grids, and industrial machinery) under variable operating conditions. Indeed, processing and interpreting the massive, heterogeneous data streams from long-term monitoring (e.g., strain, vibration, acoustic, thermal, or visual data) is an urgent challenge that motivates the application of intelligent methods. In particular, advances in data-driven monitoring are enabling autonomous damage/fault diagnosis and predictive maintenance across many sectors.

In recent years, deep learning (DL) has emerged as a powerful paradigm for SHM/CM. Unlike traditional model-based or feature-engineering approaches, deep neural networks can automatically learn hierarchical features from sensor data. This allows DL models to capture complex, nonlinear patterns in structural response that may signal subtle damage or degradation. In SHM contexts, DL has enabled automated defect localization and classification in ways that traditional methods cannot. Current research explores many architectures (e.g., CNNs, RNNs/LSTMs/GRUs, autoencoders, GANs, and hybrids) for tasks such as crack detection, bolt loosening, delamination detection, and anomaly identification. Recently, there is growing interest in physics-informed deep learning, which embeds domain knowledge or physical constraints into neural networks to improve generalization and interpretability. By combining data-driven learning with physical insight, these approaches aim to enhance predictive capability even with limited or variable data. Transfer learning (TL) complements DL by addressing practical SHM/CM challenges such as scarce labeled data and variability across assets. In many monitoring scenarios, it is difficult or unsafe to collect extensive damage data on every structure or system. TL enables models trained on one structure or system to be adapted to another with minimal retraining. Domain adaptation techniques (a branch of TL) explicitly align feature distributions between source and target domains, mitigating issues caused by structural differences or environmental changes. Furthermore, digital twins—virtual replicas that continuously assimilate sensor data from a real asset—can leverage DL and TL to update health predictions in real time and under varying conditions.

This Research Topic seeks to capture these advances and their broad applications. We invite submissions of original research and review articles on all aspects of DL and TL for SHM and CM; for example, novel deep architectures, domain adaptation methods, synthetic data generation, physics-guided networks, and case studies in civil, aerospace, mechanical, renewable energy systems, offshore assets, and other engineering domains.

Dr. Phong B. Dao
Dr. Davide Astolfi
Dr. Liang Yu
Topic Editors