DEEP LEARNING FOR MODAL PARAMETER IDENTIFICATION IN STRUCTURAL HEALTH MONITORING: A REVIEW OF ARCHITECTURES, TRANSFER LEARNING, AND PHYSICS-INFORMED APPROACHES

Abstract

Deep learning has revolutionized modal parameter identification in structural systems, enabling automated extraction of natural frequencies, damping ratios, and mode shapes from vibration data. This review synthesizes 45+ peer-reviewed studies examining convolutional neural networks (CNNs), recurrent neural networks (RNNs), hybrid architectures, and transfer learning approaches. CNN-based models achieve 95-99% accuracy in fault detection, while hybrid CNN-LSTM architectures demonstrate superior temporal-spatial feature extraction. Transfer learning reduces training requirements by 75-85% across domains. Key findings reveal that physics-informed neural networks bridge data-driven and mechanistic approaches, explainable AI enhances interpretability, and multi-modal data fusion improves diagnostic reliability. Implementation challenges include computational cost, real-time processing constraints, and domain adaptation under variable operating conditions. Future directions focus on edge computing deployment, federated learning for distributed monitoring, and quantum-enhanced algorithms. This review provides researchers and practitioners with comprehensive guidance on architecture selection, preprocessing strategies, and practical deployment considerations