Fatigue behavior and vibration-based integrity management of advanced composite structures: A comprehensive review

Abstract

Composite materials are widely used in aerospace, automotive, and marine engineering because of their high strength-to-weight ratio and design flexibility. However, their long-term reliability is limited by fatigue damage, which develops gradually through matrix cracking, delamination, and fiber fracture under cyclic loading. Conventional fatigue-life prediction methods often fail to capture early degradation and typically require extensive experimental calibration. Vibration-based Structural Health Monitoring (SHM) offers a non-destructive alternative by exploiting the direct relationship between stiffness degradation and changes in dynamic properties such as natural frequencies, damping ratios, and mode shapes. This review synthesizes current research linking vibration response to fatigue damage in composite structures. Experimental methods, sensing technologies, and computational modeling approaches are examined within a unified framework. Particular attention is given to sandwich and auxetic configurations, environmental influences, and design optimization for improved fatigue resistance. The study also discusses emerging data-driven paradigms, including physics-guided machine learning and digital-twin concepts, which enable predictive maintenance rather than post-damage detection. By integrating mechanics-based understanding with monitoring and data analytics, this review provides a structured perspective on vibration-based integrity assessment and outlines research directions toward reliable in-service monitoring and extended operational life of advanced composite structures.

1. Introduction

Fiber-reinforced composite materials have become indispensable structural components in aerospace, automotive, marine, and energy applications due to their high strength-to-weight ratio and the ability to tailor mechanical properties through stacking sequence and architecture design. Unlike metallic structures, which typically fail through propagation of a dominant crack, composite structures experience distributed and progressive degradation. Matrix cracking, fiber–matrix debonding, delamination, and fiber fracture accumulate gradually and produce a continuous reduction in global stiffness rather than sudden catastrophic failure.

Because structural vibration is governed by stiffness, fatigue-induced degradation directly alters dynamic behavior. Natural frequencies, mode shapes, and damping ratios evolve continuously as damage accumulates. This establishes vibration response as a physically interpretable indicator of structural condition rather than a purely empirical monitoring signal. The mechanical relationship between damage progression, stiffness reduction, and modal perturbation forms the physical foundation of vibration-based structural health monitoring.

Research on composite structural behavior has historically developed across several partially independent directions, including vibration modeling of laminated structures, fatigue damage mechanics, sandwich-structure dynamics, sensing technologies, and data-driven analysis methods. Instead of treating these developments separately, the present review groups them into coherent research domains summarized in Table 1. These domains collectively describe how structural behavior, monitoring techniques, and predictive modeling interact in integrity assessment.

Table 1.

Overview of composite structure research domains and representative studies.

Composite domain	Key characteristics	Representative references
General Laminated Composites	Linear and nonlinear vibration behavior, bending response, mesh-free and finite-element formulations (FEM, kp-Ritz), shear deformation theories, geometric singularities, and strength/buckling optimization	1–10
Fatigue and Damage Mechanics	Fatigue life prediction, residual strength evaluation, stiffness and natural frequency degradation models, very-high-cycle fatigue behavior, stochastic loading effects, progressive damage evolution	11–32
Honeycomb Sandwich Structures	Influence of core geometry (hexagonal, re-entrant/auxetic, corrugated), nonlinear vibration response, analytical and numerical modeling, damping and impact resistance, blast performance, structural optimization	2,33–76
Vibration-Based SHM and Damage Detection	Damage localization using modal properties (frequency, damping, mode shapes), operational modal analysis, nonlinear resonance spectroscopy, debonding identification, fiber Bragg grating sensing, impact monitoring	19,32,77–100
Machine Learning and AI Applications	Physics-informed learning, deep learning-based damage detection, autonomous recognition, hybrid predictive models, transfer learning, explainable artificial intelligence, limited-data strategies	31,32,78,96–99,101–110
Manufacturing and Processing Effects	Ultrasonic vibration-assisted machining of honeycomb composites, cutting-force prediction, manufacturing-induced performance variation, additive manufacturing of composite structures	59,111–114
Damping and Acoustic Behavior	Damping ratio estimation, vibration energy dissipation, sound radiation and insulation, viscoelastic fillers, phononic crystal-based vibration reduction	43,51,58,71,73,75,115–120
Multifunctional & Specialized Structures	Functionally graded materials (FGM), shape memory alloys (SMA), CNT-reinforced composites, hybrid (riveted/bonded) joints, FRP-concrete interaction, microwave absorption, aerospace/automotive applications.	2,7,10,54,57,58,67–69,71,74,76,78,117,121–123

Advances in sensing technology have enabled measurement of structural dynamics during actual service rather than controlled laboratory conditions. Operational modal analysis, distributed optical fiber sensing, acoustic emission monitoring, and full-field optical measurements now allow continuous observation of dynamic response under environmental variability. These developments demonstrated that operational factors such as temperature, stochastic loading, and boundary uncertainty significantly influence modal features and must be incorporated into diagnostic interpretation.

The combination of measurable modal features and physics-based degradation models enabled a transition from damage detection toward structural condition assessment. Progressive damage models, probabilistic formulations, and multi-physics simulations link stiffness degradation to remaining useful life. Simultaneously, optimization and manufacturing processes were shown to influence both fatigue resistance and dynamic response, indicating that monitoring cannot be separated from structural design.

More recently, high-dimensional vibration measurements have motivated the use of machine learning and hybrid physics-informed methods. Data-driven approaches now aim not only to detect damage but also to predict structural life under uncertain operating conditions. These developments indicate a paradigm shift from inspection-based maintenance toward predictive integrity management frameworks.

Despite extensive research activity, fatigue mechanics, sensing technologies, structural dynamics, and artificial intelligence are still frequently studied as independent topics. In reality, vibration-based integrity assessment follows a single physical chain:

Damage evolution → stiffness degradation → modification of the global dynamic operator → measurable modal features → prognosis.

This review therefore organizes the literature according to this causal mechanical relationship. Rather than presenting isolated methodologies, experimental sensing, computational modeling, and data-driven prediction are interpreted as consecutive layers of one unified integrity-management framework. This perspective enables a transition from detection-oriented monitoring toward reliable life prognosis and digital-twin-assisted structural management of advanced composite structures. Overview of composite types and their fatigue/vibration characteristics discussed in this review listed in Table 1.

Unlike existing review articles that typically address composite fatigue behavior, vibration-based damage detection, sensing technologies, or machine learning methods as separate research domains, the present work organizes the field around a single causal structural mechanics chain: damage evolution → stiffness degradation → modification of the global dynamic operator → measurable modal features → data-driven prognosis. This physics-consistent perspective enables the literature to be reinterpreted not as independent methodological developments but as components of one unified integrity-management framework. Consequently, experimental sensing techniques, computational modeling approaches, and artificial-intelligence-based prediction methods are evaluated according to how they interact within this chain rather than as isolated technologies.

The contribution of this review is therefore not merely the compilation of recent studies, but the establishment of a hierarchical and physically interpretable taxonomy that links material damage mechanisms to measurable dynamic signatures and finally to predictive maintenance strategies. This integrative viewpoint provides a structured roadmap for transitioning vibration-based monitoring of composite structures from detection-oriented diagnostics toward reliable remaining-life prognosis and digital-twin-enabled structural management.

2. Fatigue damage mechanisms and modelling in composites

Composite laminates fail under cyclic loading through a cascade of micro mechanisms that differ fundamentally from those in metals. Understanding these mechanisms, the traditional life prediction frameworks, and the influence of environment and test standards is essential before interpreting any modal frequency response.

2.1. Damage mechanisms and residual strength models

Fatigue in composites is a progressive, multi-stage process involving synergistic interaction of multiple damage modes. Unlike metals where a dominant crack propagates, composites exhibit distributed damage: initial matrix cracking in off-axis plies, followed by crack coupling, delamination at ply interfaces, fiber-matrix debonding, and eventually fiber breakage.^11,30 This sequence leads to gradual stiffness reduction, which forms the physical basis for vibration-based monitoring. The residual strength approach, reviewed by Philippidis & Passipoularidis, ¹¹ models strength degradation under fatigue loading using phenomenological models. Their comparative study of various models applied to C/Ep and Gl/Ep laminates revealed that no single model consistently predicted residual strength accurately across different materials and loading conditions. They concluded that simpler models requiring limited experimental data may be preferable to complex phenomenological ones. This underscores the need for alternative, globally sensitive indicators like vibration parameters.

2.2. Progressive damage and life prediction models

Progressive damage modeling (PDM) simulates the sequential failure of composite constituents through cycles, incorporating stress analysis, damage evaluation, and property degradation (Figure 1 as an experiment). Rafiee & Elasmi ¹⁸ employed PDM based on stiffness degradation to predict fatigue lifetime of composite pipes under internal cyclic pressure, validating against experimental data. They further extended this to stochastic fatigue analysis of GFRP pipes under variable amplitude loading, treating loading parameters as random variables. ²³ Similarly, Peng et al. ²² developed a stiffness degradation model for in-situ fatigue life prognosis of composites using Bayesian inference with piezoelectric sensor data. Akbulut ⁹ optimized progressive failure behavior of notched composite plates using a zeroth-order optimization tool, demonstrating capability for exact determination of maximum stiffness and minimum strain.

Figure 1.

Damage mode of CFRP specimens (a) E specimen (b) F specimen. ⁸⁶

The Fawaz-Ellyin parametric fatigue life prediction model has been extensively coupled with optimization algorithms for laminate design. Ertas & Sonmez ¹² integrated it with Direct Simulated Annealing (DSA) to find optimal fiber orientations for maximum fatigue life under in-plane loads. Subsequent work employed Particle Swarm Optimization (PSO) ¹⁴ and a variant of simulated annealing ¹⁵ for similar objectives. Ertas ¹⁷ further validated the Fawaz-Ellyin model for optimization problems and presented improved methodologies considering experimental results.

For very high cycle fatigue (VHCF), Wu et al. ¹⁶ proposed a natural frequency degradation model for woven fabric composites based on matrix crack density evolution, verified through on-line frequency measurements. Huang et al. ²⁰ combined stiffness degradation with infrared thermography (IRT) to predict S-N curves for CFRP laminates, using stabilized temperature rise to determine fatigue limits. These approaches demonstrate the growing integration of vibration-based metrics with traditional fatigue modeling and summarized in Table 2.

Table 2.

Summary of key fatigue life prediction models and their characteristics.

Model type	Key features	Limitations	Representative refs
Residual Strength	Phenomenological, strength degradation curves	Limited consistency across materials	11
Progressive Damage	Simulates damage sequence, stiffness degradation	Computationally intensive	9,18,23
Stiffness Degradation	Links stiffness loss to cycles, vibration-based	Requires stiffness-cycle relationship	16,20,22
Fawaz-Ellyin	Parametric, coupled with optimization	Needs model calibration	12,14,15,17
Machine Learning Hybrid	Data-driven, accounts for complex interactions	Requires large datasets	78,83

2.3. Environmental and stochastic effects

Environmental factors significantly influence fatigue behavior. Zhang et al. ⁷⁸ addressed hygrothermal-load coupling effects on CFRP/Al joints using a physics-guided machine learning (PGML) framework, combining a nonlinear damage accumulation model with residual neural networks. Temperature effects on natural frequencies of composite girders were investigated by Poudel et al., ⁸⁹ showing linear relationships enabling temperature effect separation. Rafiee^18,23 emphasized the importance of stochastic modeling for real operational conditions, where loading parameters exhibit inherent randomness.

Ma et al. ⁶⁰ studied fatigue of composite honeycomb sandwich panels under random vibration load, identifying core shear failure and face-core delamination as dominant modes. They proposed a life prediction method based on core shear stress, validated against test results. These studies highlight the necessity of accounting for environmental and operational variability in fatigue assessment.

3. Vibration fundamentals and damage-sensitive features

The reduction of a structure’s natural frequencies under cyclic loading is a direct manifestation of stiffness degradation. Because the eigen-value problem that governs vibration is mathematically coupled to the global stiffness matrix, any damage that lowers K will be reflected in the modal spectrum.

The dynamic response of a structure is governed by its mass [M], stiffness [K], and damping [C] matrices. Fatigue damage primarily reduces stiffness, leading to measurable changes in natural frequencies fi, mode shapes{ϕ}i, and damping ratios ζi. The fundamental relationship can be expressed in Eq. (1).

$\frac{\mathrm{\Delta }{f}_i}{f_i}\approx -\frac{1}{2}\left(\frac{\mathrm{\Delta }{K}_i}{K_i}\right)$

(1)

The modal parameters are obtained by solving the eigenvalue problem derived from Eq. (2).

$\left[M\right]\left\{u^{¨}\right\}+\left[C\right]\left\{u^{\dot{}}\right\}+\left[K\right]\left\{u\right\}=\left\{F\left(t\right)\right\}$

(2)

Equation (1) originates from the first-order perturbation of the structural eigenvalue problem. Since natural frequencies are proportional to the square root of the stiffness-to-mass ratio (f ∝ √(K/M)), a small reduction in stiffness due to fatigue damage produces a proportional frequency decrease approximately equal to half of the relative stiffness loss. This relation establishes modal frequency as a physically interpretable global damage indicator rather than an empirical correlation.

For free vibration (F(t)=0) and negligible damping, Eq. (2) reduces to the classical eigenvalue problem in Eq. (3)

$\left[K\right]\left\{{\varphi }_i\right\}={\omega }_i^2\left[M\right]\left\{{\varphi }_i\right\}$

(3)

from which modal parameters are obtained. Therefore, any localized material degradation modifies the global stiffness matrix, perturbs the eigenvalues ω_i, and becomes observable as measurable shifts in modal frequencies and mode shapes. This provides the theoretical basis for vibration-based structural health monitoring ¹²⁴ .

Operational Modal Analysis (OMA) extracts modal parameters from output-only vibration data under operational conditions, crucial for in-service monitoring. Pacheco-Chérrez et al. ⁷⁹ used OMA with Stochastic Subspace Identification (SSI) for damage detection in rotating wind turbine blades under noisy conditions, reliably detecting small cracks down to 40 dB SNR. They further applied OMA with Continuous Wavelet Transforms (CWT) for damage detection in composite and plastic thin-wall beams, ⁸⁵ achieving sub-grid resolution accuracy.

Van Vondelen et al. ⁹⁵ reviewed OMA techniques for damping identification of offshore wind turbines (OWTs), addressing challenges posed by rotor speed harmonics near structural modes. They classified algorithms by suitability criteria and identified future paths combining harmonic isolation with statistical methods.

For composite structures, Zhang et al. ⁸⁰ employed frequency shifts with artificial neural networks (ANN) and surrogate-assisted genetic algorithms (SAGA) to detect delamination in curved composite plates (Figure 2), with SAGA showing better performance. Khan et al. ¹⁰¹ developed a deep learning framework using convolutional neural networks (CNN) on vibration spectrograms to assess delamination in smart composite laminates, achieving 94.5% test accuracy.

Figure 2.

Mode shape changes due to delamination in curved composite plates, (a) no delamination, and (b) mid-plane square delamination of 1/3 width of plate. ⁸⁰

4. Experimental techniques and sensor technologies for SHM

4.1. Conventional modal testing

Traditional modal testing using accelerometers and impact hammers remains widely employed. Abo-Elkhier et al. ²¹ used impact hammer excitation with frequency response function (FRF) measurements on GFRP beams to track damping and frequency changes during fatigue. Sadiq et al. ³⁴ conducted free vibration tests on aluminum honeycomb sandwich panels with simply supported boundaries using accelerometers and impact hammer, interfaced with LabVIEW for data acquisition.

Suryawanshi et al. ⁸¹ utilized Fast Fourier Transform (FFT) analyzer to identify artificial voids in honeycomb composite beams, showing decreased natural frequencies with increased defect percentage. These methods provide reliable global dynamic characterization but may lack spatial resolution for local damage.

4.2. Advanced sensing: FBG, AE, DIC, and optical fibers

4.2.1. Fiber Bragg Grating (FBG) sensors

Nicolas et al. ⁸⁸ deployed 780 FBG sensors (Figure 3) on a 5.5-m carbon-composite aircraft wing to determine deflected shape and out-of-plane loads, with calculated displacements within 4.2% of measurements.

Figure 3.

FBG sensor array configuration on composite aircraft wing for load and shape determination. ⁸⁸

Liang et al. ⁸⁴ used FBGs for strain and temperature monitoring of composite tanks under cryogenic conditions, proposing temperature compensation based on thermal stratification. Sánchez-Botello et al. ⁹³ embedded FBGs in a rotating shaft-disc assembly to measure strain distribution and identify natural frequencies and mode shapes in air and water.

4.2.2. Acoustic Emission (AE) and Digital Image Correlation (DIC)

Wang et al. ⁸⁶ combined AE and DIC to monitor fatigue damage in CFRP open-hole laminates, using AE hit rate and absolute energy to characterize stiffness degradation and principal component analysis with K-means++ to identify damage modes. Molina-Viedma et al. ⁹⁴ employed high-speed 3D DIC for full-field operational modal analysis of an aircraft composite panel after multi-impact excitation, characterizing multiple modes in a wide spectrum.

4.2.3. Embedded optical fibers and other sensing modalities

Batte et al. ⁷⁷ embedded optical fibers in graphite/epoxy laminates to characterize delaminations from low-velocity impacts, correlating strain distribution with ultrasonic C-scans. This demonstrates the potential for integrated health monitoring. Alqam & Dhingra ⁸⁷ investigated indirect load identification using optimum placement of strain gages and accelerometers based on frequency response functions. Chakrapani & Barnard ²⁵ applied nonlinear resonance spectroscopy to detect early fatigue damage in carbon/epoxy laminates, tracking nonlinear frequency shift and modal damping ratio. Advantages and limitations are depicted in Table 3.

Table 3.

Comparison of sensing technologies for vibration-based SHM of composites.

Technology	Key advantages	Limitations	Typical applications	Refs
Accelerometers	Standardized, wide frequency range	Point measurements, wiring needed	Global modal analysis	21,34
FBG Sensors	Multiplexing, immunity to EMI, small size	Temperature sensitivity, specialized equipment	Distributed strain, shape sensing	84,88,93
AE Sensors	High sensitivity to active damage	Noise susceptibility, complex signal processing	Damage initiation and growth	86
DIC	Full-field, non-contact	Requires surface preparation, computational cost	Strain fields, mode shapes	94
Embedded Optical Fibers	Integrable, sensitive to internal damage	Manufacturing complexity, fragility	Internal damage detection	77

4.3. Nonlinear and full-field vibration methods

Nonlinear vibration techniques exploit the nonlinear behavior of damage interfaces. Chakrapani & Barnard ²⁵ used nonlinear resonance spectroscopy and slow dynamic diagnostics on quasi-isotropic CFRP laminates under 4-point bending fatigue, showing non-monotonic increase in nonlinearity with fatigue cycles. Molina-Viedma et al. ⁹⁴ demonstrated full-field modal analysis from impact response using DIC, offering comprehensive mode shape characterization without sensor arrays.

5. Computational and analytical modelling of composite dynamics

The substantial volume of fatigue-induced modal frequency data generated through the experimental programme (Section 4), together with complementary physics-based modeling efforts, has catalyzed a rapid expansion of data-driven methodologies. These approaches can be broadly categorized into two principal paradigms: (i) the extraction and engineering of physically meaningful features from raw modal responses, followed by their mapping to damage-sensitive indicators; and (ii) end-to-end learning frameworks that seek to directly infer the underlying frequency–damage relationship while explicitly addressing challenges associated with measurement noise, environmental variability, and limited availability of labeled data. The following sections structure the current state of the art into a coherent methodological pipeline, spanning feature construction, model development, and deployment within real-time digital twin–enabled prognostic frameworks. In doing so, they also identify critical methodological limitations and open challenges that motivate continued research in vibration-based fatigue assessment and integrity management of composite structures.

5.1. Finite element and smoothed techniques

Finite Element Analysis (FEA) is the cornerstone for simulating vibration response of complex composite structures. Cui et al. ³ developed a novel triangular composite plate element based on edge-based smoothing technique for bending and vibration analysis of laminated plates, employing discrete shear gap (DSG) method to mitigate shear locking. Dai et al. ¹ presented a mesh-free method using moving least-squares shape functions for static and free vibration analysis of laminated plates based on third-order shear deformation theory (TSDT).

Bletzinger et al. ⁴ introduced the Discrete Shear Gap (DSG) method for locking-free finite elements for shear deformable plates and shells, applicable to triangular and rectangular elements of arbitrary polynomial order. Onyibo & Safaei ³⁷ reviewed FEA applications to honeycomb sandwich structures using ANSYS and ABAQUS, providing guidelines for researchers.

For progressive damage simulation, Akbulut ⁹ optimized failure behavior of notched composite plates using FEA with a progressive failure approach. Akbulut et al ⁵ compared simulated annealing (SA), genetic algorithm (GA), and particle swarm optimization (PSO) for optimal stacking sequence of laminated plates with respect to buckling load and natural frequencies, finding PSO outperformed others..

5.2. Higher-order and shear deformation theories

Higher-order shear deformation theories (HSDT) account for through-thickness shear effects crucial for thick composites. Lei et al. ⁷ used first-order shear deformation theory (FSDT) with kp-Ritz method for free vibration analysis of laminated FG-CNT reinforced rectangular plates. Al-Khazraji et al. ⁴⁹ applied simplified HSDT to laminated honeycomb sandwich panels (LHSP), deriving four vibration differential equations and validating with Nayak and Meunier models and ABAQUS.

Li et al. ² employed FSDT and von Kármán geometrical nonlinearity for nonlinear free vibration of functionally graded fiber-reinforced composite hexagon honeycomb sandwich cylindrical shells, using Hamilton’s principle and multiple scale expansion method. Eipakchi & Nasrekani^47,48 used FSDT and Hamilton’s principle for vibrational analysis of composite cylindrical shells with auxetic honeycomb core under moving pressure and for asymmetric free vibration with adjustable Poisson’s ratio.

5.3. Analytical and semi-analytical methods

Analytical models provide efficient solutions for specific geometries. Li et al. ²⁰ analyzed vibration characteristics of fiber-reinforced composite sandwich (FRCS) cylindrical-spherical combined shells with hexagon honeycomb core (HHC) using FSDT, multi-segment decomposition, virtual spring technology, Jacobi-Ritz approach, and transfer function method. They found vibration suppression enhanced by reducing HHC thickness ratio and cell wall thickness ratio.

Zhang et al. ⁵³ derived equivalent elastic parameters for zero Poisson’s ratio honeycomb cores and analyzed free vibration and flutter of sandwich plates under supersonic flows using first-order piston theory. Dong et al. ²⁵ developed an analytical model for vibration of composite sandwich cylindrical shells with corrugated-honeycomb blended cores in inhomogeneous thermal environments using macroscopic homogeneity approach and Rayleigh-Ritz method.

For auxetic structures, Liu et al. ⁴¹ investigated nonlinear vibrations of auxetic honeycomb composite plates using modified Gibson function and multiple scale method. Peng et al. ⁴⁰ analyzed free flexural vibration of composite sandwich plates with reentrant honeycomb cores using 2D homogenized plate model via variational asymptotic method.

6. Design, optimization, and manufacturing for enhanced fatigue performance

6.1. Optimization of laminates and structures

Optimization algorithms are increasingly used to design composites for improved fatigue and dynamic performance. Ertas & Sonmez ¹² coupled the Fawaz-Ellyin fatigue model with Direct Simulated Annealing (DSA) to optimize fiber orientation angles for maximum fatigue life under in-plane loads. They extended this to Particle Swarm Optimization (PSO)^14,125 and a simulated annealing variant, ¹⁵ demonstrating effectiveness across problems. Akbulut et al. ⁵ compared SA, GA, and PSO for optimal stacking sequence of laminated plates to maximize critical buckling load and natural frequencies, finding PSO outperformed others. Akbulut ⁹ optimized progressive failure behavior of notched composite plates using Nelder-Mead algorithm for maximum stiffness and minimum strain.

For damping optimization, Kern et al. ¹¹⁶ optimized viscoelastic properties of unidirectional carbon fiber/epoxy composites with coated fibers using finite element models and Hervé-Zaoui micromechanical model, finding maximum loss for ∼10 nm coating thickness. Maneengam et al. ⁵⁸ used finite element method coupled with genetic algorithm to identify optimal ply orientations of CFRP sandwich shells with MWCNT/GFRP honeycomb for enriched natural frequencies and loss factors.

6.2. Influence of machining and processing

Manufacturing processes significantly affect structural integrity and dynamic performance. Zarrouk et al.^111,113 developed 3D finite element models for rotary ultrasonic machining (RUM) of Nomex honeycomb composites (NHCs), showing ultrasonic vibrations reduce cutting forces up to 50% and improve surface quality. Mughal et al. ¹¹⁴ experimentally validated RUM of NHC structures, finding cutting forces decrease with increased ultrasonic amplitude and spindle speed, and increase with depth of cut and feed rate.

Cao et al. ¹¹² established a finite element model for cutting force prediction in ultrasonic vibration-assisted machining of paper honeycomb materials with disc-cutter, validated with maximum 13.2% error. These studies (Table 4) highlight how processing parameters influence final structural performance.

Table 4.

Optimization algorithms applied to composite design for fatigue and vibration performance.

Algorithm	Key features	Applications	Performance	Refs
Direct Simulated Annealing (DSA)	Stochastic global search, avoids local minima	Fatigue life maximization	Effective for laminate optimization	6,12
Particle Swarm Optimization (PSO)	Metaheuristic, population-based	Stacking sequence, fatigue life	Outperforms SA and GA in studies	5,14
Genetic Algorithm (GA)	Evolutionary, crossover/mutation	Multi-objective optimization	Good but computationally intensive	5
Nelder-Mead	Zeroth-order, simplex-based	Progressive failure optimization	Suitable for stiffness/strain objectives	9
Response Surface Methodology (RSM)	Statistical, design of experiments	Parameter optimization	Efficient for multi-variable problems	34,54

7. Dynamics of advanced composite structures: Case studies

7.1. Honeycomb and sandwich structures

Honeycomb sandwich structures are ubiquitous in aerospace for their high stiffness-to-weight ratio. This section classifies honeycomb cores based on their geometric configuration and mechaical response, followed by a dedicated discussion of failure mechanisms in Section 7.1.5. A general view of the selected studies on honeycomb cores are shown in Table 5.

Table 5.

Vibration characteristics of different honeycomb core types.

Core type	Key features	Natural frequency trend	Damping characteristics	Representative refs
Regular Hexagonal	Isotropic in-plane, high shear stiffness	Increases with core height, decreases with cell size	Moderate damping	34,36,60
Auxetic (Reentrant)	Negative Poisson’s ratio, energy absorption	Lower than regular honeycomb, decreases with cell angle	In most cases, enhanced energy dissipation, though this is dependent on specific geometric configurations 76	40,41,44
Functionally Graded	Tailored property distribution	Controllable via gradation parameters	Can be optimized for damping	2,53
Corrugated-Honeycomb Blended	Combined bending and shear resistance	Higher than single core types	Improved vibration suppression	63,76

7.1.1. Conventional honeycomb cores (non-auxetic)

This category encompasses cores with positive Poisson’s ratio, including standard hexagonal geometries. Sadiq et al. ³⁴ optimized vibration characteristics of aluminum honeycomb sandwich panels using Response Surface Methodology (RSM), finding maximum natural frequency (1665.7 Hz) at 25 mm core height, 25 mm cell size, and 0.2 mm wall thickness. Zhou et al. ³⁶ analyzed vibration and aeroelastic stability of hexagonal honeycomb core sandwich panels in supersonic airflow, showing core thickness majorly influences critical dynamic pressure.

7.1.2. Auxetic honeycomb cores (negative Poisson’s ratio)

Auxetic cores exhibit negative Poisson’s ratio, leading to unique deformation characteristics. Liu et al. ⁴¹ investigated nonlinear vibrations of auxetic honeycomb composite plates using modified Gibson function, finding natural frequencies decrease with increased cell angle regardless of Poisson’s ratio sign. Peng et al. ⁴⁰ showed natural frequencies of reentrant honeycomb sandwich plates are lower than traditional honeycomb due to negative Poisson’s ratio effect. Quoc et al. ⁴⁴ analyzed free vibration and dynamic response of sandwich plates with auxetic honeycomb core using smoothed finite element model based on FSDT.

7.1.3. Functionally graded and hybrid cores

This category includes cores with spatially tailored properties or hybrid material compositions. Li et al. ² studied nonlinear free vibration of functionally graded fiber-reinforced composite hexagon honeycomb sandwich cylindrical shells, providing parameters to weaken nonlinear hardening-spring behavior. Dong et al. ⁶⁷ investigated nonlinear forced vibration of hybrid fiber/graphene nanoplatelets/polymer composite sandwich cylindrical shells with HHC in hygrothermal environment using multiple scale method.

7.1.4. Corrugated and blended core configurations

Emerging core designs combine multiple geometric features for enhanced performance. Lan et al. ⁷⁶ conducted a comparative study on cylindrical sandwich panels under blast loading, showing through numerical analysis that while auxetic honeycomb cores generally demonstrated superior blast resistance due to negative Poisson’s ratio effects, the performance advantage was not absolute. Parameter studies indicated that increasing curvature and face sheet thickness enhanced blast resistance across all core types, with regular hexagonal honeycomb cores capable of exhibiting similar or superior energy absorption depending on the specific configuration. These findings guide theoretical understanding and optimal design of cylindrical sandwich structures subject to external blast loading.

7.1.5. Failure mechanisms and detection

Damage mechanisms such as debonding represent a distinct category from core material classification and are addressed here. Zhou et al. ⁹⁰ proposed a two-step strategy based on the uniform load surface (ULS) curvature method for debonding identification in Nomex honeycomb sandwiches, utilizing local constraints to enhance sensitivity. Lajili et al. ⁹¹ investigated wave propagation in honeycomb sandwich plates through wavenumber-space identification, evaluating robustness against measurement uncertainties.

A comparative study by Lan et al. ⁷⁶ on cylindrical sandwich panels under blast loading further illustrates this configurational dependency. Their numerical analysis using ABAQUS-Explicit showed that while auxetic honeycomb cores generally demonstrated superior blast resistance due to negative Poisson’s ratio effects, the performance advantage was not absolute. Parameter studies indicated that increasing curvature and face sheet thickness enhanced blast resistance across all core types, with regular hexagonal honeycomb cores capable of exhibiting similar or superior energy absorption depending on the specific configuration. These findings guide theoretical understanding and optimal design of cylindrical sandwich structures subject to external blast loading.

7.2. Plates, shells, and curved panels

7.2.1. Cylindrical and spherical shells

Li et al. ³⁸ analyzed FRCS cylindrical-spherical combined shells with HHC, finding vibration suppression enhanced by reducing HHC thickness ratio and cell wall thickness ratio. Eipakchi & Nasrekani ⁴⁷ studied vibrational behavior of composite cylindrical shells with auxetic honeycomb core under moving internal pressure, determining critical velocities and dynamic response. They further analyzed asymmetric free vibration of such shells with adjustable Poisson’s ratio, ⁴⁸ showing frequency decreases with decreasing Poisson’s ratio.

7.2.2. Curved plates and panels

Zhang et al. ⁸⁰ detected delamination in curved composite plates using frequency shifts with ANN and SAGA algorithms. Kumar et al. ⁵⁴ investigated non-linear vibration and parametric optimization of sandwich composite curved shell panels with graphene reinforced skins and auxetic honeycomb core using finite element method and response surface method.

7.2.3. Trapezoidal and special geometries

Rahmani et al. ⁵⁰ analyzed free vibration of cantilever trapezoidal sandwich plate with auxetic reentrant honeycomb core and laminated nanocomposite face sheets using differential quadrature method, finding optimal core-to-total thickness ratio for highest natural frequencies. Ameri et al. ⁴² studied effect of honeycomb core on free vibration of FML beams compared to conventional composites using higher-order deformation theories and Rayleigh-Ritz method.

7.3. Beams, columns, and tubular structures

7.3.1. Beams

Abo-Elkhier et al. ²¹ used modal testing on GFRP cantilever beams to predict fatigue life from damping and frequency changes. Thanikasalam & Ramamoorthy^59,61 investigated free and forced vibration of composite sandwich beams with single-core and dual-core 3D printed honeycomb, showing dual-core provides higher frequencies. Fallah & Mohammadimehr ⁶⁸ studied free vibration of Timoshenko sandwich microbeam with honeycomb core and nanocomposite face sheets integrated with sensor/actuator layers using modified couple stress theory.

7.3.2. Columns and tubes

Akbulut et al. ²⁸ investigated strength loss in buckling of composite columns subjected to fatigue loading, finding samples with no cutout preserve buckling strength up to 40% of total fatigue life. Akbulut & Ertas ⁸ studied connection between fatigue life and modal characteristics in composite structures with geometrical singularities using progressive fatigue approach. Rafiee & Elasmi ¹⁸ modeled fatigue lifetime of composite pipes under internal cyclic pressure using progressive damage modeling.

8. Emerging paradigms: Data-driven prognosis and integrity management

8.1. Machine learning and hybrid models

Machine learning approaches are revolutionizing fatigue prognosis by handling complex, high-dimensional data.

8.1.1. Physics-Guided Machine Learning (PGML)

Zhang et al. ⁷⁸ developed a PGML framework (Figure 4) for fatigue life prediction of CFRP/Al joints under hygrothermal-load coupling, combining a nonlinear damage accumulation model with residual neural network, achieving R²=0.97 with most predictions within 2× error band.

Figure 4.

Physics-guided machine learning framework for fatigue life prediction under hygrothermal-load coupling. ⁷⁸ Chen et al. ²⁶ proposed Physics-informed Transfer Learning (PITL) for fatigue life prediction of IN718 superalloy, integrating transfer learning with equivalent strain theory for improved accuracy with limited data.

8.1.2. Deep learning

Khan et al. ¹⁰¹ designed a CNN to automatically extract features from vibration spectrograms for delamination assessment in smart composite laminates, achieving 94.5% test accuracy. Liang et al. ⁸³ used Support Vector Machine (SVM) and Artificial Neural Network (ANN) with multimode frequencies to predict residual fatigue life of FRP composites, showing ML algorithms outperform single-mode semi-empirical models. Peng et al. ²² implemented Bayesian inference with stiffness degradation model for in-situ fatigue life prognosis of composites, integrating uncertainties in the prognosis framework. Sha et al. ⁸² used Bayesian data fusion for damage localization in beams from natural frequency changes.

8.2. Digital twins and real-time monitoring

Digital twins—virtual replicas updated with real-time sensor data—enable predictive maintenance. Vettori et al. ⁹² developed an adaptive-noise Augmented Kalman Filter (AKF) for input-state estimation in wind turbine blades, using modal expansion for process noise covariance adaptation (Figure 5). This approach allows simultaneous full-field response and unmeasured input prediction. For real-time monitoring, Liang et al. ⁸⁴ implemented FBG-based strain monitoring with temperature compensation for composite tanks under cryogenic conditions. Nicolas et al. ⁸⁸ demonstrated real-time wing shape and load determination using FBG arrays on aircraft wings. These systems form the sensor layer for digital twin implementations.

Figure 5.

Digital twin framework with adaptive Kalman filter for wind turbine blade monitoring. ⁹²

8.3. Artificial intelligence for vibration-based SHM

Artificial intelligence methods are increasingly employed to interpret vibration responses of composite structures. Unlike classical modal threshold approaches, AI models learn complex relationships between modal features and damage states. Supervised learning techniques such as artificial neural networks and support vector machines are widely used for damage classification, while convolutional neural networks directly process spectrograms for automatic feature extraction.

Recent research demonstrates that hybrid physics-informed models significantly improve robustness under environmental variability by embedding stiffness-degradation laws into learning architectures. Transfer learning and domain adaptation methods further enable prediction across different structures and loading conditions. These developments indicate a transition from detection-based monitoring toward predictive structural integrity management.

8.3.1. Supervised and unsupervised learning for damage detection

Supervised learning approaches require labeled data representing both healthy and damaged structural states. Krishna et al. developed feedforward neural networks (FNNs) trained on numerical modal analysis data from Kevlar composites, achieving exceptional predictive performance with an R² score of 0.9312 and average localization error below 5% for crack parameters. ⁹⁶ Their methodology integrated experimental modal analysis (EMA) with numerical simulations in ANSYS Composite PrepPost, demonstrating excellent agreement with average frequency errors below 5%. The FNN architecture proved particularly effective for intermediate-depth cracks of 35% thickness, achieving an error of just 1.08%, while reducing error rates by 60–75% compared to conventional vibration-based methods.

Unsupervised learning methods address the practical challenge of limited damage state data. Saramantas et al. proposed a robust methodology for populations of composite aerostructures under uncertainty, employing unsupervised algorithms for damage detection based on Healthy Subspace representations. ¹⁰² Their approach utilized Multiple Input Single Output (MISO) Transmittance Function AutoRegressive with eXogenous (TF-ARX) excitation data-driven models, with decision-making based on model parameter vectors transformed via Principal Component Analysis (PCA). Damage detection was achieved through multi-level information fusion using acceleration and/or strain sensors, while damage characterization employed a hierarchical cosine similarity-based algorithm validated through hundreds of experiments on composite coupons with delamination and impact damage under various uncertainties.

Reis et al. developed a methodology to identify and classify damage in glass fiber-reinforced plastic composite beams through vibration data and artificial neural networks. ⁹⁷ Dynamic tests were performed on healthy and damaged beams with different delamination sizes to obtain both time and frequency domain data. To address the large dimensionality of vibrational data, a dislocated series strategy was employed to reduce raw signal size into mini batches without losing important damage detection information. Results demonstrated that proper selection of artificial neural network topology and dislocated time series parameters enabled successful damage detection and classification with reduced computational cost compared to direct use of vibration-based model data.

Patro et al. provided a concise review of various methodologies employed for damage detection in composite materials with special emphasis on supervised and unsupervised machine learning techniques. ¹⁰³ Their work highlighted that composite structures suffer from various nonlinear failure modes including delamination, voids, and matrix cracking, making early detection critical. The implementation of artificial intelligence techniques has proven to be a versatile method for damage assessment, with the review outlining major observations to present a broad perspective of the state of the art related to laminated composite structural health monitoring.

8.3.2. Deep learning architectures for vibration data

Convolutional neural networks (CNNs) have demonstrated particular efficacy in processing vibration data transformed into image representations. Jung and Chang developed an advanced deep learning model for impact characterization in composite laminates, implementing piezoelectric ribbon sensors embedded within smart composite fabrics. ⁹⁸ They applied discrete wavelet transform to convert impact signals into input image data for predictive CNN-based models, with hyperparameters optimized through Bayesian optimization theory. Data augmentation secured sufficient training data, and the performance of each optimized neural network model was investigated by comparing test errors under various applied conditions.

Meruane et al. presented a deep learning framework for damage assessment of composite sandwich structures using full-field vibration mode shapes. ⁹⁹ Their methodology employed high-speed 3D digital image correlation (DIC) measurements to identify vibration mode shapes, followed by Gaussian process regression to estimate mode shape curvatures. A baseline-free gapped smoothing method computed damage indices represented as grayscale images, which were processed through a CNN-based algorithm to automatically identify damaged regions. The approach was validated using numerical and experimental data from composite sandwich panels with different damage configurations.

Fotouhi et al. exploited deep learning for quantitative assessment of visual detectability of damage in laminated composite structures, collecting a comprehensive image-based dataset containing microscale damage mechanisms (matrix cracking and fibre breakage) and macroscale damage mechanisms (impact and erosion). ¹⁰⁴ Automated classification using pre-trained AlexNet achieved 87–96% accuracy for identifying damage severity and types in reasonable computational time, demonstrating the potential for autonomous visual inspection of composite structures.

8.3.3. Hybrid and ensemble methods

Hybrid architectures combining deep feature extraction with classical machine learning classifiers have shown improved performance on limited data. Azad and Kim developed hybrid deep convolutional networks integrating CNNs and convolutional autoencoders (CAE) with support vector machines (SVM) for damage diagnosis of laminated composites. ¹⁰⁵ Their approach incorporated advantages of both convolutional operations for deep feature extraction and SVM for diagnosis using limited feature data. Validation using random vibrational signals for one healthy and two delamination states demonstrated improved damage detection performance compared to conventional methods, with lower computational costs and elimination of manual damage-sensitive feature extraction requirements.

Khayyam et al. adapted a hybrid machine learning framework to predict fatigue life of drilled glass fiber reinforced polymer composite laminates under limited and noisy data assumptions. ³¹ Their two-step approach employed an offline deterministic model using group method of data handling (GMDH) with singular value decomposition and Pareto multi-objective optimization to prevent overfitting, followed by Kalman filter updating to minimize mean and variance of error. Results showed excellent learning reliability with correlation coefficients of 97.6% and 96.5% for predicting fatigue life of unidirectional and woven composite laminates, respectively. Sensitivity analysis indicated that fatigue life was more affected by drilling feed rate compared to cutting speed.

Liang et al. focused on predicting residual fatigue life of fiber reinforced polymer (FRP) structures using vibration parameters. ³² Through modal testing and fatigue measurements on FRP beam specimens, they examined the relationship between residual fatigue life and natural frequencies. Two prediction methods—semi-empirical models derived from residual stiffness models based on bending stiffness-frequency relationships, and machine learning algorithms (Support Vector Machine and Artificial Neural Network)—were utilized. Experimental validation demonstrated that ML algorithms can use multimode frequencies unlike the single mode limitation of semi-empirical models. Results showed that ML algorithms outperform single-mode frequency inputs, with higher modes of measured frequencies improving prediction precision. An inverse algorithm based on SVM exhibited higher prediction accuracy and stability even with limited training samples.

8.3.4. Transfer learning and data augmentation

Data scarcity remains a fundamental challenge in SHM applications, as damage state data is often unavailable for real structures. Transfer learning addresses this limitation by leveraging knowledge from related domains. Azad, Kumar, and Kim proposed CNN-based pre-trained transfer learning using ResNetV2 (RNV2) models for delamination detection in CFRP laminates with limited experimental data. ¹⁰⁶ Their approach eliminated the need for developing models from scratch, requiring only fine-tuning on the target composites dataset containing multi-class wavelet-transformed vibrational data. Results demonstrated that pre-trained RNV2 models can effectively perform SHM even under limited data conditions.

Generative adversarial networks (GANs) offer another solution through synthetic data generation. Azad, Kim, and Kim introduced a novel multi-class data augmentation method employing multi-class generative adversarial networks (MC-GAN) for the first time to generate synthetic data for multiple damage classes without excessive experimentation or simulation. ¹⁰⁷ Their MC-GAN-CNN model for autonomous data augmentation and delamination detection, validated using experimentally obtained vibrational data for laminated composites, achieved mean accuracy of 99.72 ± 0.08% with rigorous 10-fold cross-validation assessment. The proposed framework enabled autonomous delamination detection without requiring hand-crafted statistical features while maintaining good generalization capability.

8.3.5. Explainable AI for trustworthy SHM

The black-box nature of many AI models poses challenges for safety-critical aerospace applications where reliability, trustworthiness, and validation capability are paramount. Pedraza, del-Río-Velilla, and Fernández-López addressed this challenge by applying explainable artificial intelligence (XAI) techniques to evaluate an Impact-Locator-AI model before embedding it in aerospace structures. ¹⁰⁸ Their case study demonstrated that XAI methodologies enable critical evaluation of AI model reliability and potential suitability for aerospace applications, establishing groundwork for future research at the intersection of XAI and impact location in SHM.

Azad and Kim further advanced this direction by proposing an interpretable deep learning model based on an explainable vision transformer (X-ViT) for reliable damage detection in polymer composite structures. ¹⁰⁹ Their approach validated on CFRP composites with multiple health states exhibited better damage detection performance compared to existing methods. Critically, the X-ViT approach effectively highlighted areas of interest related to each health condition prediction through patch attention aggregation, emphasizing their influence on decision-making and providing improved diagnostics with increased transparency and reliability.

8.3.6. Comprehensive reviews and future directions

Recent comprehensive reviews synthesize the rapidly expanding field of AI-enabled SHM for composites. Saleh, Jaber, and Hussein reviewed advanced damage detection techniques covering non-destructive assessment methods, metaheuristic optimization techniques, machine learning approaches, and advanced signal processing algorithms applied to SHM of composite structures. ¹⁰⁰ Their analysis highlighted significant progress in integrating sensing, signal processing, and computational methods, with hybrid models showing improved robustness in noisy environments. However, challenges persist in real-time monitoring, data scarcity, environmental variability, and large-scale deployment. Future development should focus on embedded sensor networks integrated with edge computing, adaptive hybrid models, and scalable cloud-based SHM frameworks incorporating real-time localization, severity classification, and predictive maintenance.

Scarselli and Nicassio presented a comprehensive examination of recent advances in ML-based SHM methods tailored to aerospace applications, covering supervised, unsupervised, deep, and hybrid learning techniques. ¹¹⁰ Their review highlighted capabilities in processing high-dimensional sensor data, managing uncertainty, and enabling real-time diagnostics, with particular attention to challenges of data scarcity, operational variability, and interpretability in safety-critical environments. Emerging directions such as digital twins, transfer learning, and federated learning were explored, providing a roadmap for future research and identifying key enablers needed to bring ML-based SHM from laboratory development to widespread aerospace deployment.

Altabey provided a comprehensive treatment of composite materials and structures, addressing fundamental principles and advanced applications. ¹²² The work encompasses theoretical foundations alongside practical considerations for structural health monitoring and damage assessment in composite systems. Shangguan et al. investigated structure optimization and vibration fatigue analysis of aluminum and composite brake control boxes for EMU applications. ¹²³ Through static and fatigue testing of T700/5429 CFRP, combined with stress analysis and vibration fatigue life predictions using Dirlik’s theory, their lightweight design approach integrated structural and material optimizations. Results demonstrated noteworthy reduction in damage to the composite box following lightweight modification, successfully satisfying stiffness and strength criteria essential for EMU safety and reliability. ¹²³

8.4. Future outlook and challenges

Despite advances, several challenges persist:

1. Environmental Robustness: Models must account for coupled hygrothermal-mechanical effects.^18,78

2. Scaling and Validation: Methods validated on coupons need demonstration on full-scale structures. ¹⁸

3. Data Quality and Quantity: ML approaches require large, high-quality datasets.^78,83

4. Real-Time Processing: Edge computing for real-time prognosis remains challenging. ⁹²

5. Multiphysics Integration: Combining vibration data with other NDE techniques (thermography, AE).^20,86

Future research should focus on: (1) Developing standardized benchmark datasets for composite fatigue; (2) Creating explainable AI models for engineering trust; (3) Implementing edge-AI for real-time onboard prognosis; (4) Establishing certification frameworks for data-driven SHM systems; (5) Exploring self-sensing composites with integrated sensing capabilities.

9. Conclusions

This review has synthesized current knowledge on fatigue behavior and vibration-based integrity management of advanced composite structures, encompassing references across experimental, computational, and data-driven approaches. Key conclusions are:

1. Vibration-based SHM is a powerful paradigm for non-destructive fatigue assessment in composites, leveraging the intrinsic relationship between stiffness degradation and changes in natural frequencies, damping ratios, and mode shapes.^13,16,21 This approach has been validated across diverse composite configurations, from conventional laminates to advanced sandwich structures with honeycomb, auxetic, and functionally graded cores.^{2,7,34,36,37,40,41}

2. Advanced sensing technologies—particularly FBG sensors,^84,88 AE, ⁸⁶ and DIC ⁹⁴ —enable high-fidelity monitoring of damage progression. When combined with computational methods like FEA,^3,5 mesh-free techniques, ¹ and higher-order shear deformation theories,^7,47 they provide comprehensive tools for vibration analysis of complex, damaged structures.

3. Fatigue life prediction has evolved from phenomenological models ¹¹ to progressive damage approaches^9,18 and hybrid methods integrating physical models with machine learning.^78,83 Optimization algorithms (PSO, GA, SA) are increasingly used to design laminates for maximum fatigue life.^{5,12,14,15,125}

4. Honeycomb sandwich structures exhibit complex dynamic behavior influenced by core geometry, face sheet material, and bonding integrity.^34,36,60 Auxetic cores with negative Poisson’s ratio offer unique vibration and energy absorption characteristics,^40,41 while functionally graded designs enable tailored dynamic performance.^2,53

5. Environmental factors—temperature, ⁸⁹ moisture, ⁷⁸ and variable amplitude loading ¹⁸ —significantly affect fatigue behavior and must be accounted for in prognosis models. Stochastic approaches are necessary for real operational conditions. ²³

6. Emerging data-driven techniques—particularly physics-guided machine learning, ⁷⁸ digital twins, ⁹² and transfer learning ²⁶ —offer promising pathways for accurate fatigue life prediction under complex multi-physics conditions. However, challenges remain in model interpretability, data requirements, and real-time implementation.

7. Manufacturing processes significantly influence structural integrity and dynamic performance.^111,113,114 Process optimization, particularly through techniques like rotary ultrasonic machining, can enhance final component quality and fatigue resistance.

A critical next step is the development of standardized benchmark composite materials and open vibration-fatigue datasets to enable objective validation of monitoring algorithms across industries.

Future research should prioritize: (1) Development of standardized benchmark problems and datasets; (2) Creation of explainable AI models for engineering adoption; (3) Implementation of edge computing for real-time prognosis; (4) Exploration of self-sensing composite materials; (5) Establishment of certification frameworks for data-driven SHM systems. As composites continue to penetrate critical applications, robust vibration-based integrity management will be essential for ensuring safety, reliability, and cost-effective operation throughout service life.